GIVE EACH AI A CLEAR JOB

1. Give the researcher a question and ask for evidence, assumptions and source locations.
2. Give the builder one implementation area and clear acceptance criteria. Avoid two agents editing the same files at once.
3. Give the reviewer the final change and the behavior to verify. Ask for concrete failures with file locations and a way to reproduce them.
4. Have the main task collect findings, resolve disagreements and run the relevant checks. Review the combined result yourself.

Use subagents for independent work on [task]. One investigates [question]; one reviews [area]. Keep both read-only. Return evidence with file locations. Wait for both, then implement the agreed fix in the main task and verify [acceptance criteria].

1. The investigations can run independently and the review findings can change the decision.
2. Use one agent for a small change where delegation would add more coordination than value.