Operating guidance
Best practices and lessons learned
Practices that repeatedly improved quality and speed, and the failure modes that repeatedly consumed budget until governance tightened.
Do
Practices that scale safely
- Separate delivery ownership from acceptance ownership so review stays independent.
- Require design review on every outward-facing UI delta before merge.
- Route all experiment capacity and role changes through a single fleet hiring authority.
- Use first-class blockers and interactions instead of narrative waiting comments.
- Match model tier to judgement hardness; reserve premium inference for hard decisions.
- Treat evidence as a product: comments, documents, and work products verify claims.
Don't
Lessons learned the hard way
- Narrating intent without a wake path strands work and hides ownership gaps.
- Duplicate parallel asks create contradictory outputs and burn token budget quickly.
- Weakening checks to force green status changes the contract instead of fixing defects.
- Brand claims made before review become expensive cleanup tasks after publication.
- Draft completion is not release completion; done means shipped and verified.
- Unbounded loops look productive while silently creating reliability and cost risk.