Kitchen Crew

Count what you've been typing every time you finish a change, all the way back to Lesson 5: test it locally. Ship it. Verify it live with your browser. Make a save point. Five orders, same shape, every single time. By Lesson 8 you could type them in your sleep — which is exactly the problem. Routines you run from memory get skipped steps on tired nights, and the step you skip is always the verify.

Restaurants solved this ages ago: procedures live on laminated cards, not in anyone's head. Closing checklist, opening checklist — any competent crew member follows the card and the result is the same on a Tuesday as on New Year's Eve. Your agent can hold laminated cards too. They're called skills, and they're the last capability this course hands you before graduation.

What you'll plate today

A closing-time skill: one short order — "close up" — runs your entire test-ship-verify-save routine, every step, every time. Plus the mental model for when rules, skills, and plans each earn their keep.

Ingredients

• Your dinner-service folder and its well-worn routine
• About 25 minutes

Cook

1. Sort the three kinds of remembering

You now own three memory tools, and confusing them is the classic graduate mistake, so here's the sort:

• CLAUDE.md — rules. Always on, shape every action. "Announce before changing." "Secrets never in files."
• PLAN.md — state. What's done, what's next, what guests asked for.
• Skills — procedures. Multi-step routines with a name, run on demand. The laminated card.

The tell: if you keep correcting the agent, it's a rule. If you keep re-explaining where things stand, it's the plan. If you keep dictating the same sequence of steps, that's a skill announcing itself — same reflex as Lesson 4 of First Kitchen, one level up.

2. Laminate the closing routine

Create a skill for this project called "close-up" — my closing
routine, written down once, properly:

1. Run the app locally and check the basics work.
2. Ship the new version.
3. Do the guest pass from CLAUDE.md with your browser on the
   live app — front page, one test submission, phone width —
   and REMOVE your test card via the pass when done.
4. Make a save point with a plain-English note about what
   changed, mirrored to GitHub.
5. Report back in four lines max: what shipped, what you saw,
   anything that smells off, and the save point name.

Set it up so I can trigger the whole thing with one short order.
Then show me the skill file you created and walk me through it —
I want to read my own laminated card.

Read the file it shows you. This is the part worth slowing down for: a skill is just instructions in a file — the same plain language you'd have typed, written down once, stored in your project. No code, no magic. You are, at this point in the course, someone who writes procedures for a junior colleague. Because that's what this is.

3. Prove the card works

A procedure you've never run is a rumor — fire-drill rule, Lesson 1. Make a small real change:

Make the wall show how long ago each card arrived — "2 days ago"
instead of the raw date.

Then, the entire closing ceremony, in one order:

Close up.

Now watch what you used to type five orders for: local check, ship, browser pass, test card in and back out, save point, four-line report. If it skipped a step, the card is wrong, not the night — fix the skill file, run it again. Procedures improve by editing, which is exactly why they beat memory.

4. One more crew member (you'll feel clever)

The closing skill serves the kitchen. Add one that serves the host:

Create a second skill, "front-desk": read every card that arrived
since the last time I asked, give me the gist in plain lines, and
flag anything that looks like junk or like a guest REQUEST we
should log in PLAN.md's FEEDBACK section. One short order runs it.

Tomorrow morning: "front desk." Your app reports to you now. Notice you're assembling a crew — closer, front desk — each one a procedure you wrote once. (When you're ready to see how deep this goes, our skills-and-agents recipe shows the grown-up version — this blog runs its own publishing on exactly this pattern.)

When it burns

• "Close up" does something random instead of the routine — the skill didn't register; a restart usually wakes it: /exit, claude, try again. Still lost? "List the skills available in this project" — the answer tells you what it can actually see.
• The skill ran but skipped the verify — edit the card, not the vibe: "Open the close-up skill file and make step 3 unskippable — the routine isn't done until the browser pass is reported."
• It over-engineered the skill file into a novel — same dragon as the bloated CLAUDE.md in First Kitchen: "Cut this skill down to the steps I dictated, in plain words." Lean cards get followed; fat ones get improvised around.
• You're tempted to make everything a skill — the pantry rule from First Kitchen 9, resurfacing. Two corrections is a rule; two dictations of the same sequence is a skill; everything else is just cooking. Earn the laminate.

Order up

□ You can say which job CLAUDE.md, PLAN.md, and skills each do
□ "Close up" runs the full routine — including the browser pass
□ You read your own skill file and understood every line
□ "Front desk" briefs you on new cards and flags requests
□ The skill files are in save points, mirrored like everything else

Next up — Lesson 10: Grand Opening. No recipe, no training wheels. You choose the app, plan it, build it with your crew, ship it, and put it in the hands of real people. Everything from nineteen lessons across two courses, on one plate. Yours.

Stuck on a step? Question box below.