Mouth moving in a silent shot
A character's mouth opens and closes when nobody is speaking, which reads as mumbling or panting. The frame check marks every frame where the mouth looks like it is talking in a shot that has no dialogue.
Free Claude Code skill · MIT licence
Claude Video Review Skill
Claude Code can make a video, but it cannot watch or hear the result. This skill reviews the finished video before you publish it.
The skill measures the audio and picture, checks every frame, and asks Gemini and GPT to review the video. Claude then receives a list of problems with timestamps and a fix for each one.
We email you the install guide, an example report and the full checklist.
- Checks every frame
- Reviews picture and sound
- 16-group checklist
- Free and open source
.video-review/report.md
VERDICT: REJECT 1 blocker · 2 majors
720 of 720 frames checked · Gemini and GPT reviews · 7 personas
- 0:05.0 Mouth moves in a silent shot
→ add "lips closed" to the prompt; generate again - 0:15.0 Shot slowed: 25% near-duplicate frames
→ generate a longer take at normal speed - 0:26.4 Audio click at an edit
→ add a short crossfade at the cut
Keep: opening hook · music entry at 0:12
Can Claude watch videos?
No. Claude reads text and still images, but it cannot play a video file with its soundtrack. Claude Code can render a video with Remotion, HyperFrames or ffmpeg, or assemble clips from Seedance, Veo or Kling, but it cannot check how the finished file looks and sounds. That review is a separate step, and without it problems reach your audience.
The Claude Video Review Skill fills that gap. Claude runs the review: it writes a short brief, runs the measurements, sends the video to independent reviewers, checks their findings against the frames, and fixes the problems that are confirmed. Gemini watches the video with sound, GPT inspects still frames, and measurement scripts check what models miss.
| Claude alone | Claude with the skill | |
|---|---|---|
| Sound | Not heard | Voice, music and effects reviewed; loudness and clicks measured |
| Motion and timing | Not seen | Slowed shots, frozen frames and uneven motion measured |
| Frames | Only stills you extract | Every frame: 720 frames in a 30-second video at 24 fps |
| Story | Only the script | A reviewer who never saw the script retells the story |
| Output | A description | A verdict and a list of fixes with timestamps |
Problems the review finds in AI videos
AI video tools make mistakes that are easy to miss in a quick look and obvious to viewers. These are the problems the skill is built to find, and how it finds each one.
- in a 30-second video at 24 fps. The skill checks every frame.
- 720 frames
- in a 30-second video at 24 fps. The skill checks every frame.
- is the usual length of an AI glitch. Spot checks miss them.
- 1 to 4 frames
- is the usual length of an AI glitch. Spot checks miss them.
- is the loudness target for social and YouTube video. LUFS measures perceived loudness.
- -14 LUFS
- is the loudness target for social and YouTube video. LUFS measures perceived loudness.
Slowed or frozen shots
A clip is slowed down or held still to fill time or fit the music. The skill finds slowed shots from the share of near-duplicate frames, and frozen shots from frames that do not change.
A character that changes between shots
A face, hairstyle or outfit drifts from one shot to the next, or a different person appears. Each frame is compared with the character's reference sheet.
Audio clicks and uneven loudness
Clicks at edits, music that drowns the voice, and a mix that is too quiet or too loud for the platform. These are measured from the audio file, not guessed.
Text that is cut off or too small
Captions that cover a face, words cut at the edge of a box, and text too small to read on a phone. GPT checks still frames and phone-size copies.
A story that does not come across
A reviewer who has not seen the script retells the video in five sentences. If the retelling misses the main character, what that character wants or how the story ends, the story needs work.
How the Claude video review works
No single AI model sees and hears everything in a video. Models judge the overall viewing experience well but miss short clicks and one-frame glitches. Measurements catch those but cannot judge a story. The skill uses both and confirms each problem before Claude fixes it.
- 1
Measure the audio and picture
Tools: ffmpeg, Python, OpenCV and Whisper
Scripts measure the file directly, so short problems that models miss are still found:
- Overall loudness, against a target of -14 LUFS, the standard unit for perceived loudness.
- The highest audio peak, with a limit of -1 dBTP (decibels true peak): one decibel below the digital maximum.
- Clicks, silences, black or frozen frames, and repeated frames that make motion stutter.
- Sudden jumps where two clips meet.
- The spoken words compared with your script.
It also builds a contact sheet, a grid of frames for a quick visual check.
- 2
Check every frame
Tool: a vision model, one frame at a time
A vision model receives each frame, the frame before it, and a reference sheet for each character in the shot, an image showing how that character should look. It checks:
- Who appears in the frame.
- Whether each face matches its reference sheet.
- Whether each mouth is closed or appears to be talking.
- Whether hands or faces look malformed.
- Whether any text is garbled.
- What changed suddenly since the previous frame.
- 3
Get independent reviews from Gemini and GPT
Gemini reviews the video with sound. GPT reviews still frames.
Gemini watches the full video with its audio and returns findings with timestamps. GPT, from a different company, inspects still frames. Neither reviewer sees the script or how the video was made, so they judge only what a viewer would see and hear.
