Brief, critique, ship

The director's brief is for long or unattended films. CLAUDE.md holds studio rules; the brief holds only film-specific facts: one-line goal, audience, CTA, brand source, fictional data source, references (and what to take from each), beat sheet, deliverables.

Unattended runs. Claude can't wait on a clock for you, so "continue if I don't answer in 10 minutes" doesn't work. Instead:

• Ambiguous choices go into decisions.md, with the rejected alternative.
• Run-stoppers go into BLOCKED.md instead of being guessed: unclear licences, a need for real customer data, or a check failing three times.

The critique loop, done properly

1. Objective checks first. check.py must pass before any critique. It covers determinism, sync, state geometry, text visibility, loudness, clipping and the small-speaker test.
2. A fresh reviewer. Research finds AI evaluators recognize and favor their own outputs (self-preference bias). The reviewer is a separate subagent that sees only the brief, the contact, phone and spectrum images, and the check report.
3. Evidence-backed scores. Any score under 8 must cite a timestamp and an image. Maximum 3 rounds.
4. A human last. Taste, humor and music are judged by a person.

Packaging. The whole pipeline is a project skill at .claude/skills/motion-film/, containing SKILL.md plus brief, critique and spec templates. One rule in it matters most: never edit a check to make it pass.

Fact-check of the article

• ✅ PDoom: Opus wrote ANIMATION_GUIDE.md for parallel subagents and STORYBOARD.md, with nine chapters (confirmed in the repo).
• ⚠️ Self-scoring alone is lenient; use a separate reviewer.
• ⚠️ "$1,000 per video, delivered in an afternoon" is one anecdote, not market pricing.