demo-review

Prompt file imported from yousuf7474/demo-studio (.claude/commands/demo-review.md). Copyright stays with the author.

/demo-review

Resolve the project: use the named slug, else the single directory under projects/, else list slugs and ask. Read projects/<slug>/gates.json.

Prerequisites. Refuse and name the exact command to run first unless ALL hold: G17 RENDER_INTEGRITY, G18 CAPTIONS, G19 LEGIBILITY_SAFE_AREA are pass and fresh (input_digest matches current render) — else "run /demo-render first"; G20 SELF_EVAL has a recorded run against the current draft — else "run /demo-verify first"; G10 CLAIM_TRACEABILITY is pass and fresh — else "the script changed since G10 last passed; run /demo-narrate or /demo-gate G10 first."

Round cap. Count existing projects/<slug>/review/review-*.md files. If this would be round 4, do not proceed — write review/ESCALATION.md ending in a small, costed, enumerated choice set (never "what should I do?") and stop.

Assemble the packet as projects/<slug>/review/review-NNN.md (zero-padded, next number):

• Draft video: edit/preview.mp4.
• verify/contact/ composites into qc/contact_sheet.png — one tile per scene at its 40% point, labeled scene id / beat / timecode / primary claim.
• verify/timeline/ composites into qc/timeline_view.png — filmstrip + waveform + word labels + shaded silences + scene rules + claim ribbon + red bands at advisory timecodes.
• The G10 traceability TSV: read the evidence path of the latest passing G10 run in gates.json and cite it directly — never regenerate a summary.
• Checklist items, each derived (never freehand): one line per beat (from plan.md / storyboard.json scenes), one per facts.json claim with risk:"high" or confidence:"low", one per open advisory (non-blocking WARNs across the ledger, e.g. G09/G11/G20 items under 3), one per brief.json assumptions[] entry, plus the 6 fixed items ending in "would you send this to a customer?"

Present the packet and ask for the human's response in prose. Translate it into proposed feedback.schema.json items — {id, timecode, scene, beat, kind, severity, note, proposed_fix, repair_class, affects[]} — using the remediator subagent to assign each repair_class when not obvious (it may never propose skip/waive). Show the human this structured proposal and get explicit confirmation before writing anything — that confirmation is the gate record.

Write review/feedback-NNN.json ({review_id, draft, reviewer, at, verdict, items[], accepted_advisories[], checklist{}}), validate it, then append the G21 run to gates.json: pass if verdict has no blocking items, else fail with the item list. Report the round number, packet paths, feedback path, and the next command — /demo-rerun --feedback review/feedback-NNN.json if items exist, /demo-ship if clean.