What should the next shift know?
Could a small shared maintenance handover log make unfinished work, responsibility and the next action clearer between shifts?
Sets the decision for the investigation.
↳ AGENTIC BRAINSTORMING AND IDEATION IN PRACTICE
Agentic brainstorming in practice: discover a direction, test the reasoning and shape a specification a builder can use.
Follow the fragments through research, competing explanations and a considered scope. Inspect the prompts, evidence and outputs, including specifications, presentations and hardware concepts. The cases preserve their open questions and review notes so you can judge what is ready and what still needs work.
Each case gives you the inputs, the platform steps and an inspectable outcome. Search the index for a capability, follow the recorded execution and adapt the method to your own evidence. The shorter recipes explain reasoning patterns you can reuse.
Create an investigation and connect a model in Model access. Review processing consent and limits there. Reading this cookbook and downloading its example use no model calls. Repeating the research or package generation uses your connected provider and the platform allowance shown in your workspace.
Explore real platform workflows, from a bounded investigation to connected data, verified experiments, editable presentations and mechanical designs. Each case explains its inputs, what actually ran and what the result establishes.
11 of 11 cases shown

Turn a fictional offline inspection decision into an eight-slide editable PowerPoint with source-bound metrics, charts, a workflow, stop rules, and actual PPTX page previews.
Research and evidence
A research agent reads a retained NASA source excerpt, cites a passage, and turns a narrow claim into an automatically checked result.
Documents and presentations
Pin an exact interview and human decision, revise a product brief, then freeze and download an immutable draft archive whose missing checks stay visible.
Research and evidence
A typed Clarify → Explore path is previewed, reviewed, and saved as a private template that imports with fresh, unconfigured inputs.
Experiments and data
A replayed JSON feed becomes an immutable dataset, a verified numerical test, and a saved 70% versus 85% comparison.
Software
Investigate three supplied cards, challenge the software idea, and export the first planning package.
Software
Review ownership, acknowledgement, backup and scope before handing a small web application to a builder.
HardwareInspect a multipart mechanism, its editable CAD, assembly, dimensions and proposed fit checks.
Mixed productsFollow mechanical interfaces, dimensioned parts, a status PCB concept and the staff inspection handoff.
Hardware
Start with a single printed part, exact dimensional constraints, native CAD and a small prototype plan.
Earlier case · open the execution notesMixed products
Keep a proposed job board, physical marker and staff workflow consistent through an independently reviewed contract.
Earlier case · open the execution notesNo cases match those filters. Try another capability or reset the search.
Live model runs, deterministic protocol replays and supplied design inputs are identified in their respective execution notes. Automatic checks establish only the assertions and versions they tested.
Turn a fictional offline inspection decision into an eight-slide editable PowerPoint with source-bound metrics, charts, a workflow, stop rules, and actual PPTX page previews.

A fictional 12-technician service region is considering offline inspection capture. The deck asks for one four-week pilot, with no automatic expansion.
The assumed 8.4 team hours per week, fictional six-person prompts and calculated 4.29-hour savings scenario come from the downloadable investigation. The build checks source pointers and arithmetic before generating slides.
The platform's trusted template and native PPTX projector create editable text, metrics, bars, a risk table, a timeline, shapes and attached workflow connectors, plus a PDF and portable source bundle.
The guarded isolated renderer captures all eight PPTX pages and checks browser, PDF and native views. The montage below comes from the rendered PowerPoint pages.
The recommendation, fictional assumptions, proposed workflow, stop rules and three expansion gates live in an eight-slide editable deck. The renderer validated the deliverable; real pilot results remain unmeasured.


The story, source arithmetic, native export path, validation scope and reproducibility steps.
This case follows a fictional product decision for Northline Service. Twelve field technicians inspect pumps where connectivity is unreliable. The proposed change is offline capture with a visible local queue and explicit reconciliation. The deck asks for a four-week, one-region pilot; wider rollout waits for measured gates.
The PowerPoint deck is the finished artifact. Its text, metric panels, bars, comparison table, timeline, shapes and workflow connectors are native editable objects. The PDF and slide montage make it easy to review without PowerPoint. The montage uses captures of the actual exported PPTX pages, rather than browser previews or slide screenshots inserted into a deck.
The investigation is deliberately synthetic. It assumes a 12-technician region spends 4.1 team hours per week retrieving photos, 2.6 on duplicate entry and 1.7 on sync rework, totalling 8.4 hours. A planning scenario assumes reductions of 50%, 60% and 40% respectively: 4.29 team hours saved and 4.11 remaining per week. None is an observed customer result or a forecast. The invented interview counts—5, 4 and 3 out of six fictional prompts—identify design questions, not statistical evidence.
Eight slides move from the bounded authorization request through the assumed baseline, fictional interview signals, a proposed capture-to-review flow with an explicit conflict hold, the calculated savings scenario, stop rules, a four-week pilot protocol and expansion gates. The gates are 95% reconciliation within 24 hours after connectivity returns, a 30% median handoff-time improvement and ownership plus resolution state for every missing attachment. Median handoff time is a separate pilot measure from the weekly team-hours scenario. A lost record or silent overwrite pauses the pilot. The complete typed authoring input is deck-input.json; its source references and speaker notes make the assumptions traceable.
The repository script scripts/cookbook-decision-presentation.mjs constructs typed slide blocks from the investigation, checks the arithmetic and source pointers, and calls the platform's renderTrustedDeck, measurePptxLayout and renderEditablePptx modules. The platform template defines the layouts; the native projector produces the editable PowerPoint. The local build also creates browser previews, HTML, CSS, licensed font inputs and the portable source bundle with its render recipe.
The deck then ran through the current guarded platform image built from the-confused-engineer/rendering/Dockerfile, with no network and read-only renderer input. That isolated run produced the final PPTX and PDF, embedded two licensed static Space Grotesk font faces, validated the OOXML package, rendered the actual PPTX pages through LibreOffice, and checked browser, PDF and PPTX views. The validation summary records 48/48 executed checks across eight slides. The full renderer receipt keeps the hashes and per-view evidence. The validated package contains 130 text boxes, 48 other shapes, four attached connectors and no raster slide images.
The receipt remains pending-independent-vision. We visually inspected the exported pages for this cookbook case, but no commissioned SH21 package review or separate release approval was run. Rendering checks establish export integrity and visible page fidelity; the fictional business assumptions still need a real pilot.
From the repository root, run node scripts/cookbook-decision-presentation.mjs --authoring-only to check the typed input and its source pointers without touching the published files. Run node scripts/cookbook-decision-presentation.mjs --stage for a full local build in outputs/manual-20261006-native-presentation/presentations/decision-presentation/stage/. It uses the repository-pinned Space Grotesk fixture and a local Playwright Chromium installation, and writes renderer-input/ for the isolated run.
For the qualified render, build the current the-confused-engineer/rendering/Dockerfile image. Invoke its offline.mjs entrypoint under the renderer's documented containment with renderer-input/ mounted read-only at /input, an empty writable /exports, and no network. The image includes LibreOffice Impress, Poppler and Python fontTools. Import the output into the stage with node scripts/cookbook-decision-presentation.mjs --stage --import-qualified=/absolute/path/to/exports. The import requires exact deck, HTML and CSS digests in both the staged render recipe and qualified receipt, verifies all 48 checks and the native PPTX hash, then replaces the stage artifact and previews with actual PPTX-rendered pages. Review those pages before copying an accepted stage to the public case.
The downloadable source kit and recipe record the renderer source hashes for this instance. Its exact renderer hash is also in the qualified receipt.
Northline Service, interview prompts, times and targets are synthetic example inputs. The downloadable investigation and typed deck identify assumptions and proposed gates. The PPTX, PDF, actual PowerPoint page captures and 48-view receipt were generated locally through the guarded platform renderer. Its reviewState remains pending-independent-vision; this is neither a commissioned SH21 approval nor a measured customer result.
A research agent reads a retained NASA source excerpt, cites a passage, and turns a narrow claim into an automatically checked result.

The person asks whether NASA GISS reported a 2020 anomaly of at least +1.00 °C relative to 1951–1980. The question stays on My canvas.
The agent dispatches search_public, then read_source. The source reader executes against a labelled offline replay of a selected NASA excerpt and stores the response hash, passage ID and coverage. The official page was separately fetched with the web tool.
The agent proposes a finding on Agent canvas. Synatrail keeps a quote only when it appears verbatim in the exact read passage; a deliberately invented sentence is rejected in the receipt.
A structured compile_test_plan call is validated by Synatrail. Its deterministic sweep computes 1.02 − 1.00 = +0.02 °C; the measurement meets the compiled threshold. The result names the source version and test revision.
The Agent finding remains pending. A person can inspect the retained excerpt, accept or challenge the finding, and request a fresh full-source read before using it elsewhere.
The replayed excerpt supports a +1.02 °C figure, its citation survives the automatic quote check, and Synatrail derives a met verdict from the threshold result. The official page was separately fetched; the platform runner used the retained excerpt replay. No paid model call or human acceptance is claimed.
Source, automatic checks, result and unresolved review.
Question. Check whether the NASA GISS 2020 annual global anomaly reached at least 1.00 °C relative to the 1951–1980 baseline.
Agent proposal. NASA GISS reported 2020 at +1.02 °C against the 1951–1980 mean. It is pending human review on the Agent canvas.
Source. NASA Earth Observatory. The official page was separately fetched with the web tool on 2026-10-05; its retained extracted lines are source-web-retrieval.txt, SHA-256 73fa70a2f807eec9529d1417d38b0f720eb09dd19b3e84d40454eb7cc4e10dec. The agent's Synatrail source read used a selected-excerpt replay in source-response.txt, SHA-256 135cd968d22b1432b147e74e05365df00d562c278d83c80bdeb6bd116b147ac6, because shell DNS was unavailable. The replay quote was checked against line 268 of the official web retrieval. The platform reader executed against the retained replay bytes and produced passage paragraph-2-0. Neither file claims to be the full raw NASA HTTP response.
Automatic evidence check. The real quote verifier kept one exact quote and rejected a sentence absent from the passage (not-verbatim).
Compiled test. The cited 2020 anomaly is at least 1.00 °C above the 1951–1980 mean. The real compiler accepted a sweep computation and an executable measurement. Temperature anomalies are differences, so the compute engine used kelvin intervals; a 0.02 K interval equals 0.02 °C. The engine returned a margin of 0.02 °C against +1.00 °C, so the narrow threshold test is met for the retained excerpt.
Next review. A person must inspect the source and decide whether to accept or challenge the pending finding. A fresh full NASA data table read and its uncertainty notes remain open.
Exact command, fixture boundaries and generated files.
This case asks a narrow question: did NASA GISS report a 2020 global surface temperature anomaly of at least +1.00 °C relative to the 1951–1980 mean? The retained NASA Earth Observatory page extraction includes its +1.02 °C statement. A Synatrail research agent follows a source lead, reads a selected excerpt, proposes a cited finding, and asks for a numerical test. Synatrail rejects a fabricated quote, compiles the plan, computes the margin, and derives the result met. The Agent finding remains pending for a person to review.

The source path has two deliberately distinct records. The official NASA article was fetched through the web tool on 5 October 2026; the retained web extraction identifies the article, line 268 and its extracted text. This is an actual fetched page extraction, not raw HTTP response bytes. The local shell could not resolve NASA's host, so the platform read_source step used source-response.txt, a selected excerpt whose quote is checked against the independently fetched page. The case saves the replay's original bytes, SHA-256, source version, passage IDs, read coverage and date in source-receipt.json. The selected excerpt is shorter than the full article. Re-running this case does not fetch the current NASA page.
The agent calls are a deterministic OpenAI Responses protocol fixture. They pass through Synatrail's actual callResearchProvider dispatch, which checks the allowed tool, emits a function-body authority, and parses each function call. There are six recorded calls: search_public, read_source, propose_finding, assess_evidence_passages, compile_test_plan, and finish_research. No paid model call is made, and the script does not claim the fixture independently discovered the source or chose the claim. Its purpose is to make the platform's research and verification path reproducible without a credential. Prompt hashes, selected tools, source read and results appear in investigation.json.
The agent's pending finding cites sourceVersionId:passageId, so a reader can return to the exact retained text. The evidence verifier accepts one copied phrase from that passage and drops one invented sentence with not-verbatim. The accepted quote supports the reported number; it does not by itself settle scientific uncertainty or prove causation. A person has not accepted the finding onto My canvas.
