convez.ai
SkillEscrita & ConteúdoIntermediário
HarishDvs

Grizzly

HarishDvs/Grizzly

Editor IA para ficção: diagnóstico, estrutura e memória — você mantém a caneta

Ver no GitHub
InicialPluginScore · de 0 a 100

Editor de IA especializado em ficção que funciona como assistente editorial, não gerador de texto. Oferece diagnósticos, análise estrutural e memória de manuscrito através de propostas cirúrgicas e verificações de continuidade, mantendo a autoria e voz do escritor intactas.

Instalação

$npx convez-ai HarishDvs/Grizzly

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisão editorial de romances e novelas
  • Auditoria de continuidade em séries de capítulos longos
  • Diagnóstico estrutural de cenas e arcos narrativos
  • Análise de voz de escrita para calibração de sugestões

Veredito

Grizