A Human-in-the-Loop Writing Assistant

Positioning

For whom: individual authors who publish under their own name and want the writing to sound like them.

What it replaces: "generate and post" tools, whose output shares a uniform AI voice readers spot at once. The most important decision in this system is that the AI never finalises: the model may generate, but the person whose name is on it must actually have edited it.

Why it earns its place: it turns "remove the AI voice" into an explicit, tool-backed step instead of relying on the author's self-discipline.

Main modules

• Find the angle: cut a topic through a library of mental-model lenses and rank the angles.
• Draft: a structural skeleton and first draft for the chosen angle — explicitly a skeleton, not a final.
• Strip the AI voice, tune the register (the core): a regex pre-scan flags five classes of translationese and AI-tells; a person rewrites each; the register lands between academic and plain.
• Typeset: Markdown → inline CSS, because the target editor ignores external stylesheets.
• Voiceover (optional): local TTS, offline and free.
• Knowledge loop: topics come from the LLM Wiki; finished pieces flow back in.

Stack

• A chain of Claude Code skills, one per step, each callable alone.
• Python for the pre-scan and typesetting.
• Local TTS with zero-shot voice cloning, no cloud.
• LLM Wiki, shared with the workbench.

Constraint and status

An unverified personal account has no publishing API, so the last step is copy-paste by hand. That isn't unfinished — the road is closed, and saying so beats faking automation. The system is complete and in use.