Practice
Practice is the implementation layer between What is…? concepts and Constitution enforcement. These guides explain what works in production, in what order, and with what tradeoffs.
New here? Start with Where to start.
Optimization pyramid
Work through these tiers in order. Skipping measurement or prompt hygiene to jump straight to model swaps is the most common optimization mistake.
Guides by tier
| Tier | Guide | Constitution | Typical impact |
|---|---|---|---|
| 0 | Where to start | Article I | Prerequisite |
| 1 | Prompt hygiene | Article V | ~20–25% on high-volume templates |
| 1 | Model routing | Article II | 60–95% depending on task mix |
| 1 | Prompt caching | Article I | 41–80% on cached input |
| 2 | Context hygiene | Article III | 40–60% in agentic workloads |
| 2 | Output and RAG | Article III, V | 15–40% on verbose or over-fetched workloads |
| 3 | Semantic caching | — | Eliminates full inference on cache hits |
How this relates to the Constitution
The Constitution defines what must be enforced in production. Practice explains how to implement those constraints — with prompt structure patterns and savings ranges you can benchmark on your own data.
Measure first. Optimize in order. Enforce what works.