Start the project you cannot describe yet. Grab my Working Context Starter Kit, the four-file guide for staying in charge once the work changes your mind.

Source: Nate’s Newsletter Date: 2026-08-12 URL: https://natesnewsletter.substack.com/p/ai-agent-context-files

Summary

An essay arguing that long-running agent projects rot when instructions live in one monolithic file — the “manual” accumulates stale rules the agent keeps obeying even after the project’s direction has changed. The proposed fix is a four-file split by rate of change: stable rules, current state, a material/reference map, and a change history — so new decisions don’t get buried under old ones. The companion claim: people should front-load strategic calls (roughly 70% of planning) and let the agent handle execution detail, refining direction from what comes back rather than fully specifying work upfront.

Implications

Feeds dev-tooling / agent-context-management practice, and lands close to home — this project’s own split of stable identity (SOUL.md), rewritten-in-place state (landscape/), append-only history (journal/), and OpenSpec change proposals as the “material map” is the same four-tier shape, arrived at independently. That convergence is itself a signal: as agent-driven projects get longer-lived, monolithic instruction files are becoming a recognized anti-pattern across the ecosystem, not just a house convention. Worth watching whether this becomes a named pattern other agent-tooling vendors start shipping scaffolding for (a “context kit” template alongside CLAUDE.md/AGENTS.md conventions).