Stop prompting. Start engineering.
A working engineer’s course on getting real software out of coding agents — Claude Code, Codex, Cursor, Copilot, Aider, whatever you run next year. The tools churn; the process underneath them does not. This course teaches the process.
You have probably already had the good day: a feature that would have taken an afternoon landed in twenty minutes. And you have had the bad one: three hours of an agent confidently rewriting the same file, a diff you could not read, and a branch you quietly deleted.
The difference between those two days is almost never the model. It is whether you managed the agent’s context, scoped the work to something a single session could hold, gave it a way to check itself, and reviewed what came back. That is an engineering discipline, and it is learnable.
Every module below is written to be worked, not watched. Bring a real repository — ideally one you know well enough to spot a bad change in.
Curriculum
Before We Start
Set the frame: what this course is for, the two ways agent-assisted work goes wrong, and the workbench and repository you will use throughout.
0/6 02Concepts
The mental model everything else rests on: what a model is, what a harness adds, how context behaves as it grows, and why agents fail the specific ways they do.
0/9 03Getting To Know Your Agent
Hands on the tool. Configure a harness properly, run a session end to end, read what came back, and choose between pairing with the agent and leaving it running.
0/7 04Fundamentals
The core moves, practised on real code: exploring, interrogating your own requirements, scoping sessions, building feedback loops, testing, debugging, and refactoring.
0/8 05Steering
Stop repeating yourself. Encode conventions, capabilities, and standards into files the agent reads — and into checks it cannot talk its way past.
0/8 06Shipping
Work that is bigger than one session: specs, tickets, phase decomposition, handoffs, unattended runs, review at volume, and a full feature start to finish.
0/8How to work through it
- Do the exercises. Each lesson ends with one, sized to a single agent session. Reading about context management teaches you nothing; blowing a session and recovering from it teaches you everything.
- Keep a scratch log. One file where you note what you asked, what came back, and what you had to correct. It becomes the raw material for your own
AGENTS.mdin Module 05. - Pace it. Three focused days, or two weeks at an hour a night. Modules 02 and 04 are the load-bearing ones; do not skim them to get to the shipping material.
Before Lesson 1.1
Open a terminal, pick a repository you actually work in, and start a session with whichever agent you use. Ask it one question: “Explain what this project does and how it is structured, based only on files you have read.” Save the answer. You will grade it in Lesson 1.5.
When you have finished this
The Next Steps — the paid follow-up
This course teaches a process you can trust. The Next Steps is what comes after it: 128 lessons across 16 modules on the mechanics underneath, orchestration across many sessions, verification you can rely on, legacy systems, security, cost, teams, debugging at depth, and six complete worked projects — roughly five times the length of what you are reading now.
Sixteen lessons of it are free to read, one in every module, in full.
See the curriculum →A course by Pieter Zandbergen