All work
03Agent workflow

StoryWeaver

A Claude Code plugin that makes planning, continuing, and checking Chinese fiction projects beyond 300,000 characters repeatable.

Year
2026
Status
Open source · v1.2.3
Role
Independent — product, system design, and engineering
StoryWeaver — Agent workflow

Long fiction is a state-management problem

At 300,000 characters and beyond, a writing assistant cannot rely on whatever happens to fit in one prompt. Characters evolve, promises accumulate, timelines branch, and a small continuity error can contaminate later chapters.

StoryWeaver packages the work as a Claude Code plugin: explicit commands, specialized agents, durable project state, and a quality gate that can be rerun.

One repeatable chapter loop

The central workflow is deliberately visible:

plan → brief → continue → extract → verify → gate

Planning establishes intent. A brief assembles only the context needed for the next chapter. Continuation produces the draft. Extraction writes new facts back into distributed JSON state. Verification checks the draft against thirteen consistency dimensions before the quality gate allows the loop to continue.

This separation makes failures easier to locate. A context problem, a prose-generation problem, and a state-extraction problem no longer look like the same vague “AI made a mistake.”

Five agents, twenty-seven commands

The five agents divide responsibilities instead of asking one general agent to remember everything. Twenty-seven commands expose the system’s operations—from project setup and planning through continuation, extraction, checking, and maintenance.

Layer Product responsibility
Commands Give the writer explicit, repeatable actions
Agents Separate planning, writing, extraction, and verification roles
Distributed state Keep long-running project facts outside a single context window
Strict validation Stop malformed or inconsistent state from silently propagating

Severity that supports judgment

Consistency findings are grouped from P0 to P3, so the writer can distinguish a blocking canon conflict from a lower-priority polish issue. The thirteen checks turn “does this chapter still fit?” into a review surface rather than a feeling.

Version 1.2.3 is open source and installable today. The project received an Excellence Award at the MiraclePlus Vibeathon—but the stronger proof is the inspectable workflow itself: every major step has a command, an artifact, and a place for human judgment.

Toolbox

Claude Code · Markdown · JSON · Python · GitHub Actions

Next case studyFashionAI Studio

Contact

Ready to ship the next product?

I'm open to AI product collaborations, open-source work, and prototype-to-production conversations.

Email me