- 4
Confirm the problems and decide
Claude checks the evidence
Claude confirms a problem when a measurement and a reviewer point to the same moment, when two independent reviews report it, or when the extracted frames show it. The verdict is SHIP, FIX, REJECT or INCONCLUSIVE, and each fix names the step to change: script, voice, generation, edit, mix or export.
Seven passes in the independent review
Gemini and GPT examine the video in seven passes. Pass B uses the 16-group checklist further down this page, and pass G reviews the video from seven viewer and craft perspectives.
Pass A
First reaction
Gemini watches with sound and describes its reaction and any moments that felt wrong, without a checklist to steer it.
Pass B
Checklist review
Gemini reviews the video against the checklist and gives a timestamp for each problem.
Pass C
Audio only
The picture is removed so the reviewer judges the voice, pronunciation, music and mix on their own.
Pass D
Still frames
GPT inspects random and flagged frames, close-ups of faces and phone-size copies for anatomy, text and layout problems.
Pass E
Story retell
The reviewer watches once and retells the story in five sentences. The retelling is compared with what the video was meant to say.
Pass F
Sound off
Most social video plays muted. The reviewer watches without sound and reports where the story stops making sense.
Pass G
Persona panel
Seven reviewer personas assess the video from different viewpoints. Each says whether they would keep watching, names three strengths and three problems with timestamps, and gives a score out of 5.
The reviewers never see the script or the build history. The Claude session that made the video knows what it intended, so it would judge the plan rather than the result. Claude coordinates the review and makes the fixes; independent models do the judging.
Seven reviewer personas
Pass G reviews the video from seven points of view, because a film editor and a casual viewer notice different problems.
01
Target viewer
The audience you describe in the brief, watching on the platform you name.
02
Social media scroller
Decides in 3 seconds whether to keep watching.
03
Film editor
Pace, how long shots are held, where cuts fall, transitions and direction of movement.
04
Animation director
Acting, eyes, timing, staging and whether characters stay consistent.
05
Sound designer and composer
Sound layers, silence, where the music starts and builds, and playback on a phone speaker.
06
Experienced critic
Has seen thousands of AI videos and names every cliche and sign of AI generation.
07
Personal video viewers
For birthday or wedding videos, the person shown in the video and the person receiving it.
AI video quality checklist: 16 groups
Each group is checked in two ways: as a viewer experiences the video, and as an editor inspects how it was made. Every item ends as PASS, FAIL, UNKNOWN or N/A, so nothing is skipped without a record.
- 0Match to the video brief
- 1Story and comprehension
- 2Hook and first frame
- 3Character and design consistency
- 4Acting and expressions
- 5Motion and physics
- 6Staging, camera, light and colour
- 7Pace and editing
- 8Transitions and continuity
- 9Voice and sound design
- 10Music
- 11Lip sync and audio-video sync
- 12Text, captions and graphics
- 13Cliches and signs of AI generation
- 14Ready for the publishing platform
- 15Persona verdicts
Examples of checks
- Speech is no more than 45 milliseconds ahead of the mouth movement or 125 milliseconds behind it, checked by stepping through the frames around a spoken b, p or m.
- Every shot plays at normal speed. Shots are not slowed down or frozen to fill time.
- There is no unplanned silence longer than 0.7 seconds.
- Overall loudness stays within one loudness unit (LU) of -14 LUFS.
- The highest audio peak stays at or below -1 dBTP.
- Impact sounds can still be heard on a phone speaker.
- Music starts when the story changes or a character reacts, not at a random cut.
- Cuts follow Walter Murch's priorities in order: emotion, story, rhythm, where the viewer's eye is looking, direction of movement on screen, and continuity of details between shots.
Problems that block publishing
- A first-time viewer cannot retell the story.
- A character's mouth moves in a silent shot.
- A face changes suddenly at a cut.
- A shot is slowed down or frozen to fill time.
- The audio has audible clicks.
- Motion stutters.
- Lip sync falls outside the timing limits above.
- A caption changes the meaning of what is said.
Example repair list
Each review writes report.md and report.json in a .video-review folder next to your video, together with the evidence: frame strips, audio clips and a contact sheet with the problems marked. Claude works through the repair list and renders the video again.
repair_list.md · example-explainer.mp4 · 1,032 of 1,032 frames checked at 24 fps
| # | Shot | Time (s) | Problem | Fix | Severity |
|---|---|---|---|---|---|
| 1 | 02 | 5.00 to 9.25 | Character's mouth moves as if talking in a silent shot | Add "lips closed" and a physical action to the prompt; generate the shot again | major |
| 2 | 03 | 10.00 to 10.50 | Character does not match the reference sheet | Generate again with the reference sheet attached | major |
| 3 | 04 | 15.00 to 20.00 | 25% of frames are near-duplicates: the shot was slowed | Generate a longer take and keep it at normal speed | major |
| 4 | 06 | 26.40 | Audio click at an edit | Add a short crossfade at the cut | blocker |
| 5 | 07 | 31.10 to 33.00 | Caption covers the speaker's face | Move the caption lower, clear of the face and the platform buttons | minor |
| 6 | 09 | 39.46 to 43.00 | Picture frozen for 3.5 s under the title | End on moving footage and trim the music to the picture | major |
Add automatic video review to your Claude workflow
After you install the skill, add this instruction to your video project's CLAUDE.md file or to the Claude skill you use to make videos. Claude then builds the video, reviews it, fixes the confirmed problems and reviews it again, for up to three rounds, before it hands you the file with its report.
