$ ./first-rule --present

[init] loading context...

[init] follow-up: the-new-dimension

[init] audience: engineering

[init] duration: 25 min

[████████████████████████████████████] 100%

THE FIRST RULE

building your compound system

press [SPACE] to begin

last time, we drew the map.

MANUAL
AI-ASSISTED
COMPOUND
AUTONOMOUS

Today: how to build the system.

Five primitives. One running analogy.

We're going to talk about kitchens.

$ play kitchen-brigade.mp4

Watch the head chef at the pass. He's not cooking.

What you just watched is a compound system.

Every station knows its job.

Tickets route work automatically.

The kitchen runs because someone designed the system.

Not because someone does every task.

COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK
COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK

The Recipe Card → Commands

Kitchen: A recipe card at every station. No improvisation. No "I think it was..." The knowledge is in the system, not in someone's head.

Engineering: Encode a task once. Anyone can invoke it. Consistent every time.

# .claude/commands/scaffold-feature.md Create branch, scaffold structure, add failing test, create draft PR.
Verify: branch exists, test runs and fails, PR is created.
COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK

Prep for Service → Skills

Kitchen: Prep isn't a recipe. It's a workflow that makes decisions. 120 covers or 40? What's seasonal? What takes longest? Same job title. Different execution every night.

Engineering: A skill is a command with judgment. It reads context, assesses the situation, and adapts.

# .claude/commands/implement.md Read the spec. Assess blast radius. If 3+ modules, confirm plan first.
Follow existing patterns. Test everything. Fix failures before reporting.
COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK

The Sous Chef → Agents

Kitchen: "You've got fish tonight." That's the entire handoff. They own the station. They deliver the result.

Engineering: An autonomous session. Give it a goal. Go work on something else.

# kick off and move on claude -p "Analyze src/auth/**. Write tests for untested paths.
  Run suite after each. Continue until 90% coverage." &
COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK

The Brigade → Agent-Teams

Kitchen: Garde manger, saucier, pâtissier. Each autonomous but coordinated. "86 the salmon." Together, they serve 200 covers.

Engineering: Multiple specialized agents. Same codebase. Running in parallel.

# three reviewers, one PR — each in an isolated worktree Security reviewer & Performance reviewer & Correctness reviewer
DISCUSSION

Which of these could you use this week?

What workflow would you encode first?

COMMAND
SKILL
AGENT
AGENT-TEAM
HOOK

The Ticket Printer → Hooks

Kitchen: The POS fires tickets to the right stations automatically. No one reads every order. The system routes work. This takes the head chef off the line.

Engineering: A trigger that removes you from the loop. Events fire. The system responds.

// .claude/settings.json PostToolUse: lint every file on Write|Edit
Stop: run full test suite when session ends

five primitives. one rule:

Never correct the same thing twice.

explain it and it works → command

command needs judgment → skill

skill should run alone → agent

one agent isn't enough → team

tired of invoking it → hook

The head chef doesn't cook every dish.

The head chef built the kitchen.

Your move.

this week — pick one thing you corrected the AI on. turn it into a command, a rule, or a hook.

the first rule isn't "automate everything."

Don't correct the same thing twice.

bring it back next week.

OPEN FLOOR

$ open discussion

Let's talk.