Lesson 6 · 5 min
Iterating Prompts Like Code
Version, test, and review prompts.
Store prompts in version control. Pair every prompt change with eval results. Treat prompt.md like schema.sql: production-critical, reviewable, rollback-able.
Production scenario
Real-world example: Prompts-as-code at a 50-person AI team
The team treats prompts the same way they treat database schemas:
- Every prompt lives in
prompts/*.md, version-controlled. - Every PR that changes a prompt runs the eval harness on 50 golden tasks.
- A bot posts the diff: "success 78% → 82%, cost +12%, p95 latency stable."
- Rollback is one revert away.
When prompt A's success drops below B's, the merge is blocked. When it improves and cost is acceptable, it merges and ships behind a feature flag.
Why this matters: prompts have production impact. Treat them with the rigor you'd give application code, including review, testing, and rollback.
Knowledge points in this lesson
- Treat prompts like code: version them
- Pair prompt changes with eval results
- Rollback should be supported
- Don't keep prompts in admin UI history-less
- Review prompts like schemas
Quick check
Prompt EngineeringSelect one
Where should you place the HARDEST few-shot example?