shot-review
Skill0 starsChanged 3 months agoMITView on GitHub
Reviews generated footage before it reaches the edit — builds contact sheets from every clip, reads them against a rejection checklist, and decides accept, re-roll with a changed prompt, or downgrade to a graphic. Use after generating video shots, when deciding whether a take is usable, or when a finished film has a shot that feels wrong but the reason is unclear.
In this repository6 other files it gives its agents.
- final-cutSkill
- make-filmSkill
- score-filmSkill
- shot-designerSkill
- story-architectSkill
- veo-renderSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 32
Run every clip against this.
- Distinctive: line 57
Not quite the shot you imagined, but it serves the beat
- Very specific: line 67
Trim: If only the tail is bad, set in and the shorter duration and keep it.
- Very specific: line 82
Downgrade: If a shot has been re-rolled twice and still fails, it is telling you it was never a veo shot.
- Very specific: line 86
Keep rejected takes in shots/rejected/ with a one-line note of why in the shot's notes.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Shot review
This is a looking stage. Not a reading stage, not a reasoning-about-the-prompt stage. You generated a clip; now you have to actually see what came back, because generated video fails in ways the prompt gives no hint of.
Build the sheets, then read them
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/contact_sheet.py" \
--shots-dir shots/ --out qc/ --frames 6Then Read each PNG in qc/. Every one. Six frames across a clip catches almost
everything that matters; when something looks off, re-sheet that clip alone at
--frames 12 and look again.
For anything ambiguous — drift, morphing, a camera that reverses — extract the actual moment rather than guessing:
ffmpeg -ss 4.2 -i shots/03.mp4 -frames:v 1 /tmp/look.pngThe rejection checklist
Run every clip against this. Any single yes is a rejection, no matter how good the rest of the shot is.
Fatal — re-roll or downgrade, never ship
- Legible-ish text, letters, numbers, logos or watermarks anywhere in frame
- A user interface, dashboard, chart or diagram that a viewer might try to read
- Hands with the wrong number of fingers; faces that shift identity mid-shot
- An object that morphs, melts, or gains/loses parts across the clip
- A camera move that reverses direction or stutters
- Anything that reads as a real brand, a real person, or stock footage
Serious — usually re-roll
- The action completes in the first two seconds and the rest is dead
- Continuity break with its neighbours: different light, different world, different lens, different grade
- Too busy to read at the length it will actually be on screen
- Emotion is wrong for the beat — cheerful footage under a stakes line kills it
- The last second falls apart (very common; sometimes fixable by trimming)
Acceptable — note and move on
- Slight softness in the corners
- A minor physics oddity nobody watching once will catch
- Not quite the shot you imagined, but it serves the beat
That last category matters. The bar is does it serve the beat at the length it is on screen, not is it the shot in my head. Chasing the imagined shot is how a budget disappears.
Then decide, per shot
Accept. Set status: "approved", record the seed in render. Move on.
Trim. If only the tail is bad, set in and the shorter duration and keep it.
Free, and it works more often than people expect.
Re-roll. Change exactly one thing in the prompt, keep everything else. Changing three things teaches you nothing about which one mattered. Common single fixes:
| symptom | the one change |
| busy, unreadable | remove a subject; make the camera slower or locked-off |
| hallucinated UI/text | remove any screen, monitor or sign from the prompt; add to negative |
| dead after 2s | describe a slower action, or a camera move rather than a subject action |
| wrong mood | change the light, not the adjectives — light does the work |
| breaks continuity | add or fix reference_images; check prompt_suffix was applied |
| hands wrong | frame it wider, or crop the hands out of the shot description |
Downgrade. If a shot has been re-rolled twice and still fails, it is telling
you it was never a veo shot. Rebuild it as motion or still. This is a
correct outcome, not a defeat — and it is always faster than a third roll.
Keep rejected takes in shots/rejected/ with a one-line note of why in the shot's
notes. The rejection log is the most useful thing you will have on the next film.
After the cut, review again
Sheet the finished film too:
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/contact_sheet.py" \
--film out/film.mp4 --out qc/ --frames 8A whole-film sheet shows things no single clip can: whether the palette holds, whether two shots are near-duplicates, whether the middle sags into sameness, whether there is any visual variety in framing at all. If every frame is a medium wide, the film is monotonous regardless of how good each shot is.
And then watch it end to end, with sound, once. Every film has three things wrong with it that only appear on a full watch. Find them, fix them, ship.
References
references/failure-catalogue.md— annotated failure modes with the prompt change that fixes each, and the three that are not worth fixing
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