The test compiler checks the metric, units, threshold, sample rule, falsifying observation, parameter bindings, and deterministic tool configuration. The compute engine runs a one-scenario sweep for 1.02 − 1.00 = +0.02. Temperature anomalies are differences, so the computation uses kelvin intervals; 0.02 K is the same interval as 0.02 °C. Synatrail's result path extracts the computed metric and sample count, then cleanResult derives observedVerdict: "met" against the registered gte 0 threshold. test-receipt.json contains the compiled plan, evidence check, computation hashes, metric, sample and platform-produced result. This verifies only the arithmetic threshold for the retained source excerpt.
The canvas image is rendered with Synatrail's actual ResearchNode, Relationship and React Flow components from the recorded case data. It is an authored presentation of the run, not a live account screenshot. The source inspector image is a separate capture of the current app: an isolated signed-in board ingests the same retained excerpt as a text file and opens its real source viewer. That reproduction has its own capture receipt; it is not a claim that the agent research receipt created that temporary board.
To reproduce from the repository root:
node scripts/cookbook-agent-evidence.mjs
The script writes only this case folder. It uses the repository's cached Chromium and an ephemeral 127.0.0.1 listener for the native canvas and inspector captures. The app's temporary database is removed after capture. Run from a workspace that permits local listening and has the existing Synatrail browser dependencies installed. It asserts the official web extraction contains the replayed quote; that the source reader makes exactly one bounded replay read; that the cited passage exists; that the fabricated quote is rejected; that the compiler accepts the executable test; that the compute and platform result paths produce the measured value, sample and met verdict; and that the native graph has no clipped text or label collisions. Browser capture checks for page errors and unexpected external requests.
The next human step is to inspect the original NASA page and readable finding, then accept or challenge the pending Agent finding. A fresh full data-table read and NASA's uncertainty notes remain open if the question grows beyond this one reported value.
Executed with real Synatrail callResearchProvider, readCanvasURL, verifyEvidenceFindings, cleanCompiledTest, runComputation, computationMeasurement and cleanResult; deterministic OpenAI protocol and selected source-response fixtures. Official NASA page independently fetched by web.run; the Synatrail source reader used replay bytes. Native canvas components and current app inspector rendered locally. No external model call.
Pin an exact interview and human decision, revise a product brief, then freeze and download an immutable draft archive whose missing checks stay visible.

A fictional support-lead interview is saved as an exact source version. The one-team pilot decision is selected at its investigation revision, and the commission captures both references.
The software-product preflight records a zero-credit manual-draft path and shows unavailable automatic-generation capabilities. The saved job remains paused; no provider is called.
A clearly labelled local human seed for SH04 is edited through the platform revision API. The original revision stays stored, while revision 2 is blocked and marked unvalidated.
The release API assembles a downloadable ZIP and manifest with the manual edit and 27 missing outputs. The archive-bound agent handoff is available; ready publication and approval remain blocked.
The draft ZIP preserves the pinned scope and unvalidated human brief for a later team or agent. Its manifest and workbench expose missing artifacts; the platform rejects a ready release and draft approval until an automatic run and independent checks complete.


The pinned fictional source, manual commission, unvalidated revision, immutable draft, guarded ready state, and local rebuild command.
PulseNotes is a fictional internal support-triage app. In this example, a three-person support team receives bug reports through chat. A support lead estimates that two of ten weekly reports need a second clarification because the original message was not linked. That number is an interview estimate, not telemetry. The proposed pilot records ten consecutive reports with a title, impact, owner, status, and original chat URL. Success means the team can identify each report's owner and source without revisiting chat history.
The platform stores the interview as version:interview-1 and a separate human decision to run only a one-team pilot. The commission selects that exact source version and the decision at the saved investigation revision. The pinned-context receipt records the passage digest and revision; the fictional source is downloadable so the claim can be checked.
The commission and preflight use the software-product preset with manualDraft: true, no authorized provider routes or tools, and a zero-credit quote. Preflight permits the manual draft but reports automatic generation unavailable. The saved package is paused, with no model work billed.
The local fixture seeds one clearly labelled human SH04 product-and-opportunity brief so the real revision API has an existing canonical source to edit. This seed is not a model result, validated output, or qualification. appendManualArtifactRevision saves revision 2 with a new question about source-link retention. The original source revision remains stored. The revised artifact is manual-unvalidated, its task is blocked, and it has no model receipt, local validation, or independent review. The revision lineage makes both versions inspectable.
preparePackageRelease and commitPackageRelease then create the immutable draft ZIP. Its manifest names the selected revision, the human edit file, portable package templates, and 27 missing outputs. The release download endpoint returns exactly the same ZIP bytes; the verification receipt records the SHA-256 hashes and byte comparison. The agent handoff binds its instructions to this exact archive and requires the complete archive alongside the prompt.
The workbench screenshot shows the paused job and disabled ready-archive control. The release and artifact previews are captures of the actual local app after the draft was saved. A ready-release attempt returned HTTP 409 because only an active automatic run can publish ready output. Approval of this draft also returned HTTP 409. These are real platform guards, not an assertion that the pilot package passed review.
From the repository root, run:
node scripts/cookbook-package-handoff.mjs --screenshots
The script starts a throwaway self-hosted app and signed-in local fixture account, saves the investigation, source extraction and decision, then uses the package commission, context, job, revision, release, and owner-scoped HTTP read/download APIs. It checks the archive contents and SHA-256, exercises the ready/approval guards, captures browser panels, writes the assets in this folder, and closes/removes the throwaway app. No provider credential, network research, paid tool, repository push, or external account is used. Generated IDs, timestamps, and archive hashes can differ on a fresh run; the links within each run remain exact.
To inspect the ZIP independently, run unzip -l public/the-confused-engineer/cookbook/package-handoff/draft-archive.zip. The archive's own README.md and tools/validate-package.py describe the platform package format. A later ready release would require a new qualified automatic commission, completed outputs, passed local checks, and independent review; this example stops deliberately at a durable manual draft.
PulseNotes, the interview, estimate, and decision are fictional local example inputs. The first SH04 source is explicitly fixture-seeded human content, not a generated or validated artifact. The second revision, paused job, release envelope, ZIP, manifest, download, and agent handoff were produced by the platform's actual package store and runtime in an isolated local app. Browser screenshots show that saved app state. Zero provider calls and credits; no independent qualification, ready release, repository push, or customer outcome is claimed.
A typed Clarify → Explore path is previewed, reviewed, and saved as a private template that imports with fresh, unconfigured inputs.

A paper compares two workshop bench layouts. Clarify targets constraints and unknowns; Explore targets reversible trials. Validated material-flow connections join their named ports.
The platform previews both steps in order, pins their authority and input basis, and records an exact review. The downstream step has no output from Clarify yet, so this two-step trace is not start-eligible and no model is called.
A separate Clarify-only preview becomes start-eligible after review. Editing its source note changes the pinned input basis, and the attempted start is rejected with HTTP 409.
A private template carries the two operation types and their connection into a second investigation. Two new unconfigured steps and one empty material placeholder require fresh binding and setup before dispatch.
The source path has a saved preview and review, not executed reasoning. Its single-step stale start is blocked. The new investigation has two unconfigured steps and one unresolved input; no source note, approval, answer, or provider call transfers.
The configured steps, reviewed preview, stale-input check, private template transfer, and local reproduction command.
Question: Which of two workshop bench layouts should be tested before a permanent move? One layout keeps a tool cart beside the test bench; the other moves it to a shared wall. The source note in this case is invented for the walkthrough.
This case configures a bounded Clarify → Explore reasoning path on My canvas. Clarify is asked to identify constraints and unknowns. Explore is asked to suggest reversible trials and observations that could distinguish the layouts. The two connections use the platform's named out and in ports and pass its type checks. The trace preview records the selected steps, their order, exact input basis, provider/model identity, and expected effects. Reviewing the preview does not run either step.
The two-step preview is ordered but cannot start as a single trace: Explore's input is the still-unproduced Clarify output. A separate Clarify-only preview has all inputs available; after exact review it becomes start-eligible. The reproduction then revises the source note. That makes the pinned input basis stale, and an attempted start returns HTTP 409. No model output, answer, or research receipt is invented.
The useful artifact is a private workflow template. Saving the two selected steps stores their operation types, order, typed internal connection, and a material input placeholder. It omits the source note and the step-specific intent. Applying it to a second investigation creates new step identities. Its two steps have missing-setup status, the material placeholder has no selection, and dispatch is unauthorized until somebody configures the steps and chooses fresh input. The source's note, private marker, review, and results are absent from the imported canvas.
The screenshots show the actual local app's source canvas, path preview, and applied canvas. The JSON receipt records the platform preview and review, template projection, import manifest, empty bindings, and stale-input rejection. The screenshots and receipt come from the script below; the template save and apply in the script use the same owner API as the interface controls.
From the repository root, run:
node scripts/cookbook-configured-workflow.mjs
The script starts the app in its isolated browser harness, creates the illustrative paper and two configured steps with production graph helpers, previews and reviews the path, saves and applies the template through production endpoints, captures browser screenshots, and verifies the stale-input guard. It explicitly blocks outbound fetch calls after seeding a synthetic fixture credential used only to show provider/model authority in the preview. It asserts zero provider calls, zero research runs, and zero research receipts. No live provider credential or paid model execution is required. A fresh run replaces the random local IDs in the JSON downloads, so compare their structure and statuses rather than their IDs.
Isolated local Synatrail execution on 2026-10-05 using an illustrative workshop note and synthetic fixture provider credential for preview identity only. Real platform graph validation, trace preview/review, stale-input guard, private template save/apply, and browser UI produced these artifacts. Outbound provider calls were blocked and counted: zero provider calls, research runs, and research receipts. No model output or live-source research is claimed.
A replayed JSON feed becomes an immutable dataset, a verified numerical test, and a saved 70% versus 85% comparison.

A controlled API read replays six illustrative occupancy rows into a pinned dataset version. The My canvas connector names the exact version used by the test.
A scripted agent proposal asks whether at least two rows reach 70% occupancy. Its numeric filter and count metric are fixed before the run.
At full automation the platform reads the connection, wires it to the Data port, computes three qualifying rows, and records a verified, met result with provenance.
An 85% branch uses the same dataset and retains one row. The comparison preserves the baseline and records the two-row difference.
The registered minimum of two is met. At 85%, only one fixture row remains. These are computed replay results, not measurements of a live library.
The source fixture, canvas sequence, exact verdict, comparison and local reproduction command.
Worked fixture replay. Six illustrative JSON rows stand in for a public-data endpoint. No live library, public API, or model provider was contacted. The connector, immutable dataset, numerical computation, automation lane, result verification, canvas wiring, and branch comparison are real Synatrail operations in an isolated local app. The hypothesis, direction, and compiled test proposal are scripted fixture inputs so the run is deterministic.
Can a desk use a 70% occupancy alert to catch recurring crowding? We register a falsifiable rule: at least two rows in the six-row feed must reach 70%. The fixture contains 42, 74, 87, 51, 78, and 62 percent, so three rows qualify. A stricter 85% alternative retains one row. These counts describe the fixture only; they do not establish any real branch's occupancy or operational alert quality.
library.example.test endpoint and saves the six returned rows as a versioned dataset. Its URL is illustrative and will not serve data in your browser.discovery_results, and marks the test verified. “Verified” means that the registered computation completed against the pinned fixture; it is not independent verification of the fixture's origin.The receipt records the fixture SHA-256, the connector's single read, its dataset version ID, the computation's input hash and result, the registered threshold, the verified verdict, and the branch comparison. IDs and timestamps are generated by each isolated run, so they change on reproduction. The input rows and expected counts do not.
From the repository root with Node 24 and the project's installed dependencies:
node scripts/cookbook-api-experiment.mjs
The script starts a private temporary Synatrail app on localhost, creates a signed-in fixture account, supplies the exact JSON response through the controlled connector reader, runs the real test lane, captures browser screenshots, writes the artifacts in this directory, and removes the temporary app database. The model_activation row enables the internal research computation gate for this fixture account; it does not connect a model or grant access to a live provider. Rerunning replaces these public example artifacts with the new run's IDs and timestamps.
Isolated local Synatrail execution on 2026-10-05. The API response and agent proposal are explicit deterministic fixtures. No live public API or model provider was called; the connector read, immutable dataset, numerical worker, verification, canvas state and branch comparison are actual platform outputs.
A small workshop wants the next shift to know what is unfinished, who owns it and what happens next. Investigate whether a simple shared handover log is worth testing.
