| name | autodesign-start |
| description | Start Autodesign setup from the eatsteak/autodesign skill package. Use when the user asks to initialize, start, install-check, bootstrap, validate state, inspect the artifact graph, check subskill readiness, record gates, compute dirty artifacts, generate canonical artifacts, run visual reference gates, run Pencil/DS/prototype/QA metadata records, generate frontend handoff docs, generate reconcile reports, or generate SkillOpt hardening reports. |
Autodesign Start
This is the only public skill exposed by the eatsteak/autodesign package at install time.
Current Stage
Stage 11 provides a one-shot bootstrap runtime that materializes the bundled Autodesign workspace template into a target project, deterministic manifest and artifact graph state-management scripts, private subskill readiness checks, canonical artifact generation from real project input files, visual reference prompt/candidate/selection record gates, Stage 08 Pencil/DS/prototype/QA metadata workflows, Stage 09 frontend handoff documentation, reconcile reports, Stage 11 SkillOpt hardening reports, and read-only advisory hooks.
Allowed Stage 09 generation is limited to canonical planning artifacts, visual reference records, Pencil live-check records, wireframe/prototype metadata, canvas export path records, primitive inventory, DS tokens/contracts, semantic visual QA, max-two-refinement records, JSON/Markdown frontend handoff documentation, and advisory reconcile reports. Pencil canvas edits and exports are performed by the active agent through Pencil; scripts only record real evidence and fail fast when live-check evidence, active batch_design/export_nodes evidence, an Autodesign-owned virtual .pen filePath, created/exported node bindings, or real export files are missing. Handoff must not create frontend source files or executable prototype code.
Allowed Stage 11 SkillOpt behavior is limited to reading an E2E PASS eval report, comparing skill prompt/version outputs on golden cases, recording accepted/rejected edits, and writing review-only patch proposal JSON. SkillOpt does not apply upstream skill edits, generate frontend code, call image generation, call Pencil MCP, or create design assets.
Bundled Payload
Runtime assets live under assets/payload/:
subskills/ for private Autodesign subskill contracts.
schemas/ for manifest and artifact graph JSON schema files.
scripts/bootstrap.mjs for deterministic plan/apply workspace materialization.
scripts/generate-canonical.mjs for deterministic canonical artifact generation and manifest/graph updates.
scripts/generate-visual-references.mjs for visual reference prompt, candidate, and selected-reference records.
scripts/generate-pencil-prototype.mjs for Stage 08 Pencil live-check, wireframe, primitive, DS, prototype, QA, and refinement-gate records.
scripts/generate-handoff.mjs for Stage 09 frontend handoff documentation and reconcile reports.
scripts/generate-skillopt.mjs for Stage 11 SkillOpt hardening reports and review-only patch proposals.
scripts/validate-state.mjs, scripts/check-dependencies.mjs, scripts/can-run-subskill.mjs, scripts/record-gate.mjs, and scripts/dirty-artifacts.mjs for state and contract readiness management.
hooks/ for read-only advisory hook adapters.
workspace-template/ for files materialized by the bootstrap runtime.
Read assets/payload/payload-manifest.json for the concrete Stage 09 payload inventory.
Bootstrap Command
Always plan before applying:
node autodesign-start/assets/payload/scripts/bootstrap.mjs --target /absolute/path/to/project --plan
Apply only after the user explicitly approves the bootstrap write:
node autodesign-start/assets/payload/scripts/bootstrap.mjs --target /absolute/path/to/project --apply --approve-bootstrap
If the plan reports any overwrite actions, apply only after the user separately approves overwrites:
node autodesign-start/assets/payload/scripts/bootstrap.mjs --target /absolute/path/to/project --apply --approve-bootstrap --approve-overwrite
Approval Gates
- Run
--plan first and report the create, overwrite, and preserve counts.
- Do not run
--apply unless the user explicitly approves writing bootstrap files.
- Pass
--approve-bootstrap with every --apply command.
- If the plan contains
overwrite, ask for explicit overwrite approval and pass --approve-overwrite.
