youtube-team-os

youtube-team-os

A Claude Code / Cowork plugin that runs a full 5-role YouTube production OS inside your project directory.

What it does

Orchestrates five specialist roles — producer, writer, editor, thumbnail designer, and growth manager — through a persistent filesystem-based project state. One command starts a video; the system hands off between roles, preserves all deliverables, and runs postmortems.

Skills

Skill	Invoke	Purpose
orchestrator	/youtube-team-os:orchestrator	Route work, create projects, weekly planning
producer	/youtube-team-os:producer	Idea selection, content brief, calendar
writer	/youtube-team-os:writer	Outline, script, hook, B-roll cues
editor	/youtube-team-os:editor	Pacing, cut list, edit notes, chapters
thumbnail	/youtube-team-os:thumbnail	Thumbnail concepts, title variants, CTR audit
growth	/youtube-team-os:growth	Publish plan, repurposing, analytics, postmortem

Quick Start

# Install (dev mode)
claude --plugin-dir ./youtube-team-os

# Start a new video
/youtube-team-os:orchestrator "10 pricing mistakes SaaS founders make"

# Resume an existing project
/youtube-team-os:orchestrator 2026-06-ep14-pricing

# Weekly planning
/youtube-team-os:orchestrator plan week

Project State

All work lives in state/:

state/
  projects/<id>/          ← one folder per video
    brief.md
    outline.md
    script.md
    edit-notes.md
    thumbnail-options.md
    publish-plan.md
    postmortem.md
    status.md
    activity.log
  roles/<role>/           ← per-role task queue
    inbox.md
    current.md
  notes/
    content-calendar.md
    channel-learnings.md
    weekly-plan-*.md
  records/
    session.log
    writes.log
    analytics-<id>.md

Workflows

• workflows/idea-to-upload.md — full pipeline diagram
• workflows/weekly-planning.md — Monday planning procedure
• workflows/postmortem.md — 30-day retrospective

Hooks

hooks/hooks.json logs session starts/ends and every file write to state/records/, and runs the deadline/postmortem-due reminder checks on SessionStart (see Monitors below).

Connectors

• connectors/youtube-metadata.md — YouTube analytics (manual input for now — see file)
• connectors/drive.md — Google Drive sync via bin/sync-state.ps1/.sh (rclone-based)

Install from marketplace

/plugin install <marketplace>/youtube-team-os

Scripts

Script	Platform	Purpose
bin/new-project.ps1 -Slug "my-topic"	Windows	Scaffold a new project with all state files
bin/new-project.sh my-topic	Mac/Linux	Same as above
bin/validate.ps1	Windows	Health check: manifest, files, hooks, jq, state dirs
bin/validate.sh	Mac/Linux	Same as above
bin/sync-state.ps1	Windows	Push deliverables to Google Drive via rclone
bin/sync-state.sh	Mac/Linux	Same as above
bin/smoke-test.ps1	Windows	Deterministic mechanical regression test — see Testing below

Agents

Six specialist agents in agents/ are auto-routed by skills. Each has a scoped system prompt, tool list, and model config:

orchestrator → producer → writer → editor → thumbnail → growth

Monitors

Two reminder checks run via hooks/hooks.json on SessionStart (not via monitors/monitors.json, which is intentionally empty — see CHANGELOG.md):

• content-calendar-deadlines (monitors/check-deadlines.js) — warns if a project's publish date is within 3 days and it isn't yet at PUBLISH/POSTMORTEM stage
• postmortem-due (monitors/check-postmortem-due.js) — reminds when a video is 28–35 days post-publish and its postmortem is still a stub

Both require Node.js on PATH.

Resources

• resources/best-practices/youtube-production.md — hook, retention, CTR, SEO rules
• resources/best-practices/claude-code-skills.md — skill authoring guide
• resources/metrics/youtube-kpis.md — benchmark targets with grade scale
• resources/links/tools-and-resources.md — 50+ curated tool links
• resources/scoring/ctr-audit.md — thumbnail CTR scoring rubric

MCP Connectors

.mcp.json currently defines no servers — the previously configured YouTube and Drive MCP packages were removed (one didn't exist on npm, the other is deprecated). See connectors/mcp-setup.md for the current manual/rclone-based alternatives and how to wire up a real MCP server later if one becomes available.

Testing

Two layers, because the 6 role skills/agents are markdown instructions executed by an LLM, not deterministic code — no script can drive them the way a unit test drives a function.