Decision: what should a first prototype contain, and what would make us abandon it? The cards below summarise the exact downloadable inputs, which are synthetic teaching material. A live model response can expose gaps in those inputs; it cannot establish that a real workshop has this problem.
Could a small shared maintenance handover log make unfinished work, responsibility and the next action clearer between shifts?
Sets the decision for the investigation.
A coolant drip, a chuck wobble and a repair awaiting review. The fictional notes name people, actions and unresolved work. They illustrate a workflow; they do not measure demand or savings.
Role: informs the question and constrains the idea.
One shared workshop tablet: record outstanding work, an owner, a next action and acknowledgement; export a backup. Compare it with a structured paper checklist. Operational decisions stay with the supervisor.
Role: proposes an answer, still to be tested.
Open New investigation, name the project and set its question. Use the downloadable investigation record for the exact card text. Add only those three cards; keep responses in the conversation or notebook.
The recorded live run selected the three cards together without saved edges. As an optional follow-up, connect the observation to the question and the idea to its context. Explain what each relationship supports and what remains unproven.
Select all three cards, check that the assistant scope names your selection, and use Ask. Request a decision, its basis and the cheapest next test. Keep automatic exploration paused for this bounded example.
Synatrail can add its answer as a canvas card. To keep this example at three nodes, remove that extra card while retaining the saved conversation. Review the objections and open questions, then use recipe 06 to package the original three cards and the saved answer.
Using only these three cards, assess whether a workshop handover log is worth a small prototype. Separate supplied observations, assumptions and your inferences. Compare the idea with using a structured paper checklist. Give the smallest useful scope, the strongest reason to reject it, and one test with an observable pass/fail rule. Treat the example notes as synthetic. Do not claim interviews, savings or adoption that we have not measured.
The recommendation: test the paper checklist first. The live answer proposed a two-week trial covering at least ten handovers before committing to software. It proposed targets of 90% complete entries, 80% of acknowledgements within fifteen minutes of shift start, and at most 10% ambiguous items. These are suggested pilot thresholds, not measured results.
The light package: three modules, six stories and six requirements for a local, single-device log. It remains a draft requiring review. No application was built and no comprehensive specification run was started.
What the reviewed revision addresses below: its first build wave says “parallel” although the brief asks for one builder and the UI depends on storage. Sequence storage before the UI. It also suggests local telemetry and a comparison view; decide whether those belong in the first version. A completed generation still needs this scope review.
The downloadable evidence and generated package are in English. The JSON is a readable record: recreate the three cards manually, as Synatrail has no canvas importer. The copyable prompts on this page are shorter adaptations; the exact live prompts are linked above.
Stop when: you can state what to test, why this evidence justifies that test, and what observation would reverse the decision. More cards are useful only if they can change one of those answers.
A second pass gives focused questions to the reviewers and reconciles their answers into a revised specification.
The review focuses on five concrete decisions: sequence the build for one builder; keep acknowledgement separate from ownership; remove unnecessary telemetry; make backup and restore agree; and test whether software adds value beyond the paper checklist.
The resulting light specification: scope, requirements, architecture, risks and the decisions still needing evidence.
A proposed single-device handover log, conditional on validating a structured paper process first.
Kind: software · Investigation: Workshop handover, reviewed by an orchestrator · Composed: 2026-10-01T19:54:43.304Z · Model details: see the original download
Who: Staff and supervisor in the fictional two-shift workshop.
Pain: The synthetic notes illustrate ambiguity about ownership, next action and closure; they do not establish real-world demand or impact.
Context: Investigate the paper process before building software. Operational and safety decisions remain with the supervisor outside the app.
Evidence: handover-question What should the next shift know?; sample-handover Three sample handover notes; handover-concept A single-device handover log
If the paper-process gate passes, prototype a local browser log on one shared device. Capture six core fields, show unresolved items first, record receipt acknowledgements and export JSON and CSV.
Differentiators
Non-goals
User journey
Success criteria
Entities
| Entity | Fields | Source |
|---|---|---|
| handover_item | item_id: generated identifier, equipment_label: required text, issue_short: required text, current_owner: required text, next_action: required text, status: open | awaiting_review | closed, timestamp: creation time, updated_at: last edit time, acknowledged_by: optional entered initials, acknowledged_at: optional receipt time | Six core fields proposed in handover-concept; status and ownership ambiguity illustrated by synthetic sample-handover. Identifiers and metadata are design choices. |
Evidence used as data
| Evidence | Use |
|---|---|
handover-question What should the next shift know? |
Goal and validation question. |
sample-handover Three sample handover notes |
Synthetic examples of handover ambiguity, not field research. |
handover-concept A single-device handover log |
Candidate scope, fields, acknowledgement and exports. |
Proposed vanilla JavaScript UI and direct IndexedDB, with bundled static assets served on the shared device at http://localhost:4173. No external requests or backend.
Components
| Component | Responsibility | Technology |
|---|---|---|
| Staff UI | Create, edit, list and acknowledge items. | HTML, CSS, vanilla JavaScript |
| Local store | Persist records under the fixed origin. | Direct IndexedDB API |
| Exporter | Download outbound-only JSON and CSV. | Browser Blob and download link |
| Local static server | Serve bundled assets at the fixed origin. | Device-local static HTTP server |
Interfaces
| From | To | Protocol | Purpose |
|---|---|---|---|
| Staff UI | Local store | In-browser IndexedDB calls | Read and write items. |
| Staff UI | Exporter | In-browser function call | Produce local downloads. |
| Local static server | Staff UI | Loopback HTTP | Serve bundled assets. |
Deployment: Proposed fixed origin http://localhost:4173 on one shared device; choose the device, browser and local server before persistence testing. Initial loopback asset loads are expected; normal use makes no external requests.
| Id | Module | Purpose | Depends on |
|---|---|---|---|
| m1 | Local records | Schema, validation, persistence and unresolved-first query. | — |
| m2 | Handover UI | Create, edit, list and acknowledge receipt. | m1 |
| m3 | Outbound export | Download complete JSON and readable CSV without import. | m1, m2 |
After the paper gate, implement and check m1, then m2.
Modules: m1, m2
After Wave 1, implement m3 and check the integrated prototype.
Modules: m3
S1 · Save a six-field item
S2 · Retrieve unresolved items first
S3 · Create and edit from the tablet UI
S4 · Acknowledge receipt
S5 · Export JSON
S6 · Export readable CSV
Tone: Plain and non-authoritative about safety. · Colors: High-contrast dark text on a light backg · Typography: System font; tablet legibility requires prototype review.
Up to 1 builder(s) in parallel.
Shared files (serialize edits)
Rules
| Area | Check | Method |
|---|---|---|
| Paper gate | Determine whether the structured paper checklist yields complete fields and unambiguous owners and next actions. | Proposed supervised paper pilot; agree measures, duration and pass criteria before collecting outcomes. |
| Later value comparison | Determine whether software improves on the validated paper baseline. | After the paper gate and prototype, pre-agree comparable measures and evaluate both processes separately. |
| Persistence and origin | Items survive browser restart and device power cycle at http://localhost:4173. | Proposed manual check on the selected device and browser; record origin and compare stored items before and after. |
| Ordering and receipt | Unresolved items lead; acknowledgement changes receipt metadata only. | Proposed mixed-status UI check and before/after IndexedDB record comparison. |
| Export and network | JSON parses, CSV columns and escaping are correct, and normal use makes no external requests. | Proposed fixture checks and browser network inspection; allow initial loopback asset requests. |
| Dependency | Kind | Reason |
|---|---|---|
| IndexedDB | standard | Local persistence at the fixed browser origin. |
| Device-local static HTTP server | service | Provides a stable localhost application origin. |
| Risk | Mitigation |
|---|---|
| Synthetic notes do not establish user demand or ergonomic suitability. | Run the paper gate and review the prototype before adoption. |
| Changing origin or losing the single device may make records inaccessible; export cannot restore them in-app. | Fix the origin, test on the chosen device and agree an external retention and recovery policy. |
| Shared-device acknowledgements may not reliably identify a person. | Agree an initials and supervision policy outside the app. |
| Privacy and physical suitability of the tablet or paper checklist materials are unknown. | Resolve handling, access, retention and physical-use questions before any supervised pilot. |
handover-question What should the next shift know?sample-handover Three sample handover noteshandover-concept A single-device handover logFrozen scope: my canvas revision 1
No linked decision is inferred; review the retained records before treating this draft as a commitment.
No selected named concept was present in the frozen package scope.
No selected human concept decision record.
handover-question What should the next shift know? (content 1)sample-handover Three sample handover notes (content 1)handover-concept A single-device handover log (content 1)Export state: Needs review · Specification revision: 1
Specification digest: da945df18ef33c486d5a8016dfe1d97b066bcfae2cbaac190a180a66e41a432c
No human approval for handoff is recorded for this exact specification revision. Downloading or copying this export does not approve it.
Before software work, define and run a supervised structured-paper gate; assess software against paper only in a later, separate comparison.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Store the six-field handover item locally on one device at the fixed application origin, without an external service.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Present unresolved items first and allow creation, editing and receipt acknowledgement without transferring ownership.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Provide outbound JSON and readable CSV export; exclude JSON import and in-app restore.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Keep the draft build to three modules, two sequential waves and one builder; exclude telemetry and an in-app paper-comparison view.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
Contradictions, risks and missing evidence
Risks accepted for this exact handoff
None. This revision has no handoff approval.
Unanswered questions
Assumptions
Requirements without an exact decision record
Pending claims / requirements awaiting human review
Some retained historical material is unavailable. Its frozen identifiers remain visible, but this export does not restore or disclose the unavailable private content.
The generated implementation stories and their observable acceptance criteria.
Schema, validation, persistence and unresolved-first query.
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Create, edit, list and acknowledge receipt. Depends on: m1
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Download complete JSON and readable CSV without import. Depends on: m1, m2
Acceptance:
Acceptance:
The order of work, concurrency rules and checks exported with this package.
You are building Workshop Handover Log — Draft (software). Read PACKAGE.md first; it is the complete product definition. Do not widen the scope beyond it.
Frozen specification revision 1 · Needs review. Downloading this archive did not approve it. Resolve every item marked unresolved or awaiting review instead of treating it as verified.
At most 1 builder(s) at once. Serialize edits to: src/store.js, src/app.js, src/export.js.
Every story in prd/ must meet its acceptance criteria before the wave is complete. Ask a human when an open question in PACKAGE.md blocks a decision; do not guess.
Read the worker findings and the reasoning behind the proposed product.
Real model outputs from a bounded cookbook-only orchestration. All scenario inputs are synthetic.
Integrated the workers into a conditional, unbuilt software draft. The three canvas cards are the only cited inputs; the handover notes are synthetic.
Agree the paper gate and operational policies before authorizing a prototype; no files were generated, code compiled or tests executed.
The orchestrator’s synthesis, corrections and unresolved questions, preserved with the output.
These decisions are the actual final orchestrator output. They describe a light draft, not a completed build or human handoff approval.
Integrated the workers into a conditional, unbuilt software draft. The three canvas cards are the only cited inputs; the handover notes are synthetic.
Agree the paper gate and operational policies before authorizing a prototype; no files were generated, code compiled or tests executed.
Execution: cookbook-runner. Focused review calls were reconciled into an integrated result. This recipe-specific API script does not claim that the native Synatrail UI provides this model split.
This section was authored separately after the provider responses. It does not change the generated package or claim model approval.
Timing: statements about no files being generated refer to implementation artifacts at the time of the model response. The cookbook runner subsequently exported the planning documents and source files linked here. No product implementation or physical test was performed. Any later diagram rendering or CAD compilation has its own separate evidence.
The inputs are synthetic teaching material. This run used the cookbook’s reproducible API runner; its model split is not a model-routing control in the current Synatrail interface. The package uses Synatrail’s normal export format. Canvas views use the original three cards, with layout and explanatory relationships added for this walkthrough. The architecture and process diagrams retain the exported topology; card summaries and link labels were added for readability. The original records and diagrams remain available.
4 API calls · US$0.090311 estimated token cost. Exact model identifiers remain in the downloadable execution record.
These are the actual outputs of the workshop example’s light product specification. Read the product definition, inspect each module’s stories, and see the instructions a builder receives.


The screenshots show the saved output restored locally in Synatrail. The documents below are rendered directly from the downloadable files, in their original English. The specification is an unapproved draft; the separately authored review notes explain what to resolve before building.
The complete generated product definition, including scope, architecture, requirements, risks and open questions.
