Double-layer auto-optimization (Skill entry)

This file is the Skill package entry (SKILL.md). Normative text lives under docs/. Not tied to a specific IDE—any host that loads this Skill can use it.

Strict execution (hard constraints)

• The agent must follow program.md, docs/verifiability.md, docs/phase-01-spec.md through phase-04-output.md, and docs/execution.md in full; must not shorten the workflow to save tokens, time, or because it “feels enough”.
• Each tracks/prompt-*/ inner loop must satisfy docs/phase-03-inner.md: gates (e.g. at least r01, r02 per scheme) and main termination (all items checked in tracks/phase-01-acceptance.md; or 5 rounds max). Self-score “≥8” must not alone justify stopping (verifiability.md R3). Forbidden: claiming a track is “deliverable” and moving to phase 4 before the acceptance checklist passes.
• Forbidden: using “reduce workload” or other reasons not in the Skill to deviate; only when the user explicitly waives in the current turn may the full workflow be skipped.

Quick start (agent)

1. Read program.md (shortest path) or the checklist below; do not skip docs/verifiability.md (R1–R5).
2. Open phases in order: docs/phase-01-spec.md → phase-02-prompts.md → phase-03-inner.md → phase-04-output.md; read docs/execution.md for cost/order reminders.
3. Before phase 1 gates pass, do not rewrite delivery code; phase 1 must write tracks/phase-01-acceptance.md (R3); phase 2 must write tracks/phase-02-consistency-check.md (check + scheme–dimension table); r01 for prompt-b/prompt-c must not be a full-page copy of prompt-a; waivers must be explicit in the current user message.
4. Before claiming delivery matches the final track file, in a class-A environment actually run scripts/skill-verify.sh from the repository root (see program.md).

Execution checklist

- [ ] Read docs/verifiability.md (R1–R5)
- [ ] Phase 1: spec written, feature list non-empty; tracks/phase-01-acceptance.md exists with ≥1 item
- [ ] Phase 2: outer prompts + consistency check; tracks/phase-02-consistency-check.md written (with table)
- [ ] Phase 3: each tracks/prompt-*/ has at least r01.* and r02.*; non-a r01 is independently generated
- [ ] Phase 4: compare & output; at least one track fully checked before final delivery (R3; self-score reference only)
- [ ] Delivery claim: ran skill-verify.sh and recorded command + output (R2/R4)

Resource index

Resource	Path
Shortest entry	program.md
R1–R5 authority	docs/verifiability.md
Phases 1–4	docs/phase-01-spec.md etc.
Execution	docs/execution.md
Layout & verify	reference.md
Example	examples.md
Human guide	SKILL-GUIDE.md
Script	scripts/skill-verify.sh (run from repo root)

User waiver

Must be written in the current task message, e.g.:

This task waives the Skill / skips phases

(Synonyms allowed if equally explicit; the agent must not infer a waiver.)

High-risk domains

Does not replace regulation, clinical procedures, or human review; model self-scores and scripts are not sole safety evidence.