Task → Contract → Guidance → Draft → Critique → Final.

Task

It starts from the discussion you already had. You are not asked to repeat what is known. The skill notes the target agent (Claude Code or Codex) and the executor model only when the conversation actually establishes them.

Contract

Outcome, context, decisions, constraints, non-goals, acceptance criteria and permissions are extracted, and each requirement is classified:

EXPLICIT_REQUIRED

Something you asked for. Anchored to your own words and kept in the final prompt.

EXPLICIT_OPTIONAL

Something you mentioned as a preference. It stays optional.

INFERRED

A reasonable reading of the discussion. Context only — never a mandatory instruction.

UNRESOLVED

A decision nobody made. Left open, or one focused question is asked.

FORBIDDEN

Something you ruled out, including withheld permissions such as merge or deploy.

Text you mark as immutable is copied verbatim. The mode you asked for — discussion, investigation, proposal or implementation — is preserved; a question never becomes edit authority.

Official guidance

Only documentation from the AI developers themselves is consulted, and only where it applies: agent-level notes for the known target, model-specific advice only for an established model. Retrieved pages are treated as untrusted reference data.

• Live-verified — Fetched and read during this use, with the retrieval date.
• Dated cached — From the ledger snapshot, shown with its date.
• Stale — Older than the source’s refresh policy allows.
• Unavailable — Could not be retrieved; no claim is made from it.

See all 17 recorded recommendations →

Draft

First, the original task is critiqued. If it is already suitable, the answer is “no material revision is necessary”. Otherwise only the defective span changes: your operative statements stay the core, unresolved choices stay open, and new ideas go under “Optional suggestion”. Small tasks get small prompts.

Critique

One critique and revision pass against a compact rubric: requirement coverage, assumptions, contradictions, authority, realistic tools, acceptance checks, missing context, privacy and verbosity. A second pass happens only for an unresolved material defect. Same-model review is called self-review, not independent verification.

Final

A two-way drift check: every original required or forbidden item must reach the final prompt, and every final operative clause must trace back to the original. New requirements, strengthened suggestions and unauthorized actions are removed or made optional. The result is one copyable prompt.

An optional offline checker can verify annotated provenance; a clean result still needs semantic review.