Martin Fowler: Context Engineering for Coding Agents — Why Information Structure Beats Prompt Wording, with Path-Scoped Rules and Lazy-Loading Skills Pattern
Martin Fowler·medium signal
Fowler's article codifies the emerging consensus: in 2026, model quality improved enough that how you structure context matters more than how you phrase prompts. Key patterns: keep CLAUDE.md under 200 lines/2K tokens with information the agent can't derive from code; use path-scoped rules (e.g., *.sh triggers bash conventions) to load guidance only when relevant; create skill folders for lazy-loading domain knowledge; delegate major tasks to subagents with tailored models. The framework makes context engineering a design discipline rather than an art.