publish-prep-review

name	publish-prep-review
description	Score the script, validate the rendered video, and produce publish metadata so the agent can decide whether a short is ready to upload.
allowedTools	["shell","read","write"]
entrypoint	node --import tsx scripts/harness/publish-prep-review.ts
inputs	[{"name":"videoPath","description":"Path to the final rendered MP4.","required":true},{"name":"scriptPath","description":"Path to the source script.json artifact.","required":true},{"name":"outputDir","description":"Directory that will receive the review bundle.","required":false},{"name":"platform","description":"Target platform such as tiktok, reels, or shorts.","required":false},{"name":"captionExportPath","description":"Optional captions.remotion.json sidecar for rendered caption-sync review.","required":false},{"name":"assetLedgerPath","description":"Optional asset ledger for rights, license, attribution, generated-asset, and audio claim review.","required":false},{"name":"mediaIndexPath","description":"Optional media-index.v1.json with metadata.provenance or metadata.audioSource entries.","required":false}]
outputs	[{"name":"validate.json","description":"Video validation and gate report."},{"name":"score.json","description":"Script and readiness score report."},{"name":"publish.json","description":"Publish metadata and checklist."},{"name":"provenance.json","description":"Optional rights and asset provenance review when provenance inputs are supplied."}]

Publish Prep Review

Use When

• The user asks whether a render is ready to post.
• The agent needs a serious readiness gate before human review.
• Claude Code or Codex needs upload metadata plus validation in one bounded step.

What This Skill Owns

• Final readiness judgment for a short.
• Combining script quality, video validation, and publish metadata in one review step.
• Failing bad outputs closed instead of pretending they are publishable.
• Pointing the agent back to the right fix path using short-form-production-playbook.
• Applying content-machine policy on top of reusable analyzer evidence from @45ck/video-evaluator when generic video facts, contact sheets, layout safety, or OCR evidence are available.

Core Approach

1. Judge the rendered short, not the intent behind it.
2. Treat validation warnings as editorial signals, not just technical noise.
3. Reject source-text collisions, unreadable cadence, freeze, and weak pacing even if the MP4 is technically valid.
4. Reject archetype violations: pure gameplay modes cannot contain random stock clips, split-screen lanes cannot preserve accidental black gutters, and cards/receipts cannot be cropped beyond readability.
5. Use the review bundle to decide whether to rework script, audio, visuals, or captions.
6. When the bundle fails, route the fix through short-form-production-playbook instead of retrying the same bad plan.

Inputs

• final video.mp4
• source script.json
• target platform
• optional validation preferences
• optional assetLedgerPath and mediaIndexPath provenance evidence

Outputs

• validate.json
• score.json
• publish.json
• optional provenance.json
• optional packaging metadata

Output Contract

• Writes validate.json, score.json, and publish.json under the requested outputDir.
• Optionally writes packaging.json if packaging generation is enabled.
• Optionally writes provenance.json if assetLedgerPath or mediaIndexPath is supplied. Any provenance error fails the final passed result.
• Validation can include cadence, visual quality, temporal quality, audio signal, freeze detection, flow consistency, and automatic rendered-caption sync checks in addition to the base format/resolution/duration checks.
• When the render shipped a captions.remotion.json sidecar, this step should OCR the final MP4 and verify the burned-in caption timing against the expected caption export instead of trusting the pre-render caption plan.
• If the lane used FFmpeg or any local fallback assembly, generate the caption sidecars first with node --import tsx scripts/harness/caption-export.ts and pass the resulting captionExportPath into review. Missing sidecars should be treated as a lane failure, not waved away.
• This step reviews the final render. It does not retroactively make already-captioned source footage acceptable; source-text hygiene belongs in the visuals selection step.
• The goal is a trustworthy go/no-go decision, not paperwork.

Native-Feel Hard Rejects

• Reject reddit-post-over-gameplay if any random clip, stock B-roll, generated scene, support footage, or top lane appears.
• Reject visible black gutters, boxed intermediates, or default blank template backgrounds unless the format explicitly calls for them.
• For public examples, run npm run review:demo-videos and inspect the aggregate contact sheet before linking the MP4 from docs/demo or the README.
• Demo review keeps the content-machine output contract but delegates .layout.json sidecar checks to @45ck/video-evaluator when the package or a built sibling checkout is available.
• Reject any crop that cuts off the primary face, card text, receipt, UI target, gameplay affordance, or other key subject.
• Reject missing, silent, near-silent, or music-only audio when narration is expected.
• Reject missing caption sidecars, failed OCR caption sync, unreadable captions, or captions outside safe zones.
• Reject if the first 1s to 3s does not look native: no visible stakes, no hook asset, or only generic gameplay/stock without context.

Rights And Provenance Gate

• This gate verifies the existing ledger, source notes, media index, or provenance files. Do not rerun scouting here; fail with a fix path if evidence is missing.
• Runtime review accepts assetLedgerPath with assets[] or items[], plus mediaIndexPath whose entries store evidence under metadata.provenance or metadata.audioSource.
• generate-short emits provenance/asset-ledger.json automatically and passes it into publish-prep unless a custom publishPrep.assetLedgerPath is supplied.
• Public examples need source notes or an asset ledger for every external model, texture, clip, image, icon, font, audio track, component copy, and AI-generated asset.
• Fail public readiness when rights are unknown, source evidence is missing, attribution is required but absent, or the asset is editorial-only, non-commercial, no-derivatives, watermarked, scraper-only, or likely to trigger a platform claim without proof.
• Fail public readiness when YouTube-origin audio lacks ownership, explicit permission, official download/export evidence, or compatible license plus permitted access.
• Music, SFX, ambience, and extracted audio need Content ID risk notes plus license certificate or permission evidence when the source provides it.
• AI-generated media needs provider, model, prompt/reference, job id or workflow, settings, and output hash so the render can be audited later.

Optional Runtime Surface

• Repo-side runner: node --import tsx scripts/harness/publish-prep-review.ts
• Supporting code: src/harness/publish-prep.ts, src/validate/*, src/score/*

Validation Checklist

• validate.json exists and passed is true.
• score.json exists and passed is true.
• publish.json exists and includes a checklist and description.
• The review outcome clearly tells you what to fix next if it fails.
• The rendered video still matches the named archetype selected in the plan.
• No black gutters, unrelated media layers, cropped-off cards, or cropped-off subjects are visible in the final MP4.