The Orchestrate Skill Set for Claude Code

This is the skill set from the video. Orchestrate puts your Claude Code session in orchestrator mode: it plans the work, hands the building to subagents, then verifies each result itself and sends back anything that misses the bar. Orchestrate does not work alone. It calls seven sibling skills and three agents by name, and this set ships all of them.

Nothing is held back and nothing is gated. Pick the install you like.

Way 1: paste this to Claude Code

Open Claude Code in any project, paste the prompt below, and send it. Claude clones the repo, runs the installer, and shows you what it put in place.

Install the orchestrate skill set from GitHub into my Claude Code setup.

1. Clone https://github.com/cgraves09/claude-orchestrate-skill.git into a temporary directory.
2. Run its ./install.sh. It copies skills/*/SKILL.md into ~/.claude/skills/<name>/ and
   agents/*.md into ~/.claude/agents/. It skips any file that already exists there. Do not
   pass --force unless I say so.
3. Show me the script's output and list exactly which files were installed and which were
   skipped.
4. Delete the temporary clone.
5. Tell me to start a new Claude Code session and type /orchestrate followed by a task.

Do not change anything else in ~/.claude. If git or the network is unavailable, stop and
tell me instead of working around it.

Free, ungated, no email required. Copy it as many times as you like.

Way 2: clone it yourself

Three commands in a terminal. The installer only copies files into ~/.claude/skills/ and ~/.claude/agents/, and it never overwrites a file you already have unless you add --force.

git clone https://github.com/cgraves09/claude-orchestrate-skill.git
cd claude-orchestrate-skill
./install.sh

The repo is at github.com/cgraves09/claude-orchestrate-skill. Read every file there before you install it. That is the right habit for anything that talks to your AI.

Set it up

1. Install with Way 1 or Way 2. Either one puts the same eleven files in place.
2. Start a new Claude Code session. Skills load when a session starts, so an open session will not see them yet.
3. Type /orchestrate and describe the task. Or just describe the task. The skill engages on its own for anything bigger than a trivial edit or a question.
4. Watch it work. It restates what done looks like, splits the task, spawns subagents, then checks their output before it reports back.
5. Make it yours. The files are plain markdown in ~/.claude/skills/. Edit the model policy, the evidence bar, or the ledger format to match how you work.

What you get

Eight skills. Orchestrate is the entry point, and it loads the others by phase.

• orchestrate. The loop: frame, decompose, tests first, delegate, verify, kick back, report with evidence.
• orchestrator-delegation. When to delegate and when to do it inline, which model runs each job, and why a subagent's output is a claim until you check it.
• delegate-prompts. Templates for the prompt you hand a subagent: context, acceptance criteria, output contract, self-check.
• tdd-protocol. Behavior spec, failing tests, green gate, and how to stand up a test harness where none exists.
• workflow-patterns. The shapes of a fan-out: pipeline, generator plus validator, review lenses, tournament, parallel worktrees.
• validation-and-evidence. What counts as proof, how a validator pass runs, and how to triage severity.
• context-and-state. Where state lives between sessions, the NEXT.md ledger format, and how to resume after an interruption.
• familiarize. How to map an unfamiliar repo before you write any code.

Three agents, installed to ~/.claude/agents/.

• implementer. Builds one scoped task to a test command and returns a diff plus verbatim test output.
• researcher. Gathers cited facts from the repo and the docs. It never edits code.
• validator. Checks finished work adversarially and reports pass or fail per criterion. It never fixes anything.

What it changes

• The session is the quality gate, not the builder. A subagent's "done" is a claim. The session re-runs the tests and reads the diff before it counts.
• Tests come first. Each unit of work gets failing tests before any code. The subagent gets the tests and "make these green."
• Work that misses the bar goes back. The session names the gap and sends it to a subagent again. It does not quietly patch it or wave it through.
• Independent units run in parallel. Several subagents build at once. Review lenses come in before anything merges.
• It reports with evidence. Test counts, commands run, and what was observed. Not "looks right."

What you should know going in

• It needs Claude Code. The terminal, the desktop app, or the IDE extension all work. Claude Desktop chat has no subagents, so these files do nothing useful there.
• Global or per project. ~/.claude/skills/ applies everywhere. Copy the same folders into .claude/skills/ inside a repo to scope them to that repo only.
• Subagents cost tokens. The agents ship with a small model for building and a large one for review. Change the model: line in each agent file if you want a different split.
• Your own rules win. Anything in your project's CLAUDE.md, and anything that touches production, still stops for you.
• Uninstall is a delete. Remove the eight folders from ~/.claude/skills/ and the three files from ~/.claude/agents/.

Sources

The set is published under the Unlicense at github.com/cgraves09/claude-orchestrate-skill. Claude Code skills are documented in the official skills guide and subagents in the subagents guide.

Watch the short