| name | ship-idea |
| description | Run the 11-stage idea-to-production pipeline end-to-end for one idea. Delegates to /discovery, /pre-mortem, /implement or /crank, /security, review-daemon, /push. Stages 8 (lean gate) and 10 (wire-live + PROVE) are non-skippable. Triggers: ship idea, ship this, idea to prod, end-to-end idea, pipeline this idea, run the pipeline. |
| skill_api_version | 1 |
| context | {"window":"inherit","intent":{"mode":"none"},"intel_scope":"none"} |
| metadata | {"tier":"orchestration","dependencies":["discovery","pre-mortem","implement","crank","security","push"]} |
ship-idea Skill
Quick Ref: Run the 11-stage idea-to-production pipeline for one idea. Stages 8 and 10 are hard gates — never skip them. Output: tracked bead, PDR, impl, hardened PR, live wiring proven.
YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.
Efficiency Doctrine (enforce on every subagent dispatch)
Read ~/.claude/governance/subagent-token-efficiency.md before fanning out. Rules:
- Route C1/C2 mechanical work to
haiku/local-OL; C3 reasoning to sonnet; C4 synthesis to opus.
- Pass file references (paths), not payload dumps, to subagents.
- One artifact per agent dispatch.
- Use
codegraph_context / Explore for read-only discovery; never grep loops.
- Audit sprawl with
~/.claude/bin/agent_token_audit.py after heavy fan-out.
Execution
Given /ship-idea <idea text or bead-id>:
Stage 1 — Capture
Track the idea as a bead so it is never lost.
bd create "<idea-title>" --body "<idea text>"
BEAD_ID=$(bd current | awk '{print $1}')
Log: [S1] Bead $BEAD_ID created.
Stage 2 — Reuse-survey + PDR spec
Non-negotiable: the PDR must live in 08_PDRS/ and use the canonical template.
- Survey existing assets (codegraph or Glob) — build on, do not rebuild.
- Copy the template:
cp 08_PDRS/_PDR_TEMPLATE.md 08_PDRS/PDR_<SLUG>.md
Template path: C:/Users/kas41/chromatic-harness-v2/08_PDRS/_PDR_TEMPLATE.md
- Fill in: problem, reuse candidates, integration/actuation edge (required — this closes the "implemented-but-not-wired" gap), success criteria, and the live-observation plan (referenced in Stage 10).
Log: [S2] PDR written → 08_PDRS/PDR_<SLUG>.md
Stage 3 — Pre-mortem gate
Invoke the skill; do not skip even for small ideas.
/pre-mortem # reviews 08_PDRS/PDR_<SLUG>.md
- PASS: continue.
- WARN: log risk, continue.
- BLOCK: address the blocker, re-run pre-mortem, then continue.
Log: [S3] Pre-mortem result: <PASS|WARN|BLOCK> — <one-line summary>
Stage 4 — Decompose
Create child beads B1…Bn, dependency-ordered.
/plan
Verify with bd list --parent $BEAD_ID.
Log: [S4] N child beads created.
Stage 5 — Implement (TDD)
Choose the right delegator based on complexity (from PDR):
| Complexity | Delegator |
|---|
| C1–C2 (single file / mechanical) | /implement |
| C3–C4 (multi-module / systemic) | /crank |
Apply efficiency doctrine on all subagent dispatches. Tests must be written first; never trust agent claims of "passing" — run them yourself (Stage 6).
Log: [S5] Implementation delegated via /<implement|crank>.
Stage 6 — Verify yourself
Run the test suite; do not accept agent output as ground truth.
python -m pytest -q
All tests must be green before proceeding. Fix failures before Stage 7.
Log: [S6] Tests: <N passed, 0 failed>
Stage 7 — Harden gate
mcp__review-daemon__review_run
/security
Both must pass (or findings addressed) before Stage 8.
Log: [S7] Review-daemon: <result>. Security: <result>.
Stage 8 — Lean gate [NON-SKIPPABLE]
This gate is mandatory. Answer each question explicitly in the log.