- Never treat approval to bootstrap as approval to run later Autodesign generation phases.
State Commands
Validate manifest and graph JSON:
node autodesign-start/assets/payload/scripts/validate-state.mjs --workspace /absolute/path/to/project
Check required upstream dependencies and cycles:
node autodesign-start/assets/payload/scripts/check-dependencies.mjs --workspace /absolute/path/to/project
Check whether a named private subskill contract can be entered:
node autodesign-start/assets/payload/scripts/can-run-subskill.mjs --workspace /absolute/path/to/project --subskill interview
Compute dirty downstream artifacts from changed upstreams:
node autodesign-start/assets/payload/scripts/dirty-artifacts.mjs --workspace /absolute/path/to/project --changed canonical.requirements
Record an approval gate only with an explicit actor, timestamp, and write approval:
node autodesign-start/assets/payload/scripts/record-gate.mjs --workspace /absolute/path/to/project --gate state.record-gate --status approved --actor <actor> --at <timestamp> --approve-record
Canonical Generation
Plan before applying:
node autodesign-start/assets/payload/scripts/generate-canonical.mjs --workspace /absolute/path/to/project --plan
Apply only with explicit canonical generation approval and deterministic metadata:
node autodesign-start/assets/payload/scripts/generate-canonical.mjs --workspace /absolute/path/to/project --apply --approve-canonical-generation --actor <actor> --at <timestamp>
Generation reads real files from autodesign/inputs. UX rules, interaction model, screen-state matrix, and visual anchor proposals are blocked unless project inputs include explicit platform selection. Primary visual anchor proposals are emitted as unapproved proposals only.
Visual Reference Gates
Prompt records require generated canonical visual anchor proposals:
node autodesign-start/assets/payload/scripts/generate-visual-references.mjs --workspace /absolute/path/to/project --action prompts --plan
Apply prompt records only with explicit approval:
node autodesign-start/assets/payload/scripts/generate-visual-references.mjs --workspace /absolute/path/to/project --action prompts --apply --approve-visual-prompts --actor <actor> --at <timestamp>
Candidate records require a manually approved canonical.visual-anchor-selection gate and real generated image paths:
node autodesign-start/assets/payload/scripts/generate-visual-references.mjs --workspace /absolute/path/to/project --action candidates --prompt-id <prompt-id> --generated-output-path <path> --apply --approve-visual-candidates --actor <actor> --at <timestamp>
Selected references require explicit reference ids and approval:
node autodesign-start/assets/payload/scripts/generate-visual-references.mjs --workspace /absolute/path/to/project --action selection --reference-id <candidate-id> --apply --approve-visual-reference-selection --actor <actor> --at <timestamp>
The visual reference script never auto-selects references and never creates image files.
Stage 08 Pencil, DS, Prototype, And QA
Run generate-pencil-prototype.mjs --plan before any apply. All actions require approved selected visual references. Pencil-derived actions require a live Pencil MCP handoff: get_editor_state, batch_design, export_nodes, target autodesign/outputs/pencil/*.pen, frame evidence for canonical screens, export node bindings, and real canvas export paths.
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action live-check --pencil-live-check-source-path <path> --plan
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action wireframes --pencil-evidence-path <path> --canvas-export-path <path> --plan
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action primitives --plan
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action ds --plan
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action prototype --pencil-evidence-path <path> --canvas-export-path <path> --plan
node autodesign-start/assets/payload/scripts/generate-pencil-prototype.mjs --workspace /absolute/path/to/project --action qa --refinement-attempt 0 --qa-status pass --plan
Apply only with the action-specific approval flag, --actor, and --at. The QA action blocks refinement after two unsuccessful attempts.
Handoff contract when code cannot call MCP directly: Stage 08 is NOT READY until a live Pencil agent has targeted the Autodesign-owned virtual .pen filePath, created wireframe/prototype frames with batch_design, exported screenshots with export_nodes, and written evidence JSON containing evidenceType, mcpToolCalls, targetPenPath, liveCheckEvidenceHash, batchDesign.createdNodeIds, exportNodes.nodeIds, per-screen frames[], and exports[] with matching path/bytes/sha256/node bindings.
