Your Operator Toolkit¶
Every lesson here is built to leave you a durable tool that outlasts the idea behind it. By the end you have a small kit you carry into real work — a set of files and habits — and a version of your own codebase that an AI agent can navigate without you. This page is the running index of what you build.
The artifacts in the table below are produced across the modules, in nav order; you have not met most of them yet. Read it as a forward map of where each lesson leaves you, and return to it as you go.
What you build, module by module¶
| Module | What you build | What it does for your real work |
|---|---|---|
| 0 · Agent loop | a "failure-tells" note | you spot lost context, plausible-but-wrong edits, and thrashing at a glance |
| 1 · Context | a navigational CLAUDE.md |
the agent opens the right file on its own, every session |
| 2 · Specification | a reusable spec template | a sharp spec becomes your default instead of an act of discipline |
| 3 · Verification | a verification gate and an eval set | output you didn't write is checked by machinery rather than by your eyes |
| 4 · Agent Harness | a harness map and a custom skill | the agent gets the tools, permissions, workflow, and check it needs |
| 5 · Orchestration | a worktree workflow and a slicing checklist | you run several agents at once without losing control |
| 6 · Compound | the toolkit index, plus an AI-ready repo | each task makes the next one cheaper |
| 7 · Judgment | a cost/fit decision log | you reach for an agent only when the math favors it |
| 8 · Capstone | all of the above, applied end to end on your own repo | proof you can operate an agent like a tech lead |
What you can inspect now¶
The built Practice Repo slices show the proof trail for the core loop: a
navigational CLAUDE.md, a spec that turns a vague request into a testable change, a failing test driven
green, an eval gate, and a reusable verification skill. Those artifacts are the evidence the course asks you
to produce on your own codebase.
How to use it¶
Keep these in one place — a toolkit/ folder, or committed into the repo each one belongs to. The value comes
from each one paying a dividend on every later task, rather than from the collection itself. A CLAUDE.md that gets sharper over
time, a gate that catches a regression when you are tired, a skill you run a dozen times: that compounding is
the real return of the course, and Module 6 is where it becomes a system.