Prompt workflow
Treat prompt changes like code changes: reviewable, measurable, and deployable.
Lifecycle (recommended)
- Draft a new prompt version in dev
- Evaluate against your dataset(s)
- Review for policy, brand voice, and known failure modes
- Promote to staging
- Gate with full eval suites and latency/cost budgets
- Promote to production
- Monitor traces and rollback quickly if needed
What to include in a prompt change request
- what user behavior you want to improve
- examples where the current prompt fails (links to traces)
- evaluation results vs baseline
- any new constraints (formatting, citations, refusals)
How to avoid “prompt spaghetti”
- keep a template library (partials for policies, format requirements)
- standardize variable names across prompts
- keep system rules short and explicit