Generating Fiction with Claude

I've written three novels with Claude. Not outlines or drafts that I then rewrote — actual novels, start to finish, where the prose came out of a collaboration between me and an LLM. This is either impressive or horrifying depending on your perspective, but the books are genuinely good and I'm proud of them.

The Fiction plugin started because I kept doing the same thing manually: plan the structure, develop the characters, write chapter by chapter, review, revise, review again. So I encoded the whole workflow into a Claude Code plugin with 22 specialised agents.

The interesting bit is how opinionated you have to be. Early on, the prose came out sounding like AI — hedging language everywhere, perfectly balanced dialogue, emotions explained rather than shown. "Perhaps she felt a sense of unease" instead of just writing the scene where you can feel it. So I built an anti-slop system. The reviewer explicitly flags hedging, over-explained emotions, and that generic sensory detail that sounds like a creative writing textbook.

I also built four literary critic personas — James Wood, Stephen King, Ursula Le Guin, Roxane Gay — each with a distinct voice and set of concerns. Wood cares about sentences and consciousness. King cares about story and character honesty. Le Guin cares about world-building as meaning. Gay cares about representation and emotional truth. Running the same manuscript through all four gives you wildly different feedback, and the combination is better than any single perspective.

The writer agent itself is tuned toward a specific sensibility: Rachel Cusk's precision, Jenny Offill's fragments, Kazuo Ishiguro's earned revelations. I didn't want generic literary fiction. I wanted prose with a point of view.

Continuity is the hardest problem. A novel is 70,000+ words across months of work. Characters change hair colour between chapters. Timeline inconsistencies creep in. A character knows something they shouldn't yet. I built a continuity agent that runs on Haiku for speed and checks every chapter against established facts. It catches things I'd never notice.

The workflow matters as much as the output. /fiction:plan for architecture. /fiction:character for the Want/Need/Lie framework. /fiction:outline for chapter beats. Then write, review, revise, review again. The review step is iterative — the reviewer reads a chapter, gives feedback, I revise, it reads again. This loop is where the quality actually comes from.

One thing that surprised me: the system enforces writing for the ear. Modern books become audiobooks, so the reviewer checks for clear dialogue attribution, distinctive character names, pronoun clarity. It's a constraint I wouldn't have thought of on my own and it makes the prose better across the board.

I open-sourced the whole thing. Whether you think AI fiction is legitimate or not, the engineering problem of maintaining voice consistency across a long work is genuinely interesting. And the books are good. I'll die on that hill.