Overview
Sessions are where agents execute your task instructions.
Message quality strongly affects output quality.
Prompting pattern that works
Use short, constrained instructions:
- what to change
- what must stay unchanged
- how to validate
- how to report back
During a run
- Watch output for signs of scope drift.
- Keep follow-up prompts narrow.
- Ask for concise summaries after major edits.
After a run
Always move to diff review before deciding the task is complete.
Common mistakes
- Mixing multiple unrelated asks in one message
- Accepting completion claims without checking diff
- Letting follow-up prompts broaden scope unintentionally