The AI Filmmaking Pipeline
Working in Claude
Almost everything we make starts as a conversation with Claude — but not a throwaway chat. We work in Cowork mode (the selector at the bottom-left of the message box, on the Home tab), because a project needs a memory and a place for its files.
Why Cowork, not a plain chat
A normal chat forgets between sessions. Cowork remembers, and that changes how you work.
Memory across chats — inside a Cowork project, context carries over: the story, the characters, the look you've agreed on. You stop re-explaining the project every morning.
Local files — your assets live in the project folder on your machine, where Claude can read and reuse them. For this, work in the desktop app.
Turn Memory on in settings before you start a real project. The whole value of Cowork — context that survives between chats — only works when memory is enabled.
Cowork is no longer a separate tab. It now lives inside the Home tab — pick Cowork in the message-box selector (Chat / Cowork). If you don't see it, update the Claude app. Cowork is also available on web and mobile in beta (Max plans first), but local file access still requires the desktop app.
The clip is the setup for a reliable handoff: switch into Cowork and make the project's context available before you ask Claude to work from its files.
Approved assets get saved in two places: the local Claude folder for the project, and your
normal project folder where you keep every related file. Asset names start with @ — that's the
hook that pulls them straight into a prompt later.
The three file types you'll meet
Three formats carry the whole workflow:
- .md
- — A plain-text (Markdown) file — readable anywhere, rendered nicely in apps. We use it for notes, instructions, and skills.
- Skill
- — A folder of .md instructions (a SKILL.md) holding the rules for one topic — Claude opens it when a task fits and works to that standard.
- .jsx
- — Technically a code file, but we use it as a shotlist — a structured container for shot data (shot numbers, descriptions, timings, prompts) that Claude reads the fields of and pulls from.
A skill isn't magic — it's a pre-written handbook Claude loads at the right moment so it
stays on one standard instead of improvising. A skill might hold the rules for building Seedance
prompts. Not every .md is a skill; sometimes it's just a system prompt you attach to a chat.
One handoff, end to end
Don't ask Claude to “make the prompt” in isolation. Give it the production record first:
| Give Claude | What it anchors |
|---|---|
Scene brief .md | Story beat and shot intent |
Approved references and exact @ Element names | Character, location, and prop identity |
Relevant SKILL.md | Rules for this kind of prompt |
Current .jsx shotlist | Shot number, timing, and continuity |
Tell Claude which files are authoritative; memory helps it recall the project, but it does not choose the latest brief for you. Replace the bracketed values in this request with the shot you are actually preparing:
Review the record in Claude, then save it into the current .jsx shotlist. Cinema Studio does
not ingest that whole file: open the right project and folder, paste the prompt value into the
Prompt Box, and select every elements entry through the @ picker. For a video shot, also carry
over duration. Before you generate, the visible @ tags should match the record exactly.
That's the handoff: project files into Claude, one structured shot record out, then its prompt and Elements into Cinema Studio. The next lesson sharpens the prompt itself.