Hire the brain

The brain is one session

My orchestrator is Claude Code running on Claude Fable 5.1. It holds the plan, the decomposition, the tests, and the merge. It does not build. When it builds, it stops being the gate.

Everything else in this course hangs off that one session.

The skill set

I keep the operating rules as skills, so I never retype them. The public set is on GitHub with an installer:

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

That installs the orchestrate skill plus three standing agents: implementer, which builds to tests; researcher, which returns cited facts; validator, which checks and never fixes.

The skill turns a session into a conductor. The loop it runs is: frame, decompose, spec the behavior as failing tests, delegate the build, verify adversarially, kick back or integrate, report with evidence.

The dispatch router, dispatch.sh, calibrate.sh, and the lane report schema are not in that repo yet. They land in the repo next.

The prime directive

A subagent's output is a claim, not a fact.

Work that misses the bar goes back to a subagent with the specific gap named. I do not quietly fix it myself, and I do not wave it through. If I skip the kick back, I have added latency to solo work and nothing else.

There is a second half to that rule. A unit that fails twice is mis-scoped. I re-decompose it. I do not escalate it to a bigger model. That one took me a while to accept.

Model policy

Subagents are never Fable class. That is a hard rule, and every dispatch sets the model explicitly, because leaving it out inherits the session model. The omission is the violation.

• Claude Opus 4.8 builds and validates. Anything that writes product code, and every validation pass.
• Claude Sonnet researches: reading code, mapping a repo, summarizing, API calls.
• Claude Haiku does mechanical work: renames, formatting, moves.

Validation is always Opus 4.8, whatever built the unit. A cheap builder plus an expensive gate is much cheaper than an expensive builder with no gate.

Set it up

1. Install Claude Code and sign in on the plan you already pay for.
2. Clone the skill repo above and run the installer. Confirm the three agents appear under your agents directory.
3. Open a repo that has a real test command. Read its CLAUDE.md first, or write one if it has none.
4. Ask the session to orchestrate one small change. Watch it write the failing test before any product code.
5. When a subagent returns, re-run the test command yourself in the terminal. Compare the output to what the agent claimed.
6. Kick one thing back on purpose. Name the exact failing assertion. Notice how much better the second attempt is when the gap is named.

What you can do now

You have an orchestrator session that plans, delegates, and checks, with the model named on every dispatch.

Next module

Module 3 hires the first cheap hands: Codex CLI, its lane profile, and the one repo file that makes it read your rules at all.