A single-browser, single-device handover log (IndexedDB) capturing equipment, issue, owner, next action, status, timestamp, acknowledge and JSON/CSV export—draft for pilot validation.
Kind: software · Investigation: A clearer workshop handover · Composed: 2026-10-01T17:34:26.505Z · Model details: see the original download
Who: Two-shift small workshop staff and supervising manager responsible for safe, unambiguous maintenance handovers.
Pain: Ambiguous ownership, next action and closure at shift handover causes delays and unclear responsibilities (sample synthetic notes show this ambiguity).
Context: A teaching-case proposal: single shared tablet in the workshop capturing a 6-field minimal record to reduce handover ambiguity; no cloud, auth, notifications or integrations. Software justified only after a paper-checklist pilot.
Evidence: handover-question What should the next shift know?; sample-handover Three sample handover notes; handover-concept A single-device handover log
A minimal local-browser app on one shared tablet that stores handover items in IndexedDB, presents unresolved items first, supports an acknowledge action, and exports data as JSON and human-readable CSV for supervisor review and archival.
Differentiators
Non-goals
User journey
Success criteria
Entities
| Entity | Fields | Source |
|---|---|---|
| handover_item | item_id: uuid (system-generated), equipment_label: string (required), issue_short: string (<=140 chars, required), current_owner: string (initials or name, required), next_action: string (required), status: enum [open, awaiting_review, closed] (required), created_at: ISO8601 timestamp (system), updated_at: ISO8601 timestamp (system), acknowledged_by: string (optional initials), acknowledged_at: ISO8601 timestamp (optional) | Derived from handover-concept and sample-handover |
Evidence used as data
| Evidence | Use |
|---|---|
handover-question What should the next shift know? |
Defines problem scope and validation goal |
sample-handover Three sample handover notes |
Illustrates ambiguity and minimal fields needed |
handover-concept A single-device handover log |
Specifies single-device constraint and required fields |
Single-page web app (PWA-lite not installed) running in-browser on a single shared device/origin, using IndexedDB (via Dexie.js) for persistence, vanilla JS or a compact framework (Svelte/Preact) for UI, CSV/JSON export to file, no network calls.
Components
| Component | Responsibility | Technology |
|---|---|---|
| UI Shell | Render list, item create/edit form, acknowledge action and export buttons; sort unresolved first | Svelte or Preact, plain CSS |
| Local Storage Layer | IndexedDB schema, CRUD, migrations, export serialization | Dexie.js (or direct IndexedDB API) |
| Export/Serialization | Produce JSON and human-readable CSV and trigger file download | JSON.stringify, Blob, FileSaver.js or native anchor download |
Interfaces
| From | To | Protocol | Purpose |
|---|---|---|---|
| UI Shell | Local Storage Layer | in-browser JS function calls / Promise API | CRUD operations and queries (unresolved-first sort) |
| UI Shell | Export/Serialization | in-memory data serialization | Export current dataset to JSON/CSV file |
Deployment: Serve static HTML/CSS/JS files on the shared device filesystem or local file server; open index.html in a modern Chromium-based browser that supports IndexedDB.
| Id | Module | Purpose | Depends on |
|---|---|---|---|
| m1 | Core data model & storage | Implement IndexedDB schema, migrations, and API for CRUD and queries (unresolved-first). | — |
| m2 | UI and acknowledgement flow | List view sorted unresolved-first, create/edit form, acknowledge action recording initials and timestamp. | m1 |
| m3 | Export & pilot tooling | JSON and CSV export, simple pilot-mode telemetry (local only), and paper-checklist comparison view for supervisors. | m1, m2 |
Deliver M1 and minimal M2 list/create/acknowledge UI; verify persistence and unresolved-first ordering.
Modules: m1, m2
Deliver M3 export features, pilot view and finalize acceptance checks for supervised trial.
Modules: m3
S1 · IndexedDB schema and CRUD
S2 · Unresolved-first query and migration
S3 · Create / Edit / Acknowledge UI
S4 · Local usage ergonomics
S5 · JSON export
S6 · CSV export and pilot view
Tone: Plain, high-contrast, legible for workshop lighting · Colors: High-contrast text (dark on light), stat · Typography: System UI fonts, large touch targets for tablet use
Up to 1 builder(s) in parallel.
Shared files (serialize edits)
Rules
| Area | Check | Method |
|---|---|---|
| Persistence | Records persist across browser restart and device power cycle on the same origin | Manual test: create items, close browser, restart, verify items present |
| Export | JSON and CSV exports contain all fields and are syntactically valid | Automated unit test for JSON validity and CSV header/row counts; manual open of CSV in spreadsheet |
| Ordering and ACK | UI lists unresolved items first and acknowledge records initials+timestamp | End-to-end manual verification with sample dataset (three items) and UI observation |
| Paper comparison | App fields map 1:1 to the laminated paper checklist | Supervisor-led mapping session: take 5 recent paper checklist entries and confirm mapping to in-app items |
| Dependency | Kind | Reason |
|---|---|---|
| Dexie.js | library | Simplifies IndexedDB schema/migrations and is lightweight for single-device apps |
| FileSaver.js or native Blob download | library | Triggering reliable client-side file downloads for JSON/CSV |
| Risk | Mitigation |
|---|---|
| No real users interviewed; app may not match real ergonomic needs. | Run short supervised pilot using paper checklist mapping before software pilot; keep UI minimal and iterate. |
| Single-device usage can fail if tablet is unavailable or browser origin changes. | Provide clear device usage policy and local backup via manual JSON export; test recovery via import. |
handover-question What should the next shift know?sample-handover Three sample handover noteshandover-concept A single-device handover logFrozen scope: my canvas revision 2
No linked decision is inferred; review the retained records before treating this draft as a commitment.
No selected named concept was present in the frozen package scope.
No selected human concept decision record.
handover-question What should the next shift know? (content 1)sample-handover Three sample handover notes (content 1)handover-concept A single-device handover log (content 1)Export state: Needs review · Specification revision: 1
Specification digest: 5aeb57870ef9f453f68a5acc3b21d64a8aab07a0981f76d2c37dd0a2f5ac2058
No human approval for handoff is recorded for this exact specification revision. Downloading or copying this export does not approve it.
Implement a local-only handover_item schema in IndexedDB with the six fields (equipment_label, issue_short, current_owner, next_action, status, timestamp) plus acknowledge metadata; no network calls.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
App UI must display unresolved items first and allow create/edit of the six-field model and an acknowledge action that records initials and timestamp.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
Provide JSON and human-readable CSV export of the full dataset, downloadable from the device without network usage.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
The system must be single-device and local-origin only: deploy as static files to the device and require no backend or cloud services.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
The build must include a supervisor-facing paper-checklist comparison view and instructions to run a supervised pilot before committing to broader software development.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
Provide developer controls and tests proving persistence, export correctness, unresolved-first ordering and acknowledge behavior before pilot start.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
None.
Contradictions, risks and missing evidence
Risks accepted for this exact handoff
None. This revision has no handoff approval.
Unanswered questions
Assumptions
None.
Requirements without an exact decision record
Pending claims / requirements awaiting human review
Two stories for the local data model, persistence, ordering and migration, with the acceptance criteria the model produced.
Implement IndexedDB schema, migrations, and API for CRUD and queries (unresolved-first).
Acceptance:
Acceptance:
Two stories for creating and editing items and recording the incoming shift’s acknowledgement.
List view sorted unresolved-first, create/edit form, acknowledge action recording initials and timestamp. Depends on: m1
Acceptance:
Acceptance:
Two stories for exports and pilot support. The review notes question whether all of this belongs in the first version.
JSON and CSV export, simple pilot-mode telemetry (local only), and paper-checklist comparison view for supervisors. Depends on: m1, m2
Acceptance:
Acceptance:
The generated order of work, shared-file rules and definition of done supplied with the specification.
You are building Workshop Handover: Single-Device Local Log (Draft) (software). Read PACKAGE.md first; it is the complete product definition. Do not widen the scope beyond it.
Frozen specification revision 1 · Needs review. Downloading this archive did not approve it. Resolve every item marked unresolved or awaiting review instead of treating it as verified.
At most 1 builder(s) at once. Serialize edits to: schema/indexeddb-schema.json, data-model/handover_item.json, storage/migrations.js.
Every story in prd/ must meet its acceptance criteria before the wave is complete. Ask a human when an open question in PACKAGE.md blocks a decision; do not guess.
Separately authored review notes: the scope and acceptance conflicts to resolve before handing the draft to a builder.
The linked package is the real model output, preserved as an unapproved draft. These are editorial review notes; they are not extra provider findings or implemented changes.
These observations concern the recorded 2026-10-01 example. On a new reproduction, check each observation against the new output before using it. The companion archive adds this note as COOKBOOK-REVIEW.md; it does not alter the seven generated package files or approve the specification.
Suggested next steering prompt:
Revise this draft only. Keep one builder and make both waves sequential. Define acknowledgement as recording receipt, without changing the owner. Remove telemetry and the comparison view from the first version. Resolve the optional-import acceptance conflict explicitly. Separate the paper-process gate from a later software-versus-paper comparison. Keep the three original source cards, preserve every unresolved assumption, and do not start a build or a commissioned package run.
This revision prompt is an example for the reader. It was not submitted as another paid call in the recorded run.
Synatrail automatically adds a scoped answer to the canvas. For this example its answer card was removed with the normal graph edit, retaining the full saved conversation and receipt. The final investigation has the original three nodes. The first scope review saw that temporary answer card; the package itself used only the three selected input cards plus the saved answer. The reproducible script now removes the answer card immediately after the answer is saved.
An adjustable inspection fixture adds geometry and motion to the design problem. Focused reviews examine guide rods, jaws, a screw and assembly checks, then reconcile the findings into one light package.
Three nodes: the holding problem, a dimensional contract and the proposed manual fixture. The 240 × 180 mm base supports two guide rods, fixed and moving jaws, pads, bushings and a hand-adjusted screw.
Steer the work: make workers check the interfaces between parts. The workers found that moving the jaw 60 mm in the opening direction would hit the end support. The review preserved the conflict as an unresolved design decision instead of silently changing the dimensions.
The resulting light specification: scope, requirements, architecture, risks and the decisions still needing evidence.
A proposed manually adjusted fixture for positioning a de-energized sample during bench inspection.
Kind: hardware · Investigation: Adjustable bench inspection fixture · Composed: 2026-10-01T21:14:55.266Z · Model details: see the original download
Who: A person conducting a proposed bench inspection.
Pain: A small sample may need positioning and visual access; repeatability and holding performance have not been established.
Context: Fictional teaching case. The supplied mechanical parameters are design constraints, not measured geometry or evidence of a working fixture.
Evidence: fixture-question Hold the sample for inspection; fixture-dimensions Fixed dimensions and interfaces; fixture-concept Three modules, twenty proposed parts
Review the locked multipart fixture as structural supports, manual adjustment interfaces, and assembly with proposed checks. A separate deterministic CAD builder may later construct and measure geometry.
Differentiators
Non-goals
User journey
Success criteria
Entities
| Entity | Fields | Source |
|---|---|---|
| ParameterContract | schema, coordinateSystem, parameters, limitations | Supplied mechanicalDesign; locked input, not compiled CAD |
| InterfaceFinding | subassembly, supplied envelope, inferred overlap, unresolved interface, proposed later check | Reasoning from fixture-dimensions and the supplied contract |
| PartProposal | part name, proposed quantity, specified dimensions, allocation uncertainty | fixture-concept and supplied contract; bushing allocation remains inferred |
Evidence used as data
| Evidence | Use |
|---|---|
fixture-question Hold the sample for inspection |
Bound the fictional use case and prohibit performance claims. |
fixture-dimensions Fixed dimensions and interfaces |
Identify supplied coordinates and proposed opening, travel and clearance. |
fixture-concept Three modules, twenty proposed parts |
Bound modules and the proposed multipart count. |
A mechanical planning package, not an implemented device or CAD assembly. The locked contract governs later deterministic CAD work; this review records conflicts without modifying it.