Stage 09 Handoff And Reconcile
Run generate-handoff.mjs --plan before any apply. Handoff requires approved selected visual references, Pencil live-check and wireframe records, DS primitives/tokens/contracts, prototype package, prototype visual QA, and a valid refinement gate.
node autodesign-start/assets/payload/scripts/generate-handoff.mjs --workspace /absolute/path/to/project --action handoff --plan
node autodesign-start/assets/payload/scripts/generate-handoff.mjs --workspace /absolute/path/to/project --action reconcile --changed canonical.requirements --plan
Apply only with the action-specific approval flag, --actor, and --at:
node autodesign-start/assets/payload/scripts/generate-handoff.mjs --workspace /absolute/path/to/project --action handoff --apply --approve-handoff-generation --actor <actor> --at <timestamp>
node autodesign-start/assets/payload/scripts/generate-handoff.mjs --workspace /absolute/path/to/project --action reconcile --changed canonical.requirements --apply --approve-reconcile-report --actor <actor> --at <timestamp>
Handoff writes only autodesign/outputs/handoff/handoff-package.json and autodesign/outputs/handoff/README.md. Reconcile writes only autodesign/logs/reconcile-report.json. Neither action may mutate upstream artifacts, call image generation, call Pencil MCP, or create frontend code.
Stage 11 SkillOpt
Run generate-skillopt.mjs --plan after autodesign/logs/eval-report.json exists and reports E2E PASS:
node autodesign-start/assets/payload/scripts/generate-skillopt.mjs --workspace /absolute/path/to/project --plan
Apply only with explicit approval:
node autodesign-start/assets/payload/scripts/generate-skillopt.mjs --workspace /absolute/path/to/project --apply --approve-skillopt-hardening --actor <actor> --at <timestamp>
SkillOpt writes only autodesign/logs/skillopt-report.json, autodesign/logs/skillopt-patch-proposals.json, and state metadata. Patch proposals are manual-review artifacts only.
Advisory Hooks
The materialized hooks are read-only adapters:
autodesign-status-advisory.mjs emits status injection, lightweight state validation, and turn summaries.
autodesign-boundary-advisory.mjs emits overwrite warnings and reconcile alerts.
Hooks print JSON only in debug/explicit JSON mode and must not generate artifacts or mutate files.
Materialized Files
The bootstrap runtime copies assets/payload/workspace-template/ into the target project. Stage 09 materializes:
AGENTS.md
.codex/config.toml
.codex/hooks/*.mjs
autodesign/README.md
autodesign/manifest.json
autodesign/artifact-graph.json
autodesign/.system/install-state.json
- marker README files under
autodesign/inputs/, autodesign/outputs/, and autodesign/logs/
When Invoked
- Confirm that
autodesign-start is installed and visible.
- Locate
assets/payload/scripts/bootstrap.mjs.
- Run the plan command against the requested target workspace.
- Report planned file actions and approval gates.
- Apply only with the explicit approval flags described above.
- For private subskill requests, run
can-run-subskill.mjs and report readiness before any apply command.
- For canonical generation requests, run
generate-canonical.mjs --plan first and apply only with --approve-canonical-generation --actor <actor> --at <timestamp>.
- For visual reference requests, run
generate-visual-references.mjs --plan first. Apply prompt, candidate, or selection records only with that action's explicit approval flag.
- For Pencil, DS, prototype, or QA requests, run
generate-pencil-prototype.mjs --plan first. Apply only with the action-specific approval flag, --actor, and --at.
- For handoff or reconcile requests, run
generate-handoff.mjs --plan first. Apply only with the action-specific approval flag, --actor, and --at.
- For SkillOpt requests, run
can-run-subskill.mjs --subskill skillopt, then generate-skillopt.mjs --plan. Apply only with --approve-skillopt-hardening, --actor, and --at.