What we covered, to look up again when you work with your agent.
An invented chapter: an agent reads, plans, acts and checks, in a loop, until the task is done or it stops to ask.
Each turn the agent sees the files it opened, the commands it ran and what they printed. What it never saw, it cannot know.
Key point: give the agent a check it can run, and it corrects itself.
Start the agent in the project folder and ask it to explain the build:
$ agent "explain how this project builds and tests"
A hook runs the check after every edit, highlighted as the hub's Markdown sets it (Shiki with the brand's code palette as classes):
// The check the agent cannot skip
const check = "bun run check";
export default { hooks: { afterEdit: [check] } };
Rules written once apply to every session. Where they live decides who reads them.
| File | Read by |
|---|---|
| Project rules | Every session in the repository |
| Personal rules | Your own sessions only |
A rule the agent follows without being told goes. A rule that names a measured number stays.
Space for your own notes from the training.