
We have had Claude edit the same raw recording five times, each time in a different look. Our founder rejected two of those edits. One of the three that went out was built from a one-page look spec: white, near-black, one gray and one orange, three fonts, four kinds of transition, and one sound that only ever plays with one effect.
Working from that page, Claude gave each of 21 spoken lines its own motion technique, about 20 in all, and the edit still reads as one style. The steps below show how to write that page for your own videos, save it where Claude Code reads it every session, and switch styles with a reference video.
Short answer: write your editing style as exact values on one page, save it as STYLE.md in your project folder, and import it from CLAUDE.md so Claude Code loads it every session. Approve stills and a short motion test before the full render. To change styles, give Claude a reference video to measure and turn into the same template.
- The page covers palette, accent rules, type, captions, layout, motion, transitions, one signature device, sound, footage and a banned list.
- Pick one accent color and give it jobs. In our example an orange #D97757 line wipes scenes and strikes mistakes, and only that line makes its sound.
- Claude's vision works on still images, so a reference video becomes a set of stills first.
- Claude Code needs a paid Claude plan, from $20 a month ($17 a month billed yearly). The fonts in our example are free.
Why an AI editor needs the style written down
A human editor remembers your style between projects. Claude Code starts each session with a fresh context window, according to Anthropic's documentation, and an instruction given only in chat can drop out when a long session is compacted. The documented fix is a file Claude reads at the start of every session.
The same docs say Claude follows specific, concise instructions more consistently. So write values. "Clean and modern" can come back different every time, while "white #FFFFFF, ink #0E0E0E, no shadows" comes back the same.
Before you start
- Claude Code on a paid plan. Pro is $20 a month, or $17 a month billed yearly, and Max starts at $100 a month (pricing). The free plan does not include Claude Code.
- A project folder with a subfolder called references, holding three to five videos whose look you like.
- Your fonts. Our example uses Bricolage Grotesque, Inter and JetBrains Mono, free from Google Fonts under the SIL Open Font License.
- The editing tools, which Claude installs with you: FFmpeg for cuts and still frames, and HyperFrames by HeyGen, which turns HTML into video and needs Node.js 22 or newer.
- About an hour for your first guide, and less for the next style.
Step 1: Split the content from the look
The content layer stays put whatever the style: the takes you kept, your words, the cleaned voice and the order of the lines. The look layer is what the guide controls: layout, type, captions, transitions, motion, 3D, color, sound effects and pace.
When we asked Claude for new versions of one finished video, the brief came down to one line: reuse the assets, never the look. Put that rule in your guide too, so a new style never re-picks your takes or rewrites your words. This step takes five minutes.
Step 2: Gather the decisions you already like
- Collect the videos. Pick ones in your own format: a vertical talking video teaches you more than a film trailer.
- Turn them into stills. Claude's vision reads JPEG, PNG, GIF and WebP images, so Claude Code uses FFmpeg to pull a frame at every scene change.
- Let Claude interview you with the prompt below. It offers options with real values, one part at a time, and writes down the one you pick.
Help me write a one-page style guide for my videos, saved as STYLE.md in this folder.
First pull a still at every scene change from each video in the "references" folder.
Then interview me one question at a time: palette, accent rules, type, captions, layout, motion, transitions, signature device, sound, footage rules, banned list.
For each, offer 2 or 3 options with exact values (hex codes, fonts and weights, pixel sizes, frame counts, dB levels) based on the stills, and write down the one I pick.
Warn me if any text color is below 4.5 to 1 contrast for small text or 3 to 1 for large text (WCAG formula).
Keep the file under 80 lines. At the end, make one still in the new style.
What you should see: questions with options such as "white #FFFFFF with ink #0E0E0E" or "warm white #F8F6F2 with ink #1A1A1A", and a STYLE.md that grows as you answer. Plan on 20 to 40 minutes. If Claude writes the whole file at once, repeat "one question at a time".
Step 3: Fill in the one-page template
The template below is filled in with our worked example: a black-and-white style with one orange accent and a thin orange line through every scene. Copy it and change the values.
# STYLE.md: black and white, one orange cut line
Format: vertical 1080x1920, 30 fps
PALETTE
paper #FFFFFF (every graphic scene), ink #0E0E0E (type, rules, crop marks)
gray #6B6B66 (small secondary text), hairline rgba(14,14,14,0.14)
accent #D97757: cut line, strikes, highlights, counters,
full orange fields on the 4 or 5 biggest moments
No shadows on graphics.