Components
| Component | Responsibility | Technology |
|---|---|---|
| Structural group | Base, feet, supports, rods, jaws, pads and proposed guide interfaces. | Mechanical parts; CAD source deferred |
| Manual adjustment group | Proposed lead screw, nut and hand-knob interfaces. | Manual mechanical mechanism; thread and retention unspecified |
| Review record | Assembly order, interface questions and proposed CAD and bench checks. | Versioned planning text; no deployed software |
Interfaces
| From | To | Protocol | Purpose |
|---|---|---|---|
| Structural group | Manual adjustment group | Physical screw, nut and jaw interface, unresolved | Identify passages, restraint and retention without assuming they exist. |
| Review record | Separate CAD builder | Locked mechanicalDesign JSON | Supply unchanged parameters for later construction and geometric review. |
Deployment: None. CAD/source generation, geometric measurements, fabrication and physical checks are separate later operations.
| Id | Module | Purpose | Depends on |
|---|---|---|---|
| m1 | Structural and support parts | Bound the base, supports, feet, guides, jaws, pads and bushing interfaces. | — |
| m2 | Manual jaw adjustment and interfaces | Review screw, nut, knob, opening and travel without asserting operability. | m1 |
| m3 | Assembly and proposed bench checks | Reconcile named parts and plan conditional CAD and non-destructive checks. | m1, m2 |
Document supplied envelopes and unresolved support, guide and jaw interfaces.
Modules: m1
Review M2 before M3; stop at conflicts and do not infer a working assembly.
Modules: m2, m3
S1 · Map base, feet and end supports
S2 · Map rods, jaws, pads and bushings
S3 · Review screw, nut and knob
S4 · Review opening and travel
S5 · Reconcile parts and assembly order
S6 · Plan conditional checks
| Part | Qty | Specification | Est. cost | Supplier hint |
|---|---|---|---|---|
| Base | 1 | 240 × 180 × 10 mm | Not estimated | Material and fabrication unspecified |
| Guide rod | 2 | Ø10 × 200 mm; proposed axes Y45/135, Z32 | Not estimated | Material and sourcing unspecified |
| End support | 2 | 16 × 140 × 32 mm | Not estimated | Passages and attachment unspecified |
| Jaw | 2 | Fixed and moving; each 18 × 120 × 43 mm | Not estimated | Guidance and attachments unspecified |
| Pad | 2 | Each 3 × 100 × 30 mm; inward placement assumed for stated opening | Not estimated | Material and attachment unspecified |
| Bushing | 4 | Inferred quantity, not locked; Ø16 OD, Ø10.4 bore, 18 mm length | Not estimated | Allocation, fit and retention unresolved |
| Lead screw | 1 | Ø8 nominal × 220 mm; thread unspecified | Not estimated | Bearing and axial retention unresolved |
| Nut | 1 | Ø18 OD, Ø8.4 nominal bore, 14 mm length; thread unspecified | Not estimated | Engagement and anti-rotation unresolved |
| Hand knob | 1 | Ø32 × 10 mm | Not estimated | Screw coupling unresolved |
| Foot | 4 | Ø22 × 12 mm | Not estimated | Attachment and vertical placement unresolved |
Each part below gets parametric source, STEP/STL exports and dimensioned drawings in cad/<part>/ once its geometry is generated; the runner checks the measured envelope and mass against these values.
| Id | Part | Kind | Envelope (mm) | Material | Mass (g) | Tool | Module | BOM part |
|---|---|---|---|---|---|---|---|---|
| p1 | Base | mechanical | 240 × 180 × 10 | Unspecified | auto | m1 | Base | |
| p2 | Guide rod | mechanical | 200 × 10 × 10 | Unspecified | auto | m1 | Guide rod | |
| p3 | End support | mechanical | 16 × 140 × 32 | Unspecified | auto | m1 | End support | |
| p4 | Fixed jaw | mechanical | 18 × 120 × 43 | Unspecified | auto | m1 | Jaw | |
| p5 | Moving jaw | mechanical | 18 × 120 × 43 | Unspecified | auto | m1 | Jaw | |
| p6 | Pad | mechanical | 3 × 100 × 30 | Unspecified | auto | m1 | Pad | |
| p7 | Bushing | mechanical | 18 × 16 × 16 | Unspecified | auto | m1 | Bushing | |
| p8 | Lead screw | mechanical | 220 × 8 × 8 | Unspecified | auto | m2 | Lead screw | |
| p9 | Nut | mechanical | 14 × 18 × 18 | Unspecified | auto | m2 | Nut | |
| p10 | Hand knob | mechanical | 10 × 32 × 32 | Unspecified | auto | m2 | Hand knob | |
| p11 | Foot | mechanical | 22 × 22 × 12 | Unspecified | auto | m1 | Foot |
Tone: Specific, conditional and non-certifying. · Colors: unspecified · Typography: Plain labels and millimetre dimensions.
Up to 1 builder(s) in parallel.
Shared files (serialize edits)
Rules
| Area | Check | Method |
|---|---|---|
| Contract integrity | Compare the later builder input with the supplied JSON. | Exact structural JSON comparison; no dimension edits. |
| Opening | Review pad orientation and nominal gap. | Calculate 142 − (40 + 18) − 2 × 3 = 78 mm, conditional on inward pads. |
| Travel | Review both endpoints against supports and rods. | Compare X82..100 and X202..220 with support X212..228 and rod end X220. |
| Interfaces | Review passages, fits, thread and retention. | Mark each absent definition unresolved; defer CAD geometry review. |
| Future bench checks | Propose visual access and unloaded slow adjustment only after conflicts are resolved. | Planning only; stop at suspected interference; record no physical result. |
| Dependency | Kind | Reason |
|---|---|---|
| Supplied mechanicalDesign JSON | data | Locked parameter input for a separate deterministic CAD builder. |
| Separate deterministic CAD builder | service | Later source generation and geometric measurements; not executed in this review. |
| Risk | Mitigation |
|---|---|
| Opening-direction travel intersects the right end support. | Flag the 8 mm envelope overlap; do not call 60 mm usable or change coordinates. |
| Rod and screw crossings lack specified passages and retention. | Keep interfaces unresolved pending separate CAD review and design decisions. |
| Twenty named instances may conceal unspecified attachment hardware. | Treat four bushings as inferred and the count as provisional, not a complete assembly specification. |
fixture-question Hold the sample for inspectionfixture-dimensions Fixed dimensions and interfacesfixture-concept Three modules, twenty proposed partsFrozen scope: my canvas revision 1
No linked decision is inferred; review the retained records before treating this draft as a commitment.
No selected named concept was present in the frozen package scope.
No selected human concept decision record.
fixture-question Hold the sample for inspection (content 1)fixture-dimensions Fixed dimensions and interfaces (content 1)fixture-concept Three modules, twenty proposed parts (content 1)Export state: Needs review · Specification revision: 1
Specification digest: 484940da9d063d9e464725c080373a09824f12502de03db3496321491c07c0e7
No human approval for handoff is recorded for this exact specification revision. Downloading or copying this export does not approve it.
Preserve the supplied mechanicalDesign as a locked, proposed parameter contract.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
Document conditional opening and the proposed travel conflict.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Keep guidance and screw interfaces explicitly unresolved.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Reconcile the named-part BOM without treating the count as proof of completeness.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Sequence review and proposed checks without implying execution.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
None.
Unresolved
Contradictions, risks and missing evidence
Risks accepted for this exact handoff
None. This revision has no handoff approval.
Unanswered questions
Assumptions
Requirements without an exact decision record
Pending claims / requirements awaiting human review
Some retained historical material is unavailable. Its frozen identifiers remain visible, but this export does not restore or disclose the unavailable private content.
The generated implementation stories and their observable acceptance criteria.
Bound the base, supports, feet, guides, jaws, pads and bushing interfaces.
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Review screw, nut, knob, opening and travel without asserting operability. Depends on: m1
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Reconcile named parts and plan conditional CAD and non-destructive checks. Depends on: m1, m2
Acceptance:
Acceptance:
The order of work, concurrency rules and checks exported with this package.
You are building Adjustable bench inspection fixture — LIGHT review (hardware). Read PACKAGE.md first; it is the complete product definition. Do not widen the scope beyond it.
Frozen specification revision 1 · Needs review. Downloading this archive did not approve it. Resolve every item marked unresolved or awaiting review instead of treating it as verified.
At most 1 builder(s) at once. Serialize edits to: mechanicalDesign JSON contract.
If generated artifacts are present, cad/<part>/ holds their parametric source and exported views. The dimensions and mass in PACKAGE.md are design requirements, not measurements or proof of manufacture. Inspect each generated result and its checks before use.
Every story in prd/ must meet its acceptance criteria before the wave is complete. Ask a human when an open question in PACKAGE.md blocks a decision; do not guess.
Read the worker findings and the reasoning behind the proposed product.
Actual worker outputs; supplied dimensions are design constraints, not measurements.
Integrated both worker reviews as one bounded LIGHT proposal. Their coordinate findings agree; neither review establishes a working mechanism.
The orchestrator’s synthesis, corrections and unresolved questions, preserved with the output.
Actual model decisions for a light draft. No human approval or executed product test is implied.
Integrated both worker reviews as one bounded LIGHT proposal. Their coordinate findings agree; neither review establishes a working mechanism.
Obtain explicit interface and travel decisions before separate CAD construction and geometric review; preserve the supplied contract until any proposed revision is reviewed openly.
The runner exported these planning files after the model response. Later CAD compilation, measured geometry and rendering are separate operations with their own evidence. They are not physical testing.
The CAD author’s additional design choices, reproduction steps and limits of the recorded geometry checks.
This is actual parametric CAD compiled from a supplied, provider-reviewed mechanical contract. The API models reviewed the contract and plan. Codex authored the separate Python CAD implementation; no claim is made that the models emitted this source.
The assembly contains 20 meaningful component instances and 0 explicitly counted fastener envelopes, representing 10 distinct designs and 4 functional subassemblies. Counts come from the same inventory exported to STEP, GLB and BOM.
Validation covers the static shown pose. Digital checks are not a motion study, tolerance stack, physical prototype, structural assessment or certification. Motors, bearings, optics and connectors are nominal envelopes. Screw threads and internal bearing mechanisms are simplified. Exploded offsets are presentation transforms, not a verified assembly path.
Reproduce from the repository: node scripts/cookbook-mechanical-render.mjs --only=inspection-fixture. It uses the existing offline ce-spec-native:spec-final-2 Docker image with build123d 0.11.1, Python 3 with NumPy/Pillow, and sharp. The container sees only an isolated scratch directory and has no network access. There are no API calls or commissioned specification jobs.
5 of 5 selected clearance/interference checks pass in the shown static pose. The separate valid-solid and STEP round-trip checks pass. These are computational results, not physical testing.
Follow the specification’s proposed part references to the actual compiled components and identify what remains unresolved.
This crosswalk was authored after the model review and CAD compilation. The model proposal IDs and the CAD instance IDs are separate namespaces. Mapping a part does not mean that the model approved its later construction details.
Sources: actual orchestrator review, mechanical contract, compiled inventory, CAD implementation and checks, machine-readable crosswalk and bound checks.
The inventory contains 20 component instances and 0 fastener envelopes, with 10 distinct non-fastener designs and 4 functional subassemblies. These physical groups sit inside the three LIGHT planning modules.
| Model proposal | Meaning | CAD design ID | Actual CAD instances |
|---|---|---|---|
| BOM | Base | fixture-base |
F01 |
| BOM | Foot | fixture-foot |
F02_1, F02_2, F02_3, F02_4 |
| BOM | Guide rod | guide-rod-10x200 |
F03_1, F03_2 |
| BOM | End support | fixture-end-support |
F04_1, F04_2 |
| BOM | Jaw | fixture-jaw |
F05_1, F05_2 |
| BOM | Bushing | fixture-bushing |
F06_1_1, F06_1_2, F06_2_1, F06_2_2 |
| BOM | Pad | fixture-jaw-pad |
F07_1, F07_2 |
| BOM | Lead screw | fixture-screw |
F08 |
| BOM | Nut | fixture-travel-nut |
F09 |
| BOM | Hand knob | fixture-handwheel |
F10 |
Unresolved in the CAD author’s own manifest:
Feet extend below the base-origin plane as an explicit authored placement. The model-reviewed outer dimensions remain traceable; added features and extra designs are identified separately.
Independent inventory comparison: 34 selected bounds, positions and counts agree within 0.0001 mm numeric tolerance. This checks the compiled inventory against the supplied values; it does not repeat the CAD kernel or prove that every interface, bore, wall, stroke or attachment is correct.
The 78 mm pad gap is the compiled static pose. A 60 mm opening stroke would conflict with the right support. Closing-direction travel remains a hypothesis, and no motion simulation or retention test is claimed.
The inputs are synthetic teaching material. This run used the cookbook’s reproducible API runner; its model split is not a model-routing control in the current Synatrail interface. The package uses Synatrail’s normal export format. Canvas views use the original three cards, with layout and explanatory relationships added for this walkthrough. The architecture and process diagrams retain the exported topology; card summaries and link labels were added for readability. The original records and diagrams remain available.
