A good agent session has a shape. Skip a step and you'll usually pay for it in review.
The five steps
- Load context. The agent reads your context file automatically. Check that it did by asking what commands it would run.
- Map. Ask for a map of the modules, connections and blind spots involved. No edits yet.
- Answer questions. Every blind spot on the map is a question. Answer it before you let the agent guess.
- Change in small steps. One connection per step. Commit between steps.
- Verify. Run the verification loop before you call it done.
A starter prompt
Task: [one sentence]
Before editing, map the modules and connections this touches, and list anything
you can't see both sides of. Then stop.Signs it's going well
- The agent asks one specific question instead of eight vague ones.
- Diffs are small enough to read in a minute.
- "Done" comes with test output, not just a summary.