This is the version to keep open while you edit. Open it when you have a raw clip in front of you and work down the list in order: trim first, then transcribe, only then animate and render. The order is fixed, and it decides the result — do not skip a stage.
0. Before you start — what you need
Check all of this is in place before you drop in the first file.
- Install Claude Code and pick the variant: the desktop app (simpler interface) or the code editor (shows every project file). For editing, the app is enough.
- Have links ready to two repositories Claude Code will load skills from: Video Use (trimming, and optionally animation through Remotion plus render) and HyperFrames (a second way to build animations — cards, captions, transitions).
- Pick a transcription tool: OpenAI's Whisper (runs locally, free), a local tool, or the ElevenLabs service. All three work fine.
- If you go with ElevenLabs, create an API key (a private string that authenticates your account) and paste it into a separate
.envsettings file, never into the chat. - Have one short raw clip ready for the first run. Treat the first clip as learning the tool, not as the finished result.
1. Load the skills into Claude Code
- Paste both repository links (Video Use and HyperFrames) into Claude Code.
- Ask it to review the repositories and load the skills, for example: "review these repositories and load the skills I need, so I can hand you a raw file and you edit it."
- Confirm Claude Code has acknowledged loading the skills — only then does it know what both tools can do.
2. Trim the footage
- Drop in the raw clip and run the trim through Video Use.
- Review what the tool proposes to keep and cut (false starts, stumbles, filler, repeated takes, silence).
- Answer the taste calls it asks you about — for example, whether to leave a trailing half-word in as a natural breath or cut it.
- Check that the cuts sit tight against word boundaries with a small margin, so they sound clean.
3. Transcribe with timestamps
Do this step after trimming and before animation — skip it and the animations will not land on the right words.
- Run the transcription with whichever tool you picked (Whisper, a local tool, or ElevenLabs).
- Check that the transcript carries a timestamp on every word — the second, down to a fraction, at which that word is spoken.
- That timestamp is what lets an animation trigger at the exact moment you say the thing it illustrates.
4. Plan the animation before it gets built
- Describe exactly what should appear and when — for example, a card on the left with karaoke-style captions sliding in, or an animation that shows the error cuts themselves.
- Switch Claude Code into plan mode, so it lays out the plan first: which card it will build, what goes on it, at which second it appears — instead of starting work right away.
- Review the plan and add anything missing at this stage (an outro scene, for instance).
- Approve the build only once the plan checks out. A plan catches a misunderstanding before the tool renders something wrong, and building animation costs tokens and time.
5. Render and fix the first pass
The first pass is never perfect — that is normal, so budget a round of fixes.
- Render the first pass and watch it for problems (a card covering a face, an unwanted grid in the background, a different crop at the end).
- Describe the fixes concretely, for example: "shrink this card and trim its right edge so it stops covering the face," "remove the grid from the whole clip."
- Reach for the built-in timeline when you need it — drag, shorten, or delete elements by hand, and the change reflects into the code, which Claude Code will pick up.
- Ask the tool for screenshots of each scene, so it checks the result itself instead of declaring it done blind.
6. Lock in the style for the next clips
- Once you have edited a few similar clips, write their style down in a separate guidelines file.
- Point back to that file on every clip after that — that is when "drop in the raw file and the rest happens on its own" starts being true.
Watch for
- Keep the API key in
.env, not in the chat. In the chat it would sit in your history — a bad habit from a security standpoint. - Do not skip transcription before animation. Without timestamps there is no way to trigger an animation at the exact moment a word is spoken.
- Do not approve animation blind. The planning stage limits the risk of the tool going the wrong direction and burning tokens and your time.
- Expect real token spend. In the example described here, the whole run — prompts, successive animation passes, generated files — used roughly 238,000 tokens. The more precisely you steer the tool, the less gets wasted.
- Do not treat this as a button that does everything for you. It shifts the work: manual trimming and animating gets replaced by precise instructing and correcting. The real time savings show up once you're working through repeat clips.