A short textbook on managing what goes into a language model's context window — from tokens and system prompts to tool calls and memory strategies.
Read online: nl0.github.io/context-engineering Download PDF: Context-Engineering.pdf
Based on the outline from latentpatterns.com/courses/context-engineering, fleshed out with research into practitioner experiences, academic papers, and production systems.
| # | Chapter | Core concept |
|---|---|---|
| 1 | Tokens and Inference | Tokenization, stateless inference, the messages array |
| 2 | Context Window Size | RULER benchmark, lost-in-the-middle, smart zone vs. dumb zone |
| 3 | Messages Array | The 5 slots: system prompt, harness, project context, tools, your prompt |
| 4 | Tool Calling | Tool calls as memory allocations, token accumulation |
| 5 | Sub-Agents | Context isolation, the test runner problem, read vs. write heuristic |
| 6 | Message Passing | Erlang OTP actor model, message protocols, supervision trees |
| 7 | The Ralph Wiggum Loop | Crash-only agent design, fresh-context iteration, specs as memory |
| 8 | Context Management | malloc without free, compaction dangers, Write/Select/Compress/Isolate |
Requires Quarto and TinyTeX (quarto install tinytex).
quarto renderOutputs HTML site to _site/ and PDF. Pushes to GitHub trigger automatic builds via GitHub Actions.
Written by Claude (Anthropic), directed and edited by a human author across multiple sessions. Outline from Latent Patterns. Research drawn from 50+ sources including academic papers, engineering blogs, and practitioner reports.
