Context Engineering
Overview
The practice of designing what information goes into an LLM’s context window and how. Distinct from “prompt engineering” — context engineering is about the system that assembles the prompt, not the prompt text itself.
Key Patterns
- Progressive disclosure — show high-level structure always, drill into details on demand. File hierarchy as navigational signal. (Anthropic, Letta)
- Frontmatter as metadata — YAML headers on files that describe contents without loading full content
- Pinned vs on-demand context — some files always in system prompt (
system/), others loaded when relevant - Agent-managed context — agents reorganize their own context files (Letta’s approach)
Relevance to Our Wiki
The wiki schema is a form of context engineering. The index.md file is progressive disclosure — I read it first to find relevant pages, then drill in. Source summaries are compressed representations of full articles. The whole system is designed to give me (the LLM) the right context for any query.