Skip to content
Finalist

Grimoire

Grimoire clones any writing voice.

Code Video Cerebral Valley

Core trick

Scores each rewrite candidate against a statistical style fingerprint across seven metrics, keeping the best candidate per round.

Tags

creative/musicagents

Description

Grimoire clones any writing voice. Paste text, get a statistical fingerprint (9 metrics: syllable density, sentence rhythm, punctuation cadence, lexical diversity). See it as an interactive 3D shape. Then watch a multi-round agent rewrite your prose toward any target author's profile. The agent generates 3 candidates per round at varied temperatures, scores each against the fingerprint, keeps the best, and feeds the gap back. Over 5 rounds the output converges toward the target. A blind LLM judge validates whether the result passes as the target author. Ships with 6 pre-computed voices (Hemingway, Poe, Twain, Kafka, Woolf, Fitzgerald). All metrics compute client-side, all AI streams in real time.
TheApexWu/grimoire★ 00 forksTypeScriptCreated Mar 21, 2026 · last push Mar 22, 2026

README excerpt

# Grimoire

Agentic writing workshop. Clone any writing voice with statistical proof.

Finalist at Zero to Agent (Vercel x DeepMind), March 2026.

## What it does

Paste any text. Grimoire computes a statistical fingerprint across 7 dimensions (monosyllable ratio, sentence length, sentence variation, comma density, semicolon usage, lexical diversity, conjunction frequency) and renders it as an interactive 3D voice shape. Then an agentic loop rewrites your text toward any target author's profile.

The agent generates 3 candidates per round at varied temperatures, scores each against the target fingerprint locally, keeps the best, and feeds the metric gap back into the next round. 5 rounds of selection pressure. A blind LLM judge with position-bias cancellation validates the result.

Ships with 6 pre-computed voices: Hemingway, Poe, Twain, Kafka, Woolf, Fitzgerald.

## Stack

- Next.js 16 (Turbopack)
- Vercel AI SDK v6
- Google Gemini 3.1 Pro Preview (generation) + Gemini 2.5 Flash Lite (judge, voice reading)
- Three.js / react-three-fiber (3D voice shapes)
- Recharts (convergence visualization)
- ElevenLabs (TTS)
- Clerk (auth)
- Supabase (persistence)

## Pipeline

```
Input (paste or photo OCR)
  -> Fingerprint (7 metrics, client-side, instant)
  -> 3D Voice Shape (icosahedron deformation)
  -> Agent Loop (5 rounds x 3 candidates x 3 temperatures)
     score locally, keep best, feed gap back
  -> Blind Judge (2 swapped runs, majority vote)
  -> Dashboard (convergence chart,

Team