Keep one outcome per task
Small, reviewable tasks are easier to steer and much easier to trust. Split unrelated work into separate tasks instead of piling everything into one run.
Start with constraints, not just goals
Tell the agent:
- what to change
- what must stay unchanged
- how to validate
- how to summarize the result
This cuts down on scope drift quickly.
Treat runs as collaboration, not one-shot prompts
Start narrow, inspect what came back, and then tighten the next message. Short iterative loops are usually faster than one giant “do everything” prompt.
Ask for artifacts when behavior matters
If a task changes a UI, a workflow, or a generated output, ask for a screenshot, recording, or short report so you can review behavior before you inspect code in depth.
Keep one stable setup while you learn the tool
During first-run onboarding, stick with one provider and one model until the full loop is working. Add more providers, stronger policies, or more automation after the baseline is stable.
Diff review is the acceptance gate
Do not finalize a task because the chat summary sounds good. Finalize only after the diff matches the scope, validation is addressed, and the result looks trustworthy.