What this course actually teaches
Not prompts. A repeatable process for turning an agent into a reliable part of your engineering workflow.
There is a genre of AI-coding advice that consists of prompt templates. Paste this magic preamble, add "think step by step", promise the model a tip. That advice has a short half-life: it is tuned to one model, one harness, one month.
This course is about the layer underneath. A coding agent is a stateless model wrapped in a harness that gives it tools and a context window. Everything you can control lives in that wrapper: what goes into the context, how big the job is, what tools the agent has, what checks it runs, and what you do with the diff that comes out. Those levers do not change when the model does.
The four things that actually move results
- Context management. What the agent knows right now, and what you deliberately keep out. Most bad sessions are context problems wearing a different hat.
- Task decomposition. Splitting work so each session holds one coherent job, with clean boundaries between sessions.
- Steering. Standing instructions, skills, and project conventions that make the agent behave well without you re-explaining every time.
- Verification. Automated checks, automated review, and your own eyes on the diff — in that order, for cost reasons.
Modules 02 and 04 build the first two. Module 05 is steering. Module 06 is decomposition and verification at the scale of a real feature.
Tool-agnostic, deliberately
Examples are written for a terminal-based agent because that is the least abstracted form: you can see every tool call. Everything transfers. Where a specific product differs, the lesson says so and names the equivalent concept rather than the product feature.
Concretely, the course assumes your agent can: read and write files, run shell commands, search a codebase, and ask you for permission before doing something destructive. If yours can do that, you can do every exercise here.
What this course is not
- Not a model comparison. Benchmarks move weekly and your codebase is not a benchmark.
- Not an argument that you should use agents. You already do, or you would not be reading this.
- Not a promise that review gets cheaper. It does not. Generation gets cheaper, which makes review the bottleneck. Module 06 is mostly about that.
Try it
Write down, in one sentence, the last time an agent session went badly for you. Keep it. At the end of Module 04 you will be able to name which of the four levers above was missing — and it will usually be the first one.
Takeaways
- The model is the part you cannot control; the harness, context, and process are the parts you can.
- Prompt tricks are model-specific and expire. Context, decomposition, steering, and verification do not.
- Cheap generation moves the bottleneck to review. Plan for that from the start.
Why does this course spend so little time on prompt wording?
Because wording is the smallest and least durable lever. The same task succeeds or fails mostly on what context the agent has, how big the job is, and whether it can check its own work — all of which survive a model upgrade, while a tuned prompt often does not.
A course by Pieter Zandbergen