Double-layer auto-optimization Agent Skill — host-agnostic: use with any IDE, CLI, or Skill-capable tool (OpenClaw, editor integrations, etc.), not tied to a single vendor.
An Agent Skill workflow (entry: root SKILL.md + docs/): lock requirements first, iterate multiple Prompt phrasings in parallel, write each round to disk under tracks/prompt-*/, use an acceptance checklist (not self-score alone) as the delivery bar, and optionally hash-compare delivery vs the chosen final track file.
- Vague requirements → churn without a baseline
- Only one phrasing → no A/B on “how to ask”
- “Five iterations” with no files → not auditable
- Self-score 8/10 while features missing
- No proof delivery matches the “final” track file
This Skill addresses that with four phases and five verifiability rules (R1–R5).
- Keep this repo at your project root (or merge
SKILL.md,program.md,docs/,scripts/into an existing repo). - Optional: add host-specific rules (e.g.
.cursor/rules/, OpenClaw config) to enforce the workflow. - Ask for features in your host; the agent should follow the phases when this Skill is loaded.
- Verify from repository root:
FINAL_TRACK_FILE=tracks/prompt-a/r05.html DELIVERY_FILE=index.html bash scripts/skill-verify.sh| Path | Role |
|---|---|
SKILL.md |
Skill entry (YAML description for discovery) |
program.md |
Shortest path for agents |
docs/ |
R1–R5 and phase 1–4 norms |
scripts/skill-verify.sh |
Optional checks (run from repo root) |
MIT — see LICENSE.
