Practical thinking on building reliable LLM-powered applications.
Hardcoding prompts in your application is the obvious starting point. It's also the thing that will slow your team down the most. Here's what the alternatives actually look like.
Eval-first tools argue that prompt versioning without evals is just diff tracking. Here's why that gets the order of problems wrong.
Something broke. You suspect the prompt. Here's why rolling back is harder than it should be — and what good rollback actually looks like.
Versioning your prompts is the right instinct. Here's how to do it in a way that actually holds up in production.
Most teams treat prompts like config strings. In production, that breaks. Here's what prompt management actually requires once you're shipping real AI features.
You use semantic versioning for packages and APIs. Your prompts need the same discipline — here's what that looks like in practice.
Feature flags let you ship code safely — gradually, reversibly, and testably. Your prompts need the same pattern. Here's what prompt feature flags are and why they matter.
Prompts and code have different release cadences. Treating them the same creates bottlenecks, slows iteration, and locks non-engineers out of AI behaviour they should own.
Prompt drift is when your AI feature's outputs silently change over time — without anyone touching the code. Here's what causes it and how to stop it.
Change your AI behaviour in production without touching your codebase or triggering a deploy. Here's how prompt infrastructure makes it possible.