[How Coding Agents Work][09] Prompt Assembly as Typed Fragment Composition

How Coding Agents Work · Module 09 of 29: Prompt Assembly as Typed Fragment Composition Part II — The Six Layers Context layer. Codex's prompt is not a template string — it's a composition of roughly fifty individually named, typed fragments, each an explicit decision about what the model must see right now. The clearest worked example of context engineering in any open harness. After this module you can replace a monolithic system prompt with a composed fragment set, and justify each fragment's existence and position. Lessons in this module: 1. Three layers, two channels: base instructions ride the API instructions field while world state arrives as developer messages inside the input history 2. Where the live prompt actually comes from: the model catalog's instructions template, not the reference .md files 3. AGENTS.md discovery, precedence, trust, and byte budget 4. The fragment catalogue: fifty named decisions, read as a list of hard-won lessons 5. Reminders, warnings, and nudges: the fragments that exist because the model misbehaves 6. Ordering, stability, and keeping the prompt cache warm Prerequisites: module 8. About the course: Built from the OpenAI Codex harness at commit 8d01cd42fa.