One Cook Isn't a Kitchen

Welcome to Kitchen Brigade — the course for First Kitchen graduates who keep hearing the words agents, MCP, multi-agent and suspect, correctly, that there's something real under the noise. There is. And every course we could find on it assumes you write Python for a living. This one assumes you finished First Kitchen.

Here's the secret that makes this course possible: you already run an agent. Claude Code — the thing you've been ordering around since your first hello.html — is an AI agent: it reads, decides, acts, checks, and comes back. What you don't have yet is what a busy restaurant has and a home cook doesn't: a brigade. Stations. Specialists. One cook isn't a kitchen — and today you watch your one cook hire its first helper.

What you'll plate today

The concept, made physical: your agent takes a job it would normally grind through alone, hands part of it to a subagent, and reports back like a chef debriefing the pass. Plus the one honest rule that keeps this whole course from becoming expensive theater.

Ingredients

• Any project folder from your journey (your first-kitchen one is perfect), agent running
• About 25 minutes

Cook

1. Get the words straight (two minutes, no jargon)

Three terms this course uses constantly, translated once:

• Agent — a cook: reads the order, works the station, tastes, plates. You've run one for twenty lessons.
• Subagent — a second cook your main cook can call over: "you handle the sauce while I do the fish." Same skills, separate hands, separate attention.
• Brigade — cooks plus structure: who does what, in what order, who checks the plate. Structure is the part that makes a crew better than a crowd — and it's what lessons 2 through 10 actually teach.

That's the entire vocabulary. MCP arrives in lesson 3 with its own one-liner.

2. Watch one cook work alone (the "before" taste)

Give your agent a job with two genuinely different halves:

Two-part job, and do BOTH parts yourself, one after the other:

1. Read through this project and write a short "state of the
   kitchen" note — what's here, what's rough, what you'd fix.
2. Then switch hats completely: as a skeptical outside reviewer,
   pick the weakest three claims in your own note and challenge
   them.

Tell me honestly at the end: what was hard about doing both hats
yourself?

Read its answer to that last question. It will say something true: reviewing your own work with the same head that wrote it is exactly as compromised for AIs as it is for humans. The second hat knows what the first hat meant to say. That's the disease. Crews are the cure.

3. Now hire the second cook

Same job, brigade-style:

Same two-part job — but this time, DELEGATE part 2. Send the
"state of the kitchen" note to a subagent with fresh eyes and
this instruction: "You didn't write this. Find the three weakest
claims and challenge them hard." Then show me:

1. Your original note
2. The subagent's challenges, unedited
3. One line: where was the fresh reader right and you were wrong?

Watch what happens in your terminal — you'll see the delegation happen, a second worker spin up, and results come back. Then read the three outputs side by side. The subagent's challenges will be sharper than the self-review from step 2 — not because it's smarter (it's the same model!), but because it doesn't know what the author meant. It only knows what the note says. Fresh eyes are a structural property, not a talent. You just bought them with one sentence.

4. The honest rule (this course's spine)

Before the crew grows, the rule that keeps it honest:

Add a rule to CLAUDE.md: one cook by default. Suggest a subagent
or crew ONLY when the job has genuinely separate parts — fresh-eyes
review, parallel independent tasks, or different specialties. If I
ask for a crew where one cook would do, say so.

Here's the honesty this course owes you up front: a brigade costs more than a cook. Every extra agent burns more of your subscription's capacity, and a crew doing a one-cook job is slower and worse — coordination isn't free in kitchens or in software. The multi-agent hype skips this part; lesson 9 prices it out properly. The skill you're building across these ten lessons isn't "always use many agents" — it's knowing which jobs deserve a crew. Fresh-eyes review, you now know from experience, is one. Most jobs aren't.

When it burns

• You didn't see any delegation happen — say: "Did you actually use a subagent for part 2, or did you role-play it yourself? Redo it with a real subagent and point out where I can see it happen." Honest tooling note: the delegation view differs by version; the output structure (note vs unedited challenges) is your real receipt.
• The subagent's review looks identical in tone to step 2's — your step-2 self-review might have been genuinely good. Check the content: did fresh eyes catch anything the author-hat missed? Usually at least one. If truly nothing, your project is small — that's the "one cook was enough" rule proving itself. Also a finding!
• It spun up subagents for everything after this — the enthusiasm dragon wearing a new hat. Point at the CLAUDE.md rule: "one cook by default."
• "Isn't this just asking twice?" — sharp question. The difference is context: a re-ask shares the whole conversation (the author's memory included); a subagent starts clean. You proved the difference with your own eyes in step 3 — that's why the challenges were sharper.

Order up

□ You can say agent / subagent / brigade in kitchen words
□ One job, done solo — and the agent's honest note on why self-review is weak
□ Same job with a delegated fresh-eyes reviewer — and sharper challenges
□ You saw where the fresh reader was right and the author was wrong
□ CLAUDE.md holds the spine rule: one cook by default

Next up — Lesson 02: The Station Chart. A crew without structure is a crowd. Your CLAUDE.md, plans, and skills — the things you've been building since First Kitchen — get promoted into a proper crew handbook.

Stuck on a step? Question box below — a step that lost you is a hole in the lesson, not in you.