Run the token audit:
python ~/.claude/bin/agent_token_audit.py
Answer in order:
- Boot tax — does this add always-on overhead? If yes, is it justified?
- Poll vs event — is any polling replaceable with an event-driven Magnet at an inflection point?
- On-demand vs always-injected — is context injection on-demand or always loaded?
- Swappable — is the heavy capture component swappable (thin interface, not hardwired)?
If any answer is "yes, unjustified" — fix it before Stage 9. A WARN is logged and you continue only after the fix.
Log: [S8-LEAN] boot-tax=<ok|WARN> poll=<ok|WARN> inject=<ok|WARN> swappable=<ok|WARN>
Stage 9 — Ship (PR)
/push
bd close $BEAD_ID
Confirm PR URL and that the bead is closed.
Log: [S9] PR: <url>. Bead $BEAD_ID closed.
Stage 10 — Make it LIVE + PROVE [NON-SKIPPABLE]
This is the #1 missed step. Shipping a PR is not live wiring.
-
Wire the artifact into the runtime path. Concretely:
- Register in the relevant registry/plugin/magnets list (e.g.
magnets/plugin.py:default_registry).
- Add the integration/actuation edge described in the PDR (Stage 2).
-
PROVE it is called — via live observation, not a unit test:
- Trigger the real runtime path (run the harness, fire an event, invoke the CLI hook).
- Capture evidence: log line, trace output, or Magnet event showing the new code executed.
python 02_RUNTIME/main.py --dry-run
If you cannot produce live evidence, the stage is not done. Fix the wiring and re-verify.
Log: [S10-LIVE] wired=<where>. proof=<log line or observation summary>
Definition of Done gate: check 08_PDRS/DEFINITION_OF_DONE.md before marking complete.
Path: C:/Users/kas41/chromatic-harness-v2/08_PDRS/DEFINITION_OF_DONE.md
Record the evidence durably so session closeout can enforce it (Gap C). After
Stages 8 and 10 pass for $BEAD_ID, persist the verdict — closeout reads this file
and refuses to close a bead that is missing any gate:
python scripts/record_ship_evidence.py --bead-id $BEAD_ID --lean-ok --live-ok --dod-ok
Stage 11 — Observe
Wire a Magnet or telemetry watch so production behavior is visible.
- Prefer event-driven Magnets at inflection points over pollers.
- At minimum: confirm an existing Magnet covers this path, or create a new one.
- Document the observation plan in the PDR.
Log: [S11] Observer: <Magnet name or telemetry sink>.
Completion Report
After all 11 stages, output:
ship-idea complete for <IDEA>
Bead: <ID> (closed)
PDR: 08_PDRS/PDR_<SLUG>.md
PR: <url>
Live at: <registration point>
Proof: <one-line live observation>
Observer: <Magnet / telemetry>
Gate log:
S3 pre-mortem: <result>
S7 harden: <result>
S8 lean: boot-tax=<> poll=<> inject=<> swappable=<>
S10 live: wired=<> proof=<>
<promise>DONE</promise>
Key Rules
- Stages 8 and 10 are non-skippable. If blocked, fix the blocker and re-run the stage. Never mark either done without the required evidence.
- Stage 2 PDR must use
08_PDRS/_PDR_TEMPLATE.md and must specify the integration/actuation edge.
- Stage 10 proof must be live observation, not a unit test assertion.
- Efficiency doctrine applies to every subagent dispatch — route cheap work cheap, pass refs not payloads.
- Stage 6 self-verify — never trust agent claims; run tests yourself.
- Never open a PR without review-daemon + security passing (Stage 7).
Reference Docs
| Doc | Purpose |
|---|
.agents/tmp/idea-to-prod-pipeline.md | Canonical 11-stage pipeline brief |
08_PDRS/_PDR_TEMPLATE.md | Mandatory PDR template (Stage 2) |
08_PDRS/DEFINITION_OF_DONE.md | DoD gate (Stage 10) |
~/.claude/governance/subagent-token-efficiency.md | Efficiency doctrine for all dispatches |
~/.claude/bin/agent_token_audit.py | Token audit (Stage 8) |