StyleLoop is a self-improving AI writing engine. It learns your voice from every edit you make and compounds it into a personal style layer on top of Claude. No fine-tuning. No prompt engineering. Every draft lands closer to you than the last.
Every writer using AI spends the session rewriting its output back into their own voice. Then those edits are thrown away, and tomorrow starts from zero.
Every model has a house voice. Yours gets averaged out.
Your corrections don't carry over. Same fixes, every single draft.
Prompt tweaks feel better, but nobody checks whether they actually are.
Fine-tuning bakes a style into weights: slow, expensive, opaque, and thrown away with every new model. StyleLoop keeps Claude frozen and recursively improves everything around it: the rules, the memory, and the gate that decides what stays. It is recursive self-improvement at the harness level, measured on your own edits.
| Fine-tuning | StyleLoop | |
|---|---|---|
| Data needed | Thousands of examples | A handful of edits |
| Cost to adapt | GPU training runs | Zero GPUs, minutes per round |
| What changes | Opaque weights | Readable rules you can edit |
| New model release | Retrain from scratch | Your style carries over |
| Proof it helped | Hope | Gated on your own edits |
The model stays fixed. What evolves is a readable style adapter wrapped around it: your long-term style memory, owned by you.
A few instructions or a style guide. Claude writes the first draft.
Every change you make is captured as a labeled example of your taste.
Repeated edits become plain-language rules, each backed by the edits that support it.
A new adapter must beat the current one on drafts you already corrected, or it is discarded. Then the flywheel spins again.
Asking an AI to grade "good writing" just swaps your taste for its taste. StyleLoop holds back drafts you already corrected, has the candidate adapter re-edit them, and ships it only if it lands measurably closer to your real edits.
Share of characters the writer changed in each draft (lower is better). p = 0.008 on both comparisons. B's adapter made Writer A edit 3.6× more, so every adapter is personal. October 2026 experiment with simulated writers who had hidden preferences; the learner saw only their edits. Validation with real writers is the next step.
Claude drafts every scene with your current adapter and the story's state: facts, beliefs, open threads.
Before you edit, Claude predicts which sentences you will change. The hit rate proves whether it really knows your taste.
Claude proposes small rule edits from your corrections. The gate decides what stays. Every call is logged with its cost.
Fiction and web-novel writers, the most demanding voices to match.
Teams with a house style: video scripts, marketing, newsletters.
Plug a learned voice into any Claude-powered app.
We are opening StyleLoop to a small group of writers first.