4 API calls · US$0.099689 estimated token cost. Exact model identifiers remain in the downloadable execution record.
The advanced example is a benchtop XYZ inspection workstation: a moving gantry, focus mechanism, camera and ring-light envelopes, workholding deck, controller housing and cable carrier. Focused reviews investigate mechanical and service interfaces before an integrated design review.
Three nodes: the inspection-to-handoff question, the locked dimensional brief and the integrated concept. The 540 × 420 mm base anchors a much larger design, while the planning package remains three modules and six stories.
Steer the work: assign one worker the physical interfaces and another the local web record and human handoff. Require agreement on setup references and staff responsibilities. A changed or mismatched setup must stop record confirmation; the proposed web application does not control the mechanism.
Counts come from the compiled CAD inventory: 85 component instances and 8 fastener instances. Repeated parts are counted separately from distinct designs.
Drag to rotate, scroll or pinch to zoom, and select a part to read its name. The rotation buttons and controls also work with a keyboard.
The drawing and CAD downloads are available below.
Loading the assembly and part inventory…3D assembly loaded. Choose a subassembly or separate the parts to inspect it.Interactive 3D is unavailable here. You can still open the drawings and download the CAD files.Interactive inspection-station CAD assembly. Use the controls below to rotate, separate and isolate parts.These orthographic sheets project the saved CAD solids and label nominal millimetres. The base plate, open gantry bridge, controller shell and vented lid share part IDs with the full assembly. Material, tolerance, loads and manufacturing still need engineering review.
540 × 420 × 12 mm overall. 4 × Ø8.6 mm through holes.
440 × 60 × 30 mm overall. Open at both X ends.
155 × 100 × 58 mm overall. 149 × 94 mm open cavity.
155 × 100 × 3 mm overall. 8 × 4 × 55 mm through slots.
The individual STL files are meshes; the named STEP assembly and Python source contain the editable solid design. The source checks confirm dimensions and geometry bounds, not physical fit or fabrication readiness.
The workstation study now includes a separate 5 V status indicator concept. Its two dry-contact inputs show whether contact A or B is closed; a third LED shows power. It does not drive motors, control the camera or make inspection decisions. Electrical mounting, wiring, environmental protection and physical behavior still need review.
The KiCad exports are inspectable design files, not approval to manufacture or connect this board to the workstation. The controller housing is a nominal CAD envelope; a board mounting interface and electrical integration have not been validated. Compare the controller part drawings ↑
The mechanical package contains compiled solid geometry, part identities, drawings and recorded geometry checks. The separate electrical files describe a status-only board concept. The CAD source was authored from the dimensional contract reviewed by the model team. It is a concept assembly; purchased components are nominal envelopes, and powered operation, optics, loads and manufacturing have not been validated.
A design issue you can inspect: the focus screw intersects the camera housing in the supplied layout. The locked positions are preserved, and the measured interference is recorded in the CAD notes for the next design iteration.
The resulting light specification: scope, requirements, architecture, risks and the decisions still needing evidence.
One unchanged mechanical contract, one proposed setup manifest and one staff-controlled local inspection handoff.
Kind: hybrid · Investigation: Benchtop XYZ inspection workstation · Composed: 2026-10-01T21:19:57.622Z · Model details: see the original download
Who: Setup operators, staff reviewers and receiving service staff in a fictional teaching case.
Pain: Mechanical setup declarations, inspection records and handoffs can disagree or become stale.
Context: Planning only. Supplied dimensions are constraints, not measurements. CAD construction, software implementation and validation are separate later operations.
Evidence: workstation-question One setup, one inspection handoff; workstation-dimensions Locked workstation geometry; workstation-concept Mechanical interfaces meet the staff record
Preserve the complete supplied mechanicalDesign and organize its interfaces into two mechanical modules and one local record module. A shared identifier manifest links staff declarations to the unchanged contract without asserting physical verification.
Differentiators
Non-goals
User journey
Success criteria
Entities
| Entity | Fields | Source |
|---|---|---|
| ContractReference | contractRef: C-LOCKED-1, schema: tce-cookbook-mechanical-design-v1, product: inspection-workstation, sourceNode: workstation-dimensions, sourceContentRevision: 1, payload: unchanged supplied mechanicalDesign | workstation-dimensions; C-LOCKED-1 is a proposed local snapshot identity, not a new geometry revision. |
| MechanicalCrosswalk | SA01 base/feet: P01 base, P23 foot, SA02 Y guidance/carriages: P02 rail, P03 carriage, P04 saddle, SA03 gantry: P05 tower, P06 bridge, SA04 X guidance/carriage/screw: P07 rail, P08 carriage, P09 screw, SA05 Z guidance/screw: P10 backplate, P11 rod, P12 screw, SA06 optics envelopes: P13 camera body, P14 lens, P15 ring light, SA07 deck/workholding: P16 deck, P17 standoff; fixture hardware undefined, SA08 enclosure: P18 rear panel, P19 side panel, SA09 controller housing: P20 housing, P21 lid, SA10 cable carrier: P22 link | Proposed worker-a crosswalk integrated against workstation-dimensions; owned by the mechanical reviewer. |
| SetupManifest | manifestRef: M-SETUP-r1; manifestRevision: 1, contractRef: C-LOCKED-1; sourceContentRevision: 1, partRefs: P01–P23; subassemblyRefs: SA01–SA10, fixtureRef: FX-DECK-GRID-r1, a reference to P16 grid only, setupRef: synthetic identifier, declaredState: manual X, gantry Y and focus Z descriptions, engineeringFlags: endpoints, mounts, optics, access, carrier, loads, electrical | Proposed integration of workstation-dimensions and workstation-concept; not physical setup evidence. |
| InspectionRecord | inspectionId: synthetic identifier, setupRef; fixtureRef; contractRef; sourceContentRevision, manifestRef; manifestRevision, recordRevision: increasing integer, checks: code, outcome, responsibleRole, discrepancies: controlled synthetic codes, imageRefs: optional synthetic reference IDs, inspectionState: draft | ready | confirmed | review-required, handoffStatus: none | offered | acknowledged, declaredState; engineeringFlags; receivingRoleAcknowledgement | workstation-concept; proposed staff module schema with no names or customer fields. |
| ImageReference | imageId: IMG-ASM | IMG-EXP | IMG-SEC, owner: future separate geometry builder, status: pending | available-synthetic, sourceRef: absent until a synthetic source is registered, contractRef; manifestRef; manifestRevision | Proposed future references from worker-a; no images or CAD outputs currently exist. |
Evidence used as data
| Evidence | Use |
|---|---|
workstation-question One setup, one inspection handoff |
Revision 1; fictional scope, staff ownership and exclusion of energized operation. |
workstation-dimensions Locked workstation geometry |
Revision 1; supplied constraints and unchanged separate mechanicalDesign contract. |
workstation-concept Mechanical interfaces meet the staff record |
Revision 1; candidate manifest, local record and human handoff workflow. |
Proposed architecture: a read-only contract and mechanical crosswalk feed a versioned setup manifest. A local browser record compares selected references with the latest IndexedDB record inside a transaction before a staff-requested transition. There is no connection to machinery. Geometry below summarizes selected supplied parts only; mechanicalDesign remains authoritative. Materials, masses and unspecified envelopes are intentionally not invented.
Components
| Component | Responsibility | Technology |
|---|---|---|
| Contract and interface register | Preserve the supplied JSON, own P01–P23 and SA01–SA10, and list unresolved mechanical interfaces. | Read-only JSON and Markdown review documents |
| Setup manifest | Bind contract, fixture/grid reference, declared setup and engineering flags under M-SETUP-r1. | Versioned local JSON |
| Staff record demonstration | Present role-controlled checks, revision comparison, record states and handoff receipt. | Proposed TypeScript and semantic HTML; not implemented |
| Local persistence and export | Store synthetic metadata and explicitly export a point-in-time JSON copy. | Proposed IndexedDB transactions and browser Blob download |
Interfaces
| From | To | Protocol | Purpose |
|---|---|---|---|
| Contract and interface register | Setup manifest | Local JSON references | Reuse the unchanged contract identity and shared part/subassembly IDs; retain unresolved engineering flags. |
| Setup manifest | Staff record demonstration | Local read-only manifest selection | Compare contract, manifest, setup and fixture references before a staff-requested confirmation. |
| Staff record demonstration | Local persistence and export | IndexedDB readwrite transaction; explicit JSON download | Reject stale record revisions and persist accepted record changes; export a separate, potentially stale copy. |
Deployment: Future local-only browser demonstration in one browser/profile. This package deploys nothing. Browser clearing can erase records; downloaded exports are not authoritative live records.
| Id | Module | Purpose | Depends on |
|---|---|---|---|
| m1 | Base and motion structure | Review SA01–SA05: base/feet, Y guidance, gantry, X guidance/screw and Z focus guidance/screw. | — |
| m2 | Workholding/optics/enclosure interfaces | Review SA06–SA10: optics, deck/workholding, enclosure, controller housing and cable carrier. | m1 |
| m3 | Local staff inspection record and service handoff | Specify synthetic local metadata, reference checks, explicit staff transitions and receiving-role acknowledgement. | m1, m2 |
Review M1 then M2, publish the shared manifest and preserve all unresolved mechanical questions without CAD generation.
Modules: m1, m2
Specify M3 against the reviewed reference baseline and prepare proposed synthetic checks, without implementation.
Modules: m3
M1-S1 · Review assembly interfaces against the locked contract
M1-S2 · Review declared manual positioning and clearance questions
M2-S1 · Review fixture/grid and optics references
M2-S2 · Review enclosure, cable and service interfaces
M3-S1 · Specify staff setup-reference confirmation
M3-S2 · Specify staff handoff and receipt
| Part | Qty | Specification | Est. cost | Supplier hint |
|---|---|---|---|---|
| P01 base | 1 | 540 × 420 × 12 mm; material and mass unspecified. | Not estimated | Not selected |
| P02 Y rail | 2 | 360 × 16 × 12 mm nominal envelope; not vendor-qualified. | Not estimated | Not selected |
| P06 bridge | 1 | 440 × 60 × 30 mm; wall 4 mm. | Not estimated | Not selected |
| P07 X rail | 1 | 410 × 16 × 16 mm nominal envelope. | Not estimated | Not selected |
| P08 X carriage | 1 | 60 × 24 × 50 mm nominal envelope. | Not estimated | Not selected |
| P09 X screw | 1 | Diameter 8 mm; length and drive connection unspecified. | Not estimated | Not selected |
| P10 focus backplate | 1 | 70 × 8 × 180 mm; attachment and relief unresolved. | Not estimated | Not selected |
| P11 focus rod | 2 | Diameter 8 mm; length 120 mm. | Not estimated | Not selected |
| P12 focus screw | 1 | Diameter 6 mm; length and nut interface unspecified. | Not estimated | Not selected |
| P13 camera-body envelope | 1 | 32 × 32 × 40 mm; nominal geometry only. | Not estimated | Not selected |
| P14 lens envelope | 1 | Diameter 20 mm; length 22 mm; placement interface unresolved. | Not estimated | Not selected |
| P15 ring-light envelope | 1 | OD 48 mm; ID 26 mm; thickness 10 mm. | Not estimated | Not selected |
| P16 deck | 1 | 260 × 200 × 8 mm; grid pitch 25 mm; hole diameter 5 mm. | Not estimated | Not selected |
| P17 standoff | 4 | Diameter 20 mm; height 44 mm; attachment unspecified. | Not estimated | Not selected |
| P18 rear panel | 1 | 500 × 4 × 360 mm. | Not estimated | Not selected |
| P20 controller housing | 1 | Outer size 155 × 100 × 58 mm; wall 3 mm; eight vent features, not separate parts. | Not estimated | Not selected |
| P21 controller lid | 1 | Thickness 3 mm; footprint, attachment and fit unresolved. | Not estimated | Not selected |
| P22 carrier link | 12 | Pitch 18 mm; width 22 mm; height 16 mm; full link envelope and joints unspecified. | Not estimated | Not selected |
| P23 foot | 4 | Diameter 36 mm; height 24 mm; Z placement and mounting unresolved. | Not estimated | Not selected |
Each part below gets parametric source, STEP/STL exports and dimensioned drawings in cad/<part>/ once its geometry is generated; the runner checks the measured envelope and mass against these values.
| Id | Part | Kind | Envelope (mm) | Material | Mass (g) | Tool | Module | BOM part |
|---|---|---|---|---|---|---|---|---|
| p01 | Base | mechanical | 540 × 420 × 12 | Unspecified; mass unknown | auto | m1 | P01 base | |
| p02 | Y rail envelope | mechanical | 16 × 360 × 12 | Unspecified; mass unknown | auto | m1 | P02 Y rail | |
| p06 | Gantry bridge | mechanical | 440 × 60 × 30 | Unspecified; mass unknown | auto | m1 | P06 bridge | |
| p08 | X carriage envelope | mechanical | 60 × 24 × 50 | Unspecified; mass unknown | auto | m1 | P08 X carriage | |
| p10 | Focus backplate | mechanical | 70 × 8 × 180 | Unspecified; mass unknown | auto | m1 | P10 focus backplate | |
| p13 | Camera-body envelope | electronic | 32 × 32 × 40 | Unspecified; mass unknown | auto | m2 | P13 camera-body envelope | |
| p16 | Perforated work deck | mechanical | 260 × 200 × 8 | Unspecified; mass unknown | auto | m2 | P16 deck | |
| p18 | Rear enclosure panel | mechanical | 500 × 4 × 360 | Unspecified; mass unknown | auto | m2 | P18 rear panel | |
| p20 | Controller housing | mechanical | 155 × 100 × 58 | Unspecified; mass unknown | auto | m2 | P20 controller housing |
Tone: Plain, cautious and specific; never present a record state as permission to operate. · Colors: #172033 text, #FFFFFF background, #92400E review warning, #B91C1C reference blocker, #1D4ED8 record action · Typography: System sans-serif for prose; monospace for IDs and revisions.
