Tips for Effective Use

These patterns consistently produce better results, especially for complex tasks.

1. Provide constraints upfront

Mention data boundaries, compliance frameworks, change windows, and approval steps at the start. This avoids solutions that would never pass review.

"This system handles PII under GDPR Article 17. No data may leave the EU zone."

2. Use @-mentions for precision

Instead of describing a file, reference it directly with @filename. The assistant gets the exact content, no searching required. See @-Mentions.

3. Request a plan before implementation

For ambiguous tasks, ask for a plan first — goals, constraints, acceptance criteria. Review it, add your constraints, then say "proceed."

"First write a plan: what files will change, what tests will you add, what risks do you see?"

4. Break complex requests into steps

For very large tasks, work iteratively:

1. Ask the assistant to analyze the situation
2. Review its understanding
3. Ask it to implement changes
4. Review the changes
5. Ask it to add tests

This produces better results than a single "do everything" prompt.

5. Use subagents for large tasks

The assistant starts subagents automatically for complex work, but you can also direct it explicitly.

"Start subagents to investigate the legacy auth module and write a migration spec in parallel."

6. Steer it when it goes off track

If the answer is partially right, point out exactly where it went wrong and ask it to try again. This usually corrects the course immediately.

"You missed the approval gate in step 3 — try again keeping that constraint."

7. Use the conversation history

The assistant remembers the full conversation. You can refer back to earlier messages: "apply the same pattern we discussed above" or "use the approach from your first suggestion."

8. Start a new conversation when the task drifts

If your next question moves to a different task — even in the same part of the codebase — open a fresh conversation rather than continuing the current one. Older messages and tool results that were useful earlier can crowd out the assistant's focus and pull it toward the previous goal; a new thread will fill up with exactly the context the new task needs as the assistant explores.

"Let's continue this in a new conversation — different goal, same module."

9. Control tool permissions precisely

Use Settings → Tool Approvals to pre-approve safe operations and block risky ones. See Permissions for details.

10. Curate organizational memory

When the assistant surfaces a memory that's off-target or outdated, click Dislike and leave a reason. You're directly improving the shared knowledge base for every future conversation.

"This memory is outdated — we migrated away from this pattern in Q3."