Chapter 6 of 6
How I actually use AI
The ten lines I actually run.
The tools keep changing. My loop doesn’t:
- Put the outcome and definition of done in the first prompt.
- Ask the agent to research the repo and propose a plan.
- Correct the plan before it writes code.
- Give independent work to parallel agents.
- Pick Claude Code or Codex for the job.
- Load a skill instead of retyping a checklist.
- Review the diff, then simplify it.
- Open the product and use it.
- Test according to what can break.
- Ship it, then watch what happens.
Turns out the code is the cheap part. Direction, taste, ownership, and talking to users still belong to me.