Up to 1 builder(s) in parallel.
Shared files (serialize edits)
Rules
| Area | Check | Method |
|---|---|---|
| Contract preservation | Proposed check: all values, array order, units, origin and limitations match the supplied JSON. | Later recursive JSON comparison against the supplied payload; report every difference. |
| Mechanical interfaces | Proposed check: document conditional optics overlaps, X attachment questions and undefined motion endpoints. | Review supplied-coordinate arithmetic now; request later identified sections and endpoint checks. |
| Geometry evidence | Future targets: 60–100 placed components, at least 25 distinct non-fastener designs and at least 8 subassemblies. | Separate builder reports actual counts, bounds, assembly, exploded and section views; no result asserted. |
| Reference confirmation | Proposed cases: matching references, changed fixture, changed manifest, stale view and missing listed image. | Specify synthetic inputs and expected record-only outcomes; do not report execution or passing results. |
| Revision integrity | Proposed case: two views request transitions from the same record revision; the stale request is rejected. | Later exercise transactional latest-revision comparison with two synthetic browser views. |
| Handoff and adjustment | Proposed cases: incomplete offer, unacknowledged limitations and a setup change after acknowledgement. | Walk the role/state table; expect blocked receipt or renewed review, never physical action. |
| Local-only scope | Proposed review: metadata only, no network uploads, device APIs, real images or identity claims. | Review future schema, browser API use and export fields against the exclusions. |
| Prototype evidence boundary | Loads, fits, protection, electrical integration and prototype validation remain unresolved. | Maintain an open review register; do not substitute counts, renderings or record states for validation. |
| Dependency | Kind | Reason |
|---|---|---|
| Supplied mechanicalDesign and workstation-dimensions revision 1 | data | Authoritative locked geometry input, not measurement evidence. |
| workstation-question and workstation-concept revision 1 | data | Define synthetic scope, staff ownership and candidate local workflow. |
| IndexedDB | standard | Proposed browser-local transactional metadata storage; no implementation supplied. |
| TypeScript | library | Proposed language/toolchain for a later bounded browser demonstration. |
| JSON | standard | Shared manifest and explicit local export format. |
| Risk | Mitigation |
|---|---|
| Conditional camera-body and ring-light intersections with the unrelieved backplate. | Preserve geometry and flag conditional envelope conflicts; request explicit mounting interpretation and later sections. |
| Undefined engagement, drive connections, endpoints and attachment faces undermine motion conclusions. | Do not infer usable travel from rail lengths or nominal travel; require separately reviewed definitions. |
| A confirmed record or acknowledged handoff is mistaken for physical approval. | Use record-only labels and retain engineering limitations in the record and handoff. |
| Local persistence is cleared, or a downloaded copy is mistaken for the latest record. | Explain browser/profile limits, include revisions in exports and provide no import or cloud-sync path in this scope. |
| Complexity targets encourage unsupported parts or inflated design counts. | Count only later evidenced geometry; do not treat identifiers, vent features or repeated instances as distinct designs. |
| The controller, panels and carrier appear integrated despite undefined wiring, access and protection. | Keep electrical integration, loads, routing, guarding and prototype validation explicitly unresolved. |
workstation-question One setup, one inspection handoffworkstation-dimensions Locked workstation geometryworkstation-concept Mechanical interfaces meet the staff recordFrozen scope: my canvas revision 1
No linked decision is inferred; review the retained records before treating this draft as a commitment.
No selected named concept was present in the frozen package scope.
No selected human concept decision record.
workstation-question One setup, one inspection handoff (content 1)workstation-dimensions Locked workstation geometry (content 1)workstation-concept Mechanical interfaces meet the staff record (content 1)Export state: Needs review · Specification revision: 1
Specification digest: 5ecf6b810ad05d603a2d2f98308022779915a9b0f497c13ddfe7b3b8b50293ff
No human approval for handoff is recorded for this exact specification revision. Downloading or copying this export does not approve it.
Preserve the supplied mechanical contract and the bounded investigation scope.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Use one proposed mechanical crosswalk and an honest partial BOM.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Record mechanical conflicts without changing geometry or asserting validation.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Specify minimal local synthetic storage and record-only confirmation gates.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Define staff-controlled inspection and handoff transitions.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Keep pending image references and proposed checks distinct from evidence.
Acceptance
Decision records
No exact human decision record.
Evidence versions
Assumptions
Unresolved
Contradictions, risks and missing evidence
Risks accepted for this exact handoff
None. This revision has no handoff approval.
Unanswered questions
Assumptions
Requirements without an exact decision record
Pending claims / requirements awaiting human review
Some retained historical material is unavailable. Its frozen identifiers remain visible, but this export does not restore or disclose the unavailable private content.
The generated implementation stories and their observable acceptance criteria.
Review SA01–SA05: base/feet, Y guidance, gantry, X guidance/screw and Z focus guidance/screw.
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Review SA06–SA10: optics, deck/workholding, enclosure, controller housing and cable carrier. Depends on: m1
Acceptance:
Acceptance:
The generated implementation stories and their observable acceptance criteria.
Specify synthetic local metadata, reference checks, explicit staff transitions and receiving-role acknowledgement. Depends on: m1, m2
Acceptance:
Acceptance:
The order of work, concurrency rules and checks exported with this package.
You are building Benchtop XYZ inspection workstation — LIGHT (hybrid). Read PACKAGE.md first; it is the complete product definition. Do not widen the scope beyond it.
Frozen specification revision 1 · Needs review. Downloading this archive did not approve it. Resolve every item marked unresolved or awaiting review instead of treating it as verified.
At most 1 builder(s) at once. Serialize edits to: mechanicalDesign.json, mechanical-crosswalk.json, setup-manifest.json, record-contract.md, review-register.md.
If generated artifacts are present, cad/<part>/ holds their parametric source and exported views. The dimensions and mass in PACKAGE.md are design requirements, not measurements or proof of manufacture. Inspect each generated result and its checks before use.
Every story in prd/ must meet its acceptance criteria before the wave is complete. Ask a human when an open question in PACKAGE.md blocks a decision; do not guess.
Read the worker findings and the reasoning behind the proposed product.
Actual worker outputs; supplied dimensions are design constraints, not measurements.
Integrated the two actual reviews into one bounded LIGHT package. The supplied mechanicalDesign is retained unchanged as a proposed parameter contract, not compiled CAD. No check, geometry result, physical test or application implementation is claimed.
The orchestrator’s synthesis, corrections and unresolved questions, preserved with the output.
Actual model decisions for a light draft. No human approval or executed product test is implied.
Integrated the two actual reviews into one bounded LIGHT package. The supplied mechanicalDesign is retained unchanged as a proposed parameter contract, not compiled CAD. No check, geometry result, physical test or application implementation is claimed.
Have one reviewer execute Wave 1's document review, confirm the crosswalk and record unresolved interfaces without changing the contract; then specify Wave 2's record transitions. Separately authorize CAD/source generation and validation only after missing definitions receive explicit review.
The runner exported these planning files after the model response. Later CAD compilation, measured geometry and rendering are separate operations with their own evidence. They are not physical testing.
The CAD author’s additional design choices, reproduction steps and limits of the recorded geometry checks.
This is actual parametric CAD compiled from a supplied, provider-reviewed mechanical contract. The API models reviewed the contract and plan. Codex authored the separate Python CAD implementation; no claim is made that the models emitted this source.
The assembly contains 85 meaningful component instances and 8 explicitly counted fastener envelopes, representing 47 distinct designs and 10 functional subassemblies. Counts come from the same inventory exported to STEP, GLB and BOM.
Validation covers the static shown pose. Digital checks are not a motion study, tolerance stack, physical prototype, structural assessment or certification. Motors, bearings, optics and connectors are nominal envelopes. Screw threads and internal bearing mechanisms are simplified. Exploded offsets are presentation transforms, not a verified assembly path.
Reproduce from the repository: node scripts/cookbook-mechanical-render.mjs --only=inspection-workstation. It uses the existing offline ce-spec-native:spec-final-2 Docker image with build123d 0.11.1, Python 3 with NumPy/Pillow, and sharp. The container sees only an isolated scratch directory and has no network access. There are no API calls or commissioned specification jobs.
19 of 20 selected clearance/interference checks pass in the shown static pose. The separate valid-solid and STEP round-trip checks pass. These are computational results, not physical testing.
Follow the specification’s proposed part references to the actual compiled components and identify what remains unresolved.
This crosswalk was authored after the model review and CAD compilation. The model proposal IDs and the CAD instance IDs are separate namespaces. Mapping a part does not mean that the model approved its later construction details.
Sources: actual orchestrator review, mechanical contract, compiled inventory, CAD implementation and checks, machine-readable crosswalk and bound checks.
The inventory contains 85 component instances and 8 fastener envelopes, with 46 distinct non-fastener designs and 10 functional subassemblies. These physical groups sit inside the three LIGHT planning modules.
| Model proposal | Meaning | CAD design ID | Actual CAD instances |
|---|---|---|---|
| P01 | Base | station-base |
W01 |
| P02 | Y rail | y-guide-rail |
W04_0, W04_1 |
| P03 | Y carriage | y-carriage |
W05_0_0, W05_0_1, W05_1_0, W05_1_1 |
| P04 | Y saddle | y-saddle |
W06_0, W06_1 |
| P05 | Gantry tower | gantry-tower |
W07_0, W07_1 |
| P06 | Bridge | gantry-bridge |
W09 |
| P07 | X rail | x-guide-rail |
W10 |
| P08 | X carriage | x-carriage |
W11 |
| P09 | X screw | x-lead-screw |
W12 |
| P10 | Focus backplate | focus-backplate |
W18 |
| P11 | Focus rod | focus-guide-rod |
W19_0, W19_1 |
| P12 | Focus screw | focus-lead-screw |
W23 |
| P13 | Camera-body envelope | camera-envelope |
W26 |
| P14 | Lens envelope | lens-barrel |
W27 |
| P15 | Ring-light envelope | ring-light |
W28 |
| P16 | Deck | perforated-deck |
W31 |
| P17 | Standoff | deck-spacer |
W32_0, W32_1, W32_2, W32_3 |
| P18 | Rear panel | rear-panel |
W42 |
| P19 | Side panel | side-screen |
W43_0, W43_1 |
| P20 | Controller housing | controller-shell |
W37 |
| P21 | Controller lid | controller-lid |
W38 |
| P22 | Carrier link | carrier-link |
W45_01, W45_02, W45_03, W45_04, W45_05, W45_06, W45_07, W45_08, W45_09, W45_10, W45_11, W45_12 |
| P23 | Foot | station-foot |
W02_1, W02_2, W02_3, W02_4 |
P01–P23 and SA01–SA10 come from the actual mechanical worker findings, explicitly accepted by the final orchestrator.
