Six agents, one stage each
A video pipeline from raw idea to edit plan — and the gap between rules written in a prompt and rules a tool can enforce
Six prompts, each owning one stage of making a video, chained so every stage consumes the approved output of the last. The discipline is the good part — one job per agent, a human gate after the first two, and a standing ban on inventing numbers. The part worth completing is that Claude Code has a native form for exactly this, and it can enforce the safety rules the prompts can only request.
Most multi-agent setups fail the same way: six agents all try to do the whole job, each re-deciding what the video is about, and the outputs contradict each other. This one avoids that with a single constraint — each agent owns exactly one stage and reads the approved output of the stage before it. Ideas do not write scripts. Scripts do not plan shots. The pipeline is linear on purpose, with two human gates early where changing your mind is still cheap.
Who owns what
Each agent reads PROJECT.md plus the approved files before it, and writes one file into projects/[PROJECT]/.
The folder structure
Empty files are fine to start. The structure is what keeps each stage from reaching into another one’s work.
Five minutes, most of it copying
Create the folders
Empty files are fine. The shape matters more than the contents at this point.
Paste each agent prompt into its file
One prompt per file under agents/, numbered in pipeline order so the sequence is obvious from the directory listing.
Fill in brand/ and examples/
Your voice, your rules about what you never do, and your genuinely good past work. This is the only part nobody can write for you, and it is what stops the output reading like everyone else’s.
Write PROJECT.md for the first video
One project file, read by every agent. This is what keeps six separate prompts describing the same video.
Open Claude Code in the folder and run the master prompt
cd AI-VIDEO-TEAM && claude — then paste the master prompt to start the chain.
CLAUDE.md — the standing rules
Loaded at the start of every session, so the constraints do not need restating in each prompt.
Every safety rule here lives in prompt text: never publish, never invent a number, never assume a tool is connected. Anthropic’s own documentation is explicit that CLAUDE.md is treated as context rather than enforced configuration, and that to block an action regardless of what the model decides you need a PreToolUse hook. So a CLAUDE.md line saying “never publish anything” is a strong hint, not a lock. It is worth keeping — clear, specific instructions are followed more consistently — but worth knowing it is the weakest of the three ways to stop something, behind restricting the agent’s tools and behind a hook that refuses the call outright.
Claude Code already has a native form for this
The guide has you keep six prompts in a plain agents/ folder and paste them by hand. Claude Code supports exactly this pattern natively: a subagent is a Markdown file with YAML frontmatter in .claude/agents/ for one project, or ~/.claude/agents/ to reuse across all of them. The body is the prompt you already wrote — it moves across unchanged.
What the move buys is not tidiness. Each subagent runs in its own context window with its own system prompt, its own tool access and independent permissions, so the producer’s shot-list sprawl never crowds out the script, and Claude delegates to an agent automatically when a task matches its description rather than waiting to be told.
The field that matters most is tools. The guide asks the editor agent not to claim it edited footage and asks every agent not to publish. A subagent declaring tools: Read, Write, Glob, Grep cannot publish — not because it was asked not to, but because the capability is absent. That turns the document’s best intentions into something structural. Where you need a hard stop rather than a narrow toolset, a PreToolUse hook can refuse the call outright, and subagent frontmatter supports hooks directly.
Agent 1, as a real subagent
Same prompt, four lines of frontmatter. Drop it at .claude/agents/viral-ideas.md.
Worth knowing before you convert them
PROJECT.md — one file every agent reads
The fields people fill in badly
Viral ideas
Ten angles, then three, then a human picks one. The ban on predicting performance is what keeps this from generating confident nonsense.
Hooks and packaging
Titles and thumbnails as a set rather than separately, because they are read together.
Storytelling
The separation of PLANNED STORY from ACTUAL RESULT is the line that stops a documentary becoming fiction.
Scriptwriter
Producer
Editor
The master prompt
The instruction not to let all six agents redo the same job is the one doing the work here.
The shape of a run
Raw idea → Agent 1 returns three options → you pick one → Agent 2 returns three packages → you pick one → Agents 3 to 6 run through to the end → FINAL-VIDEO-PACKAGE.md → you review and publish.
The two gates are both at the front, which is the right place for them. Changing your mind about the concept costs nothing at stage one and costs a rewrite of everything downstream at stage five. After packaging is approved, the remaining four stages are mechanical enough to run without stopping.
Claude Code has no automatic access to YouTube or social analytics, to your editing software, to publishing tools, to current trends, or to any external data. Every one of those needs an API, an MCP server, a browser or search tool, a connector or a credential actually configured. Without that the agents are working only from PROJECT.md, research/ and examples/ — which is fine, and is exactly why the prompts forbid inventing the numbers they cannot look up. The failure mode is not the tool being missing; it is an agent filling the gap with something plausible.