ACCENT RULES
On orange: type in ink, highlights in white, the cut line turns white.
Key caption word (1 in 4 or 5): orange on white, white with an orange
underline on footage, ink with an ink underline on orange.
TYPE (no italics, no serif)
Headlines: Bricolage Grotesque 800, width 75, all caps,
tracking -0.03em, line height 0.86, 180 to 340 px
Interface and lists: Inter 500 and 600
Numbers and timecodes: JetBrains Mono 500
CAPTIONS
Inter 700, 54 px, sentence case, 2 to 5 words at a time.
Each phrase rises 8 px and fades in over 5 frames. No box, no outline.
Ink on white, white with a soft shadow on footage.
Never on my face. Hidden when on-screen type says the same words.
LAYOUT
12 columns, 60 px side margins. Safe area: x 60 to 1020, y 140 to 1700.
Keep x 940 and up clear from y 1000 to 1700 (the app's buttons).
Measure every word before drawing it.
Crop marks on every framed thing: 28 px corner ticks, 3 px ink.
MOTION
In: cubic-bezier(0.16, 1, 0.3, 1), 8 to 12 frames
Out: power3.in, 6 to 8 frames, toward where the next thing enters
Shape changes: cubic-bezier(0.45, 0, 0.55, 1), 12 to 16 frames
Held elements scale slowly from 100% to 104%.
Picture lands 2 to 3 frames before its word. Overshoot: 3 in total.
TRANSITIONS (only these four)
Hard cut on a word's first sound (default), cut-line wipe (7 frames),
frame morph, match cut. No crossfades, whips, blur or smear.
SIGNATURE DEVICE
A 4 px orange cut line: wipes scenes, plays the playhead,
strikes mistakes, cleans the voice, applies styles.
Its "tsk" sound belongs to it alone.
One technique per spoken line, chosen to show what the line says.
SOUND
One effect per visual landing, each at least 14 dB under my voice.
Planned quiet after big hits. No music in the master.
-12 LUFS integrated, true peak -1.5 dBTP or lower.
FOOTAGE
Natural color, always sharp. Scale down from the 4K source, never up.
No blur, depth of field or skin smoothing on my face. Nothing over it.
BANNED
Glitch, RGB split, grain, letterbox, black-and-white grade on footage,
ghost trails, italics, other tools' logos, and my other styles'
signature moves: paper cards, dot grids, rounded panels, word flashes.
Palette and contrast
Black and white describes the graphics here. The footage stays in natural color: we rejected another version of the same video that used a dark film grade with grain and black-and-white inserts. We checked every text color with WCAG's contrast formula. That web standard asks for 4.5 to 1 for normal text and 3 to 1 for large text:
| Text on background | Contrast | Use it for |
|---|---|---|
| Ink #0E0E0E on white | 19.3 to 1 | All text |
| Gray #6B6B66 on white | 5.4 to 1 | Small secondary text |
| Ink on orange #D97757 | 6.2 to 1 | All text on orange fields |
| Orange on white, or white on orange | 3.1 to 1 | Lines, fields and big words only |
| Gray on orange | 1.7 to 1 | Never |
Accent rules
An accent can fill the screen. Our first plan kept orange to lines and highlights. After a review we added full orange fields on five big moments, so the video reads as orange and white even as a thumbnail. Each background then needs its own accent rule: an orange line vanishes on orange, so there it turns white.
Type, captions and layout
Give each font one job. Bricolage Grotesque, by Mathieu Triay, runs from weight 200 to 800 and width 75 to 100, and we use it at its heaviest and narrowest. Inter, by Rasmus Andersson, handles captions and screens. JetBrains Mono is monospaced, so counters do not wobble. If you recreate an app's screen, use your own fonts: a copy of a company's interface in its own typeface reads as official material.
The right edge and bottom of a vertical video sit under the app's buttons and caption, so the safe area keeps text clear of them. Ask Claude to measure faces too. On our example edit it checked caption against face on 614 frames: the smallest gap was 73 pixels, with no overlaps.
Motion, transitions and the signature device
At 30 frames a second, 8 to 12 frames is a quarter to two-fifths of a second, and a frame count leaves nothing to interpret. Paste easing values into cubic-bezier.com to see the curve. With four transitions allowed, new ideas go inside scenes.
