- Start in read-only plan mode, then turn what the agent found into a reviewed plan. For high-stakes plans, consider a second model as a check.
- Build in small steps and make your own git commits at logical points so changes are easy to undo. Read the diffs for security problems.
- Describe bugs by symptom, location and desired fix, and handle one problem per prompt.
- Verify with tests, builds or linters. Stricter options include a separate evaluator, a stop hook and a second-opinion subagent.
- Control permissions with allow, ask and deny rules, ideally starting from deny.
- Keep context tidy with `/clear`, a focused `/compact` and subagents for research.
This is a summary in our own words, not the article itself. Read the original for the full text.