Video Review
This skill handles the judgment of whether a generation result is worth entering the current production pipeline. It does not lower existing production constraints simply because the asset has already been generated, nor does it rewrite upstream facts to preserve an already-generated result.
1. Inspect the Actual Generation Result First
Review must be based on the actual returned images, videos, or audio, not on prompts, filenames, or assumptions about what the model "should have done."
When the current executor can directly view / play media, inspect the complete artifact; when only partial frames, screenshots, or no audio are available, explicitly note which dimensions have not been verified, and do not pass unobserved content.
To quickly read media metadata, SHA-256, uniformly sample to locate problem frames, or split strict 2×2 four-views frame by frame, prioritize using this skill's built-in scripts/media_review.py: probe reads real media properties, sample writes only auxiliary video frames into the current project's .tmp/review/, split-2x2 mechanically cuts a 1:1 image into four quadrants, and cleanup removes the review auxiliary directory. Sampled frames / split grids / manifests can never replace full playback, audio listening, lip-sync, or character / scene content verification; a mechanical 2×2 success does not mean strict four-view semantic approval.
Video review cannot rely on a single best-looking still frame; it must at least cover the beginning, main action process, and end state, and check motion / audio continuity.
2. Fixed Review Basis
First read the official source corresponding to the result:
- Reusable assets: corresponding long-term prompts, character/scene/product definitions, and necessary upstream facts;
- Shot generation results:
SHOT.md, current prompt, referenced reusable assets, and adjacent shot continuity; - Advertising products: simultaneously load product facts and brand constraints from
video-advertising; if this capability is missing, cannot claim the product Gate is complete, first return toakira-videoto supplement the optional Package.
The one-time generation package is only used to locate the current round's input, not a higher priority source of facts.
For detailed generation result / asset quality reviews, read references/TAKE-QA.md. When multiple image, audio, or video candidates are returned at once, also read references/CANDIDATE-SELECTION.md, first eliminating identity / product / space / storyboard hard errors, then comparing production suitability and aesthetics. When results are clearly failing, first locate the failure type according to references/FAILURE-MODES.md, then select adopt, post-production fix, regenerate, or rewrite upstream using references/REPAIR-DECISIONS.md; before regenerating, clearly identify the one main variable being changed this round.
3. Review Dimensions
Only check the dimensions applicable to the current object:
- Identity, costume, material, structure, and state stability;
- Scene space and key positions validity;
- Action, performance, cinematography, and timing fulfilling the shot's purpose;
- Continuity In / Out compatibility with adjacent shots;
- Product, logo, text, and industrial structure matching factual sources;
- Dialogue, lip-sync, sound, and ambient audio meeting current requirements;
- Any obvious drift likely to amplify in subsequent shots.
Do not let a single good-looking frame replace full segment video review.
4. Conclusion
Prefer one of the following clear handling methods:
- Accept: the result can enter the current production;
- Accept and Adjust Subsequent: non-critical deviations exist, but accept the actual frame, and let the next shot continue from the real visible state;
- Post-Production Fix: the main body is generated correctly, the issue is more suited to deterministic editing, compositing, subtitles, logos, volume, or other post-processing;
- Regenerate: the shot/asset definition is correct, but the current generation result failed;
- Rewrite Definition or Prompt: the failure came from
SHOT.md, long-term assets, or the prompt itself, not random execution.
5. Record Accept Decisions
Shot acceptance decisions follow references/TAKE-DECISION.md to keep concise: record the current generation result, actual exit, acceptable deviations, and items still needing post-production fixes; do not create an extra database.
For which generation result to accept for a shot, maintain a brief "current take" field or equivalent clear record in that shot's SHOT.md, and do not copy a final.mp4.
Eliminated generation results may still be kept for comparison; if the project does not need to retain failed outputs long-term, they can be deleted according to user storage policy, but must not delete the only archived generation result prematurely.
Completion Criteria
Every reviewed result has a clear basis and handling direction; accepted results can proceed to the next shot or full post-production; when regenerating, it is also clear what will be changed next time.