| Model group | Meaning | Actual CAD group |
|---|---|---|
| SA01 | Base and feet | 01_base |
| SA02 | Y guidance/carriages | 02_y_axis |
| SA03 | Gantry | 03_gantry |
| SA04 | X guidance/carriage/screw | 04_x_axis |
| SA05 | Z guidance/screw | 05_focus |
| SA06 | Optics envelopes | 06_optics |
| SA07 | Deck/workholding | 07_workholding |
| SA08 | Enclosure | 09_enclosure |
| SA09 | Controller housing | 08_controller |
| SA10 | Cable carrier | 10_cable_carrier |
These designs were supplied by the later CAD author. They do not receive a fabricated P identifier or retrospective model approval. The original proposed BOM remains unchanged.
| CAD design | Meaning | Actual instances | Group |
|---|---|---|---|
y-riser |
Rail support riser | W03_0, W03_1 |
02_y_axis |
gantry-gusset |
Triangular saddle gusset | W08_0_0, W08_0_1, W08_1_0, W08_1_1 |
03_gantry |
x-bearing-support |
X shaft bearing support | W13_0, W13_1 |
04_x_axis |
x-motor-envelope |
X motor nominal envelope | W14 |
04_x_axis |
x-motor-mount |
X motor L mount | W15 |
04_x_axis |
x-coupler |
X flexible coupling envelope | W16 |
04_x_axis |
x-lead-nut |
X lead nut | W17 |
04_x_axis |
focus-end-block |
Focus shaft end block | W20_0, W20_1 |
05_focus |
focus-slider |
Focus carriage plate | W21 |
05_focus |
focus-bushing |
Focus linear bushing | W22_0, W22_1 |
05_focus |
focus-motor-envelope |
Focus motor nominal envelope | W24 |
05_focus |
camera-mount |
Camera angle mounting bracket | W25 |
06_optics |
lens-window |
Optical window envelope | W29 |
06_optics |
ring-bracket |
Illuminator support bracket | W30 |
06_optics |
locating-pin |
Locating pin | W33_0, W33_1 |
07_workholding |
hold-down-clamp |
Slotted hold-down clamp | W34_0, W34_1 |
07_workholding |
clamp-hand-screw |
Clamp hand screw envelope | W35_0, W35_1 |
07_workholding |
calibration-coupon |
Perforated calibration coupon; uncalibrated | W36 |
07_workholding |
switch-envelope |
Manual switch envelope; not wired | W39 |
08_controller |
panel-connector |
Panel connector envelope | W40_0, W40_1 |
08_controller |
cable-gland |
Cable gland envelope | W41_0, W41_1 |
08_controller |
rear-post |
Rear panel post | W44_0, W44_1 |
09_enclosure |
carrier-end-bracket |
Cable carrier end bracket | W46_0, W46_1 |
10_cable_carrier |
M6x12-envelope |
Nominal M6 mounting screw envelope | B01_1, B01_2, B01_3, B01_4, B01_5, B01_6, B01_7, B01_8 |
11_fasteners |
FX-DECK-GRID-r1 maps to P16 / W31, the deck grid. Later locating pins, clamps and coupon are distinct author additions; the reference does not establish tested workholding.IMG-ASM can now resolve to assembled.svg, and IMG-EXP to exploded.svg. These are renderings of the compiled geometry.IMG-SEC remains unresolved: the engineering drawing provides orthographic views, not a claimed section cut or motion study.M-SETUP-r1 is still a proposed application manifest. This crosswalk does not create a deployed staff application, an inspected sample or an acknowledged handoff.Unresolved in the CAD author’s own manifest:
The supplied dimensions can agree with the inventory while parts still overlap. The CAD builder reports these failed selected geometry checks; their conflicts remain unresolved in this saved concept.
| Failed check | Parts | Reported geometric result |
|---|---|---|
| Focus lead screw intersects locked camera envelope | W23, W26 |
{"intersectionVolumeMm3":1130.97335529,"maximumMm3":0.001} |
Feet extend below the base-origin plane as an explicit authored placement. The model-reviewed outer dimensions remain traceable; added features and extra designs are identified separately.
Independent inventory comparison: 85 selected bounds, positions and counts agree within 0.0001 mm numeric tolerance. This checks the compiled inventory against the supplied values; it does not repeat the CAD kernel or prove that every interface, bore, wall, stroke or attachment is correct.
For the workstation, the original focus plate is 8 mm thick; its integrated 8 mm bosses make the finished part bounding depth 16 mm. The audit explicitly accounts for this authored addition instead of claiming that the whole finished part still has an 8 mm depth.
The inputs are synthetic teaching material. This run used the cookbook’s reproducible API runner; its model split is not a model-routing control in the current Synatrail interface. The package uses Synatrail’s normal export format. Canvas views use the original three cards, with layout and explanatory relationships added for this walkthrough. The architecture and process diagrams retain the exported topology; card summaries and link labels were added for readability. The original records and diagrams remain available.
4 API calls · US$0.895764 estimated token cost. Exact model identifiers remain in the downloadable execution record.
The earlier single-part counter stand and repair-shop service study are preserved as additional reference material.
Use this when an attractive idea has acquired more confidence than evidence. Example: “A room sensor will reduce wasted heating.”
Three ingredients: an Idea containing the claim; a Dataset with timestamped temperature and occupancy observations; a Context card explaining sensor placement, missing periods and what “waste” means. Replace the example with your actual records before drawing conclusions.
Connect: explain how the dataset supports or challenges the idea, and how the context limits that interpretation. Select the relationship to ask about that particular inference.
Audit this inference. Quote the observation that supports it and name what it does not establish. Give two alternative explanations, the missing measurement that would distinguish them, and a result that would make us reject the claim. If the selected evidence is insufficient, say exactly why.
Where available, Challenge on the relationship creates a configured reasoning step with that exact target. Save the question and required inputs, review the path and its model costs, then start it deliberately. For a direct scoped answer, use Ask with the prompt above.
“Temperature and occupancy may suggest a mismatch. Energy savings remain unmeasured. Compare energy use under equivalent conditions before claiming an improvement.” A narrower claim and a discriminating test are valuable results.
Stop when: you have one test that separates the explanations. If the model simply repeats the claim, narrow the selection and ask for the exact missing evidence. See how relationships work →
Use this when two solutions sound reasonable. For workshop handovers, compare a shared checklist with a purpose-built log before deciding to build.
Three ingredients: one Context card with the decision criteria, plus one Idea card per option. Record the same constraints for both: who enters data, who acknowledges it, how recovery works and what effort is acceptable. Link each option to those criteria.
Compare both options against the selected criteria. Use a table: criterion, evidence for option A, evidence for option B, unknown, next check. Do not invent scores or fill missing evidence with confidence. Recommend which reversible test to run first and state the condition that would change that choice.
With a configured reasoning step, choose Compare, provide the actual alternatives and approve the criteria before running. Use Run selected path to inspect exact inputs and limits first. For a numerical comparison, Branch and compare preserves a baseline while you change a parameter and rerun its experiment. A comparison table is a proposal to review; it is not proof that an option works.
Stop when: you know which option earns the next test and what is still unknown. Keep an inconclusive result if the evidence cannot distinguish them. See experiments and comparisons →
Group by the question you want to ask. A useful group name and purpose help the agent read the evidence together, and help you return to the reasoning later.
Purpose: “Assess whether missing ownership is the cause of handover failures. Keep evidence against this explanation visible.” Ask this group for the weakest inference. Useful for literature reviews, root-cause questions and challenging a product premise.
Purpose: “Evaluate the shared-checklist route against the same acceptance criteria as the custom-log route.” Use one group per route on a larger canvas. Keep common criteria in a referenced card and explicit connections, so one option does not quietly receive an easier test.
Purpose: “Collect the reviewed basis for a one-workshop prototype and the uncertainties the builder must preserve.” Useful when preparing a package. Mark the actual evidence you intend to include; being inside a frame does not make a card true or approved.
Press G, select the relevant cards, then Make a group. Fill Group name and What connects this group?, then Save group.
Use Ask this group for its nested evidence. Use Show only this group to focus the view. Use Ungroup if the framing no longer helps; its cards and connections remain.
A group frame is itself a canvas node. These arrangements illustrate larger investigations. The live example stays at three total nodes without a frame; select its three cards together to ask a scoped question.
Use Ask for an answer now. Use Steer to direct the research agent’s next work within its saved limits. A good instruction names the scope, the challenge, the output and when to stop.
Focus on handover ownership and acknowledgement. Challenge the need for custom software by considering an existing shared checklist. Keep supplied observations separate from assumptions. Return the strongest objection, the evidence needed to resolve it and one small next test. Do not expand into inventory, purchasing or predictive maintenance. Stop after this comparison and surface the unresolved decision.
| When you see | Steer towards | Review afterwards |
|---|---|---|
| Too many directions | “Only investigate the cause that could change this decision.” | Whether the next pass stayed inside that question. |
| Confident agreement | “Find the strongest counterexample and show its source.” | Whether the objection has evidence behind it. |
| Repeated findings | “State what is already known and name the one unresolved gap.” | Whether another call can add useful information. |
| A growing feature list | “Keep one user, one workflow and one acceptance test.” | Whether every feature serves that test. |
Read Why before accepting a discovery. Check Activity for what ran and its costs. Use the agent’s pause/stop controls when you have enough; a sentence asking it to stop is not a substitute for the saved automation and budget controls. See the agent watch and its limits →
A useful specification tells a builder who it serves, what to implement first, how to judge it and what is deliberately outside the scope.
In this cookbook, light package names the earlier composer's planning archive used for the saved examples: a product definition, modules, stories, acceptance criteria, risks and open questions. Those downloads and screenshots preserve that original draft. Current deployments with comprehensive packages enabled use Produce package. The steps below prepare a new bounded handoff; the saved example does not claim a comprehensive run or implementation.
Select the three evidence cards and choose Prepare a package draft, or open Produce package and review Selected investigation context. Include only the cards, reviewed findings and decisions that belong. The commissioned package pins their exact revisions.
Keep the brief below in the existing Context card before selecting it. In Requested outcome, ask for a small workshop handover-log prototype; choose Software product and the intended delivery stage. Start with Basic if its displayed scope meets your need. Resolve unknown choices, select a qualified provider and tools, and authorize processing only the selected material.
Use Examine scope and quote to inspect inclusions, blockers and price. An optional Pre-flight — 5 credits saves an explanation you can revise without another Pre-flight charge. Save manual draft without model calls keeps an ungenerated draft; automatic generation requires accepting the displayed full package price and available routes. Inspect artifact revisions and checks, then freeze a draft or ready archive and download that exact release.
If your deployment still opens the earlier Build package composer, the historical light-package route is Cards → Notebook → Earlier answers, Software, Steering, scope review, then Skip the pitch, build the package. It produces a planning archive rather than a versioned package release. See both package flows and their costs →
Define a small workshop handover-log prototype for one workshop. Keep the first version focused on recording unfinished maintenance work, an accountable owner, the next action and acknowledgement by the incoming shift. Preserve the simpler paper-checklist alternative. Exclude inventory, payments, machine control, predictive maintenance and multi-site rollout. Include observable acceptance criteria, evidence references, risks and open questions. Label the supplied notes as synthetic and any benefits as hypotheses. Deliver a light planning specification; do not claim implementation or field validation.
| Package idea | Keep the first version to | Evidence to obtain next |
|---|---|---|
| Workshop handover log | One handover workflow and a visible acknowledgement. | Whether the incoming shift can identify the owner and next action. |
| Room comfort monitor | One room, a documented measurement and a readable trend. | Sensor placement, measurement quality and whether a person acts on the result. |
| Support triage assistant | Draft one suggested category for human review. | Performance on labelled examples and which errors require abstention. |
Review rule: a story should name a user action and an observable result. “Easy to use” is unfinished. “After the incoming shift acknowledges an item, the outgoing shift can see who acknowledged it and when” gives a builder something to implement and a reviewer something to check.
Check that references lead back to your selected evidence and that assumptions remain visible. The archive is a starting point for a builder; generated acceptance criteria are still tests to perform. See the package viewer and archive →
Keep a notebook entry after a useful pass: decision, evidence, uncertainty, next test, stop condition. Include why an option was rejected. This makes a later return cheaper than asking the model to rediscover your reasoning.
Before another run, ask what new evidence or changed assumption could alter the decision. If nothing has changed, inspect the existing answer. Select only the cards needed for the next question, preserve source versions and spend the next call on the unresolved part.
Download and inspect a package before you hand it to a builder. Keep illustrative inputs labelled wherever the result travels.
“We will test the smallest handover workflow with the next shift. The benefit is still a hypothesis. If a shared checklist meets the same need with less effort, we will use that instead.”