Playbook 1

Playbooks for delegating to coding agents

Split the ticket

An agent does best with one outcome per work order. If a ticket needs the words "and also", split it. Aim for a diff a person can read in under 30 minutes.

Write the single outcome in one sentence.
List the files the agent may touch.
Move every side idea to its own order.

Before

"Make login better: faster, safer, and fix the weird error message."

After

Order A: Return HTTP 429 after 5 failed logins in 10 minutes.
Order B: Show a clear message for an invalid password.
Order C: Profile slow login query (report only).

Playbook 2

Set guardrails

Guardrails are the part of the order that says no. They keep scope, cost and risk bounded.

Scope: name files and folders; everything else is off limits.
Out of scope: list tempting neighbours, such as refactors.
Constraints: no new dependencies, API unchanged.
Budget: set an iteration limit and stop-and-report rule.
Secrets: never put credentials in the order.

## Out of scope
- Password reset flow
- Any UI change
## Constraints
- No new dependencies
- Stop after 8 iterations and report

Build your own in the Work Order Builder.

Playbook 3

Write checks a test can verify

"Works well" cannot be tested. A good acceptance check names an input and an observable result.

Weak

Rate limiting should be reasonable.

Testable

- [ ] 6th failed login within 10 minutes returns 429
- [ ] Successful login resets the counter
- [ ] Existing login tests still pass

Playbook 4

Review agent output

Compare the diff file list with the work order scope.
Run every acceptance check yourself.
Look for swallowed errors, leftover debug code and edited tests that were weakened.
Read every changed line.
Record the verdict: pass, hold or stop.

Red flags

Tests deleted or loosened to pass.
Catch blocks that hide failures.
Changes in files nobody mentioned.
New dependencies without a reason.

Use the Review Gate checklist or ask the Diff Reviewer for a second look.