If you have ever pasted a still from a movie into Claude and asked “make my video feel like this,” you already know the result. You get mood words back. You get adjectives. You do not get a shot list you can hand to an editor. That gap is what Claude video reference prompts are meant to close, and it is why the structure of the reference matters more than the reference itself.
vague ref structured ref
--------- --------------
"Blade Runner vibe" scene 04 / 00:12-00:19
cut every 2.1s average
wide hold, then push in
practical haze, no fill
hook lands at 00:03
| |
v v
adjectives shot list +
for the model cut rhythm
Why Claude Needs Better Video References
A model cannot watch a film the way you do. When you give it a title or a poster, it draws on whatever text about that film happens to be in its training. Sometimes that is useful. Usually it is generic, and it is always a summary rather than a breakdown.
What a model can actually work with is structure: how long each shot runs, what changes between cuts, where the hook lands, how the camera moves, what the light does. Give it those, and the output changes from “moody, cinematic, warm” to a sequence of shots with timings and purposes. That is the whole idea behind a cut-by-cut reference: write films down as blueprints instead of posters, so an agent stops guessing and starts planning.
This matters most in pre-production. Once you are actually editing, you have footage and constraints, and the model can react to what exists. Before that, you are asking it to imagine. Better references give it something concrete to imagine with.
The Anatomy of a Strong Claude Video Reference Prompt
A reference prompt has five parts. Skip any of them and the model fills the gap with something generic.
- Source. Which film, which scene, which timestamp range. Be narrow. A whole movie is not a reference.
- Structure. Cut count and rough durations. “Five cuts across twelve seconds, longest cut four seconds.”
- Form. Framing, movement, lens behavior. Static wide. Slow push. Handheld follow. Locked-off insert.
- Look. Light, color, texture. Practical haze. Hard key from one side. No fill. Cool shadows.
- Output shape. What you want back. A shot list with columns. A beat sheet. A caption plan. Name the format.
Here is a weak prompt next to a strong one.
Weak:
"Make my product video feel like a Denis Villeneuve
scene."
Strong:
"Use scene 04 (00:12-00:19) as the structural
reference. Five cuts, average 1.8s, longest 4s,
static wide to slow push. Practical haze, hard
side key, no fill. Output a 5-row shot list with
columns: shot, duration, framing, movement, purpose,
and mark where the hook lands."Notice the strong version names a purpose for every shot. That column is what turns a list into a plan, and it is the column most people forget.
If you want the same discipline applied to your own footage rather than a film, Creator OS ships an open-source set of Claude Code skills for that at github.com/kevinbadi/open-edits. It takes a raw talking-head clip and renders an animated short locally with Python and ffmpeg. The reference work and the edit work are different jobs, but they use the same habit: describe the sequence, not the feeling.
There is a walkthrough of how that editing stack fits together on the launch video, and more build-along videos on the KevBuildsApps YouTube channel.
Recipe Prompts for Claude Code, Codex, and Cursor
Agent tools behave differently from chat. In Claude Code, Codex, or Cursor the model can read files, write files, and run commands, so your prompt should point at a bundle instead of describing one. A workable shape:
Read refs/scene-04.md.
Produce storyboard.md with one section per cut:
- shot number
- duration in seconds
- framing and movement
- light and texture notes
- purpose in the sequence
Then produce shotlist.csv with the same rows.
Do not invent cuts that are not in the source file.
If a field is missing from the source, write UNKNOWN.Two rules make this reliable. First, tell the agent not to invent. Second, give it a marker for missing data, so you can see the holes instead of receiving confident filler. Once the bundle is a file in the repo, the reference is versioned, diffable, and reusable across projects.
Creator OS ships 14 Claude Code skills you can install next to a reference bundle like this. Run npx @creatoros/cli@latest init and you get skills such as post-shortform, post-longform, schedule-posts and analytics, so the same repo that holds your shot list can also publish the finished cut. The full list is in the CLI docs.
Turning Blueprints Into Storyboards and Shot Lists
A blueprint describes someone else’s film. A storyboard describes yours. The translation step is where most of the value sits, and it looks like this:
- Pull the cut list from the reference into a plain table.
- Replace the subject of each cut with your subject, keeping the duration and movement.
- Rewrite the purpose column for your actual goal.
- Mark which cuts are load-bearing, meaning the edit breaks if you lose them.
- Hand the table to the editor, or to an agent, as the spec.
