Prompt as Program

The pattern: prompts are code (or config, depending on the lens) — versioned, tested, PR-reviewed. An eval suite (golden inputs + expected behavior) gates prompt changes. Regressions get caught at PR time, not by users.

The trade-off: engineering discipline vs. iteration speed. Throwaway strings are fast; treating prompts as code is slower upfront but compounds. The maturity arc: prompts in code, then versioned files, then a registry with eval, then feature-flagged with A/B tests. Senior+ ML/AI work lives near the right end of that arc.

Deepens in Year 5 Phase 26: Agent Development — prompts live in basecamp git, the PR template requires eval-pass, regressions are tracked over versions. Phase 28: AIOps and services/aiops/ exercise this discipline under production pressure.

Related patterns

• agent-loop — the planner node is the prompt this pattern governs.
• tool-use-as-capability — prompts and tool schemas evolve together; both need eval coverage.
• model-lifecycle — same versioning shape applied to model artifacts.
• train-serve-skew — prompt drift is the LLM-era cousin of train/serve skew.