Critic

Two blind reviewers who stand in for the real audience, judged against a bar set first. One round.

1. Set the frame

Ask one question only if the audience is unclear.

• Audience: who really judges this (recruiter, first-time user, maintainer, investor).
• Bar: pass score out of 10. Default 8.
• Mode: one version, or 2 to 4 versions to compare.

2. Build the blind packet

1. Make a new, empty folder outside the project. A nearby v2.md gives the answer away.
2. Copy in only the work a real reviewer would see. For code, the diff plus changed paths.
3. Strip names, "v2", "new", dates, and hints of a favorite.
4. Comparisons: label A, B, C in random order (shuf); keep the key outside.
5. Running app: include the URL; reviewers use it live (main path, one bad input).

3. Run 2 reviewers

Launch 2 parallel subagents with different audience personas:

No subagent tool: do both passes yourself and do not mention it.

Each reviewer reads only the packet and checks:

• Contradictions between sections, numbers, and promises.
• Failure states: errors, bad input, a job that fails or data that goes stale, a user whose situation changes.
• Claims with no number or proof, and marketing language where a plain sentence belongs.
• Success targets: is there a number to hit?

Each returns a score per version and problems ranked P0 (blocks shipping) to P2, each with location, why it hurts this audience, and a paste-ready fix ("author to supply" if only the author has the fact).

Rules: every problem points to a spot in the work. Search the packet before calling something missing. No guesses about the business. Legal points are "likely; confirm with counsel". Never keep something that causes a problem you found. Fewer real problems beat padding. No repeated per-finding template.

4. Report

Unblind after scoring; use the lower score. A problem both raise is a must-fix; one reviewer's taste is a suggestion. Write for the author: no personas, reviewer count, method, or tools.

Verdict: Not ready for <deadline> as written; fixable in <effort>.
Fix first:
1. [P0] <where>: <problem>. Fix: <paste-ready change>.
2. ...
3. ...
Also: <P1/P2 items, one line each>
Keep: <what works, specific>
Could not check: <one line>

Pass: "Ready for <audience>" plus P1s. Comparisons add Scores: A 8, B 6.5, C 7. Pick A: <reason>. One verdict, top to bottom.

5. Stop

One round; the author decides what to fix. Re-review only if asked or after fixing a FAIL, with a fresh packet and the rejected fixes listed. Cap: two rounds.

6. Learn, quietly

A repeat mistake gets one line where it is read next time: CLAUDE.md (project rules), memory (user taste), or this skill (process gap). Never in the report.