The purpose column is worth an example. If the reference cut exists to establish scale, your version might exist to establish that the product is small enough to hold. Same duration, same framing, completely different job. That is a decision a mood board cannot make for you.
If your finished pieces are vertical and short, the same logic applies to how you cut a long video into clips. There is a worked process for that in repurposing a long YouTube video into Shorts with Claude, and the reference structure feeds directly into it.
Matching References to Hooks, Pacing, and Platform
Different platforms reward different rhythms, and your reference should match the destination, not your taste. A few practical pairings:
| Destination | Reference shape | Why |
|---|---|---|
| Short-form feed | Fast cuts, hook inside 3 seconds | The first seconds decide whether the rest is watched |
| Long-form YouTube | Longer holds, slower push | Viewers settle in and tolerate setup |
| Product demo | Locked-off inserts between motion | Detail reads better when the frame is still |
| Story or casual post | Single continuous take | Reads as unpolished on purpose |
If you publish across several networks, you can keep one reference bundle and swap the pacing rules per destination. Creator OS posts to Instagram, TikTok, YouTube, X, LinkedIn, Facebook and Threads from one place, and each network applies its own rules. Details on what changes per platform live in the post options docs. If you are scheduling the results of all this pre-production, the same applies to Instagram specifically, covered in scheduling Instagram posts with Claude.
Workflow: From Reference Bundle to Finished Edit
Here is a full pass you can run this week.
1. Pick 3 reference scenes. Narrow ranges only.
2. Write one markdown file per scene in refs/ with
cut count, durations, framing, movement, look.
3. Ask the agent for storyboard.md and shotlist.csv.
4. Cut the reference durations to match your format.
Trim the hook to land before 3 seconds.
5. Shoot to the shot list. Do not add shots on set.
6. Upload the footage once and reuse the media id
across every platform you publish to.
7. Check caption length per platform before posting.
8. Schedule, then read per-post analytics after 48
hours and compare pacing against performance.Steps 6 through 8 are where a publishing layer earns its place. With the Creator OS CLI you can upload once and reuse the returned media id across platforms:
creatoros media:upload ./final-cut.mp4
creatoros validate:post-length --text "your caption"
creatoros posts:create --text "your caption" \
--platforms instagram,tiktok,youtube \
--media med_abc123 \
--scheduledAt 2026-03-04T18:00:00Z --timezone America/Toronto
creatoros analytics:posts --platform instagram --sortBy engagement
Every account connection, plan limit and command is documented at https://www.creatoros.ca/docs, including the CLI reference. If you run the whole thing by hand in an agent, the CLI docs are the fastest starting point.
Common Mistakes When Prompting Claude With Film Refs
- Referencing a whole film. A 90-minute movie gives no usable rhythm. Use a scene and a timestamp range.
- Describing mood instead of motion. “Dreamy” tells the model nothing it can schedule. “Slow push, 4 seconds, no cut” does.
- Forgetting the output format. Without columns named, you get paragraphs. Name the table.
- Letting the agent invent cuts. Add an explicit instruction not to, and a marker for missing fields.
- Copying the reference exactly. You want its rhythm, not its content. Rewrite the purpose column for your own goal.
- Ignoring platform rules. A rhythm that works in a feed may fail in a longer video, and vice versa.
- Never closing the loop. Compare pacing against the analytics you get back, or you are guessing twice.
One more note for anyone running agents rather than chats: Creator OS ships a hosted MCP server and a CLI, so an agent can run your accounts instead of just describing them. The server is at mcp.creatoros.ca/mcp with nothing to install, and the setup steps for each app are in the MCP docs. Read-only connections only see read tools, and destructive tools are marked so the AI app asks before doing anything permanent. For an end-to-end walkthrough of an agent posting and replying on your behalf, see posting to every social platform from Claude.
Get Started With Creator OS
Better references are only half the loop. The other half is publishing what you planned and reading what happened. Creator OS handles both, from the web app, the iOS app, the REST API, the CLI, or a hosted MCP server for Claude, ChatGPT, Claude Code, Cursor, Codex, Windsurf and VS Code. API keys, the MCP server, the CLI and the agent skills are included in every plan.
The Creator plan is $19.99/month or $59.99/year, with up to 8 connected accounts: 7 social platforms plus Skool. Higher tiers add more sets of socials for teams running multiple brands.
Start at creatoros.ca/sign-up, then build your first reference bundle and hand it to an agent. If you want to see how other builders wire the same tools together, the walkthroughs live on the KevBuildsApps YouTube channel.