UseAgentBEFORE YOU RUN
Know the job and the reviewer.
This playbook is for engineering leads coordinating several kinds of agent-assisted work.
TARGET OUTCOME
A repeatable routing system that matches a task to an engine, isolates context, and leaves a result a person can verify.
Bring these inputs
Better source material makes the result easier to review and safer to reuse.
- A prioritized issue list
- Repository and local instructions
- Acceptance criteria
- Required checks and review ownership
COPYABLE STARTER
Start with this prompt.
Replace bracketed details, attach the source material, and keep the limits until your team has reviewed a few runs.
Turn these engineering issues into separate, bounded runs. For each, define the goal, relevant repository area, acceptance criteria, required tests, allowed actions, and reviewer. Recommend an engine only when the task shape gives a clear reason. Do not combine unrelated issues or merge changes.
THE PLAYBOOK
Move from context to a reviewed result.
- 01
Classify the work
Separate investigation, implementation, review, migration, and documentation because each needs different context and proof.
- 02
Bound every run
Include target files, acceptance criteria, allowed actions, and exact validation.
- 03
Keep work isolated
Use a separate sandbox or branch when runs may touch overlapping code.
- 04
Review the evidence
Read the diff and test output. A completed run is a handoff for engineering judgment.
DELIVERABLES
What the run should produce
- Routed work queue
- Run briefs
- Test evidence
- Review handoffs
HUMAN REVIEW
Check this before it leaves the team
- Each run has a narrow scope
- Tests prove stated behavior
- Overlapping edits are coordinated
- A person reviews before merge
Keep exploring this workflow