The Fog — Book as Code

Long-running AI-assisted writing can depend on chat history or large manually assembled prompts. I tested whether durable repository context could support continuity across a large manuscript while leaving creative decisions with a human.

I separated Manuscript (established prose and events), Narrative (future plans and beats), Characters (reference), World (canon and knowledge boundaries), and State (compact current continuity). AGENTS.md defines retrieval, domain-specific authority, drafting, revision, review, and State maintenance. A chapter begins with: "Write Chapter X. Follow the repository instructions." The agent retrieves relevant context selectively instead of loading everything.

The instructions distinguish established facts, future plans, reference canon, and derived State. State serves as constraint memory, including current continuity and reveal boundaries, rather than creative source material. The agent updates materially affected State after drafting. Major contradictions and consequential new canon require human review; I review, edit, redirect, reject, or request revisions before accepting chapters.

At the time of this case study, the repository held approximately 29 manuscript chapters, 164,700 manuscript words, 541 numbered narrative planning beats, 37 world-reference files, 8 character-reference files, 10 State files, and a substantial repository-level agent instruction specification. The long-running project was maintained with repository-contained context and relatively little dependence on individual chat sessions.

Durable context can live outside chat history. Different kinds of information need distinct sources of authority and selective retrieval. Context also needs structure and lifecycle rules: State summaries help with continuity but can drift unless maintained. Human judgment remains necessary for creative direction and genuine ambiguity, while natural-language agent instructions remain probabilistic rather than enforced.