Then pick one element that appears in every scene and give it several jobs. Ours is a 4 px orange line, the editor's razor: it wipes scenes, plays the playhead on a scrubbing timeline, strikes through mistakes, sweeps across the frame when the voice switches from raw to clean, and applies each new look. Around it, each line got its own technique, from a credits roll to a split screen over a live spectrogram, tied together by the line and its sound.
Sound, footage and the banned list
Our example names its effects (tsk, clack, tick, thock, whirr, brake, slide, flap, riffle, drop, air, chime). The finished mix placed 72 effects, one per visual landing, with the loudest 17.6 dB under the voice. FFmpeg's ebur128 filter, an EBU R128 loudness scanner, measured it at -12.0 LUFS integrated with a true peak of -1.7 dBTP.
The footage rules came from a failure. On a version built as a 3D studio, the face went soft in places. Claude traced it to four causes: the footage sat on a 3D panel with depth of field and motion blur, the clips were pulled at 640 by 360 and scaled up, the panel frames were JPEG, and the phone copy was 2.3 Mbps. Each cause became a rule. The banned list also names our other styles' signature moves, so the styles stay apart.
Step 4: Save it where Claude Code reads it every session
- Save the file as STYLE.md in your project folder.
- Import it from CLAUDE.md. Claude Code reads CLAUDE.md at the start of every session, and an @ with a file path pulls that file in too. Create CLAUDE.md if there is none and paste the lines below.
- Check that it loaded. Start a new session and run /context: CLAUDE.md should be listed under Memory files. Then ask "What is my accent color?" and check the answer.
Follow every rule in @STYLE.md on every video edit.
If a shot cannot follow a rule, stop and ask me before you break it.
A new style never changes my takes, my words or the order of my lines.
Keep it short. Anthropic's docs suggest under 200 lines per CLAUDE.md file, and an imported file counts too, because it loads along with CLAUDE.md. Our example is about 60 lines.
Step 5: Run the edit with your style guide
Open Claude Code in the project folder, add your raw video and paste the prompt below: our standard edit prompt with one line added at the end. Claude shows you the takes it kept before the final render.
I want you to edit my talking video. Here is the raw file: [add your video file here]
1. Turn every word I say into text, with the time of each word.
2. If I said a line more than once, keep the best take. Pick the one with no stumbles, nothing cut off, and where I look at the camera. Cut every mistake, filler word and long pause.
3. Clean my voice. Remove the echo and the room noise.
4. Add captions that match my words exactly, a few words at a time. Never put them over my face.
5. Add an animation for each line that shows what I am saying.
6. Export it as a vertical 9:16 video, 1080 by 1920.
Before you start, tell me which tools you need and help me install them. Show me which takes you kept before you make the final video.
Follow every rule in STYLE.md for the look and the sound.
Prefer a one-click copy? The same prompt, with every tool link, is on our free prompt page. Add the STYLE.md line at the end.
Claude Code runs the tools: Whisper by OpenAI for the transcript, MediaPipe by Google to check where you look, ElevenLabs Voice Isolator for the voice (paid, with a free tier for personal use), FFmpeg, HyperFrames, Blender for 3D and Robust Video Matting to cut you out of the background.
Step 6: Approve stills and a motion test before the full render
The full render is the slowest step, so check the style first, in stills and in motion.
Before the full render, show me the style:
1. A sheet of 12 to 16 stills, one per line, at the moment each visual lands.
2. A 10 to 15 second motion test of the busiest part.
Under each still, list any rule in STYLE.md it breaks. Wait for my OK before you render the whole video.
On our example, the notes from this review were small and exact, such as a bigger text-based edit and no near-blank frames after one scene. All of them went in before the full render.
Step 7: Check the finished video against the guide
Check the finished video against STYLE.md, pass or fail per part:
1. Colors in the graphics (not my footage) that are outside the palette.
2. The smallest caption-to-face gap in pixels, and any text outside the safe area.
3. Every scene change and its type. Flag anything outside the allowed four.
4. Integrated loudness and true peak with FFmpeg's ebur128 filter, and how far under my voice the loudest effect sits.
5. Any shot where my face was scaled up, blurred or covered.
Ask Claude to measure with scripts and real pixels: Anthropic's documentation says Claude's sense of where things sit in an image is approximate. Then watch it yourself, with sound, on your phone.
