Prompting tips

The agent works best when you treat it like a senior engineer pair — specific goals, scoped context, verifiable success criteria.

Do

• Name files and symbols. @auth/session.py beats “the auth file”.
• State the outcome. “Add bcrypt verification and unit tests” beats “fix auth”.
• Give constraints. “No new dependencies”, “keep Python 3.10 compat”.
• Ask for a plan first on large refactors, then approve step by step.

Avoid

• “Make it better” without criteria — the model will guess.
• Pasting secrets (API keys, .env contents) into chat.
• Asking it to rewrite the entire repo in one shot — split by module.

Example prompts

Find every place we hash passwords without bcrypt and migrate them.
Show a diff per file; do not touch tests yet.

---

Add type hints to backend/routes/auth.py only. Run pyright; fix errors
you introduce.

---

Explain why pytest collection fails in tests/auth/, then fix the import
path. Run pytest -q when done.

Max quality mode

On Plus and Pro plans, enable Max quality for hard reasoning tasks. It routes to the strongest model and consumes ~2–3× more tokens per turn. Use it for architecture decisions, not autocomplete.