Experiment
yet-another-writer
activeA long-form story generation system built around an MCP server for story arc, state, characters, locations, and chapter review. Chapters are produced sequentially by Task subagents that read only the prior chapter plus shared state, targeting novel-length output without a single mega-context.
How far can context discipline carry generation quality? Instead of stuffing the whole novel into the prompt, the system keeps the canonical state in MCP tools and hands each chapter-writing subagent only what it needs: the global guidelines, the story arc, the previous chapter, and the characters and locations that actually show up.
Questions it’s starting to answer:
- State files as memory: whether an MCP-backed story bible can carry continuity across a lot of short-context generations
- Sequential subagent chains: why generating chapters in parallel breaks continuity in ways the architecture has to forbid
- Review as a separate role:
chapter_reviewas its own MCP tool, not something the writing agent does to itself
Field notes
No posts tag this experiment yet. When a post does, add experiments: [yet-another-writer] to its frontmatter.