Switch to a new style with a reference video
Once the template exists, a new style is new values in the same slots. A video whose look you want is the quickest source.
- Pick a reference you can learn from, such as your own older edit or a public video you study. Take its layout, pacing and color logic, and leave its footage, music, logos and fonts alone.
- Turn it into stills. FFmpeg scores every frame from 0 to 1 for how likely it is to start a new scene, and its documentation calls a threshold between 0.3 and 0.5 "generally a sane choice".
- Have Claude fill in your template from measurements. Font names read off a frame are guesses, so ask for the closest free matches and pick by eye.
- Test it on your own footage with one still and a 10-second motion test before you commit.
Here is a video whose style I want for my next edit: [add the reference video file]
1. Pull a PNG still at every scene change with FFmpeg (scene threshold 0.4), plus a contact sheet.
2. Measure the style from the pixels and timing, not by eye: palette as hex codes, text sizes and positions, caption style, transitions and how often they appear, average shot length.
3. Write it as a new style guide in the STYLE.md format, saved as styles/[name].md. Mark every value you are unsure of.
4. Keep my footage rules and safe area.
5. Make one still and a 10-second motion test on my footage in the new style, and wait for my OK.
Use PNG stills. Anthropic's documentation warns that heavy JPEG compression can make text in an image hard to read, and small type is part of what you are measuring.
Keep a library of styles
- Save each style as its own file in a styles folder and point the @ line in CLAUDE.md at the one in use. Switching styles means changing one line.
- Give every style a banned list that includes the other styles' signature moves.
- If you are stuck between two directions, ask for both. For our black-and-white style, two Claude agents each wrote a proposal for every line, argued over each one, and merged the winners before building. Two competing style guides, argued and then merged, work the same way.

When the style breaks
- Claude drifts from the guide on a long edit. Run /context to check that CLAUDE.md loaded, and move anything you only said in chat into the file.
- Your face looks soft. Check the footage rules: one downscale from the source file, no blur, nothing over the face.
- It looks right in stills and wrong in motion. Adjust the easing and frame counts, then run the motion test again.
What it costs, and other ways to do it
Claude Code needs a paid Claude plan, from $20 a month. The fonts, FFmpeg and HyperFrames (Apache 2.0) are free. The style guide is a text file, and it doubles as a brief for a human editor.
Teams that edit in Descript can keep layout packs, fonts, colors and media in its Brand Studio, on the Business and Enterprise plans. If you would rather have a team build and run a style system for your brand's videos, that is part of our creative production work.
Frequently asked questions
What should a video editing style guide include?
Exact values for the palette (hex codes), one accent color and where it may appear, fonts with weights and sizes, caption rules, the canvas and safe area, motion timing in frames, the transitions you allow, sound levels, rules that protect your footage, and a banned list. Keep it to one page.
How do I find my own editing style?
Collect three to five videos whose look you like, pull still frames from them and write down what they share. Then ask Claude to interview you one part at a time with concrete options, and test the result on one of your own videos before you commit to it.
Can Claude copy the editing style of a reference video?
It can measure one and write it down. Claude Code pulls still frames with FFmpeg, measures the palette, text sizes, caption style, transitions and shot length, and fills in your template. Font names are guesses you should confirm. Learn the decisions from a reference, and leave its footage, music, logos and fonts alone.
Can Claude watch a video?
Claude's vision works on still images (JPEG, PNG, GIF and WebP), so Claude Code turns a video into still frames with FFmpeg and a timed transcript with Whisper, then works from those. It sees a sample of frames, so always watch the final cut yourself.
How long should a style guide for AI be?
One page. Claude Code's documentation suggests keeping each CLAUDE.md file under 200 lines and says specific, concise instructions are followed more consistently. Our worked example is about 60 lines.
Where do I save the style guide so Claude uses it every time?
Save it as STYLE.md in your project folder and add a line to CLAUDE.md that imports it with @STYLE.md. Claude Code reads CLAUDE.md at the start of every session. Run /context to check that it loaded.
Sources
- Claude plans and pricing
- Claude Code docs: CLAUDE.md, imports and /context
- Claude Code docs: work with images
- Claude vision docs: formats, image quality and limitations
- FFmpeg filters: select (scene score) and ebur128
- WCAG 2.2: Contrast (Minimum)
- Google Fonts metadata for Bricolage Grotesque, Inter and JetBrains Mono
- HyperFrames on GitHub
- Descript Brand Studio