Before delivering any video, run the video review skill. If the verdict is SHIP, deliver the video. If it is FIX or REJECT, fix every confirmed blocker and major problem, render again and review again, up to three rounds. If it is INCONCLUSIVE, find what the missing check needs, resolve it and run the review again. If the video still does not reach SHIP, show me the report and explain what needs to happen next.
Works with any finished video
- Remotion and HyperFrames explainers
- ffmpeg edits and podcast clips
- Seedance, Veo, Kling and Runway shots
- HeyGen and other AI avatar videos
- Course and lecture videos
- Instagram Reels, YouTube Shorts and LinkedIn video
Video review skill vs video watcher skills
Most Claude video skills help Claude describe a video by giving it a transcript and a few frames. This skill decides whether the video is ready to publish and lists what to fix.
| Video watcher skills | Claude Video Review Skill | |
|---|---|---|
| Purpose | Summarise a video or answer questions about it | Decide SHIP, FIX or REJECT |
| Sound | Usually the transcript only | Full audio review, with loudness and clicks measured |
| Frames | A sample of frames | Every frame, compared with the reference sheets |
| Story and acting | Not reviewed | Story retell, sound-off review and an animation director persona |
| Output | A summary | A repair list with timestamps and a fix for each problem |
Requirements
The full review on this page uses Gemini for the video with sound, GPT for still frames, a vision model for the every-frame check, and the free measurement tools below. Gemini and the measurement tools are enough to start.
- Required
Claude Code
Runs the skill, confirms the problems and fixes the video.
- Required
Gemini API key
Gemini watches the full video with its sound.
- Required
ffmpeg, Python 3.10+ and Whisper
Run the measurements and transcribe the speech. All free.
- Recommended
Codex CLI with an OpenAI account
GPT reviews the still frames as a second, independent reviewer.
- Optional
A vision model for frame checks
Checks every frame against the character reference sheets. The included script uses Jev by TypeSafe.
How to install the Claude Video Review Skill
- 1
Clone the skill
Copy the repository into your Claude skills folder. It is one folder under the MIT licence.
- 2
Install the requirements and add your API key
Install ffmpeg and the Python requirements, then save your Gemini API key in the skill folder.
- 3
Ask Claude for a review
For example: "Review final.mp4 with the video review skill. Audience: founders on LinkedIn."
The install guide by email has the exact commands for macOS and Linux, a brief template, the full checklist and an example report.
Source code on GitHubFrequently asked questions
Can Claude watch videos?
No. Claude reads text and images, but it cannot play a video with its soundtrack. With the Claude Video Review Skill, Claude sends the video to Gemini, which watches it with sound and returns timestamps, asks GPT to inspect still frames, runs measurements on the file, and then checks each reported problem against frames it extracts itself.
Can Claude analyze a video file?
Claude can analyse frames and transcripts that are given to it. To analyse the moving picture and the sound together, the skill uses ffmpeg to extract frames, audio and measurements, sends the full video to Gemini, and gives Claude the results to confirm.
How do I check a video I made with Claude Code?
Install the skill and ask Claude: "Review final.mp4 with the video review skill." Tell it who the audience is and which platform the video is for. For a story video, add a short story summary. Claude runs the review and returns a verdict with a repair list. Add the review instruction to your build skill, and Claude runs the review before it delivers any video.
Does it work with Remotion, HyperFrames or ffmpeg edits?
Yes. The skill reviews the finished file, so it works the same way whether the video was made with Remotion, HyperFrames, ffmpeg, Seedance, Veo, Kling, Runway, HeyGen, a screen recording or a camera.
Why not ask Gemini to review the video on its own?
Gemini judges the overall viewing experience well, but it samples the video at a few frames per second and compresses the audio, so a click lasting a few milliseconds or a glitch lasting two frames can pass unnoticed. The skill adds measurements, a check of every frame and a second reviewer from a different company, and confirms each problem before it is fixed.
Why doesn't Claude review its own video?
The Claude session that built a video knows what it meant to make, so it tends to judge the plan rather than what is on screen. In this skill Claude coordinates the review and makes the fixes, and models with no knowledge of the build do the judging.
Does it handle long videos?
Yes. Gemini watches long videos in overlapping 3-minute sections at 2 frames per second, plus one full viewing for the overall reaction, and watches Reels and Shorts at 5 frames per second. The separate frame check still examines every frame.
Is the Claude Video Review Skill free?
Yes. The skill is free and open source under the MIT licence. You pay only for your own use of the Gemini API and, if you add them, OpenAI and a vision model for frame checks.
Get the install guide
Enter your email and we will send the install guide, a brief template and an example report. Setup takes about ten minutes, and every video Claude makes after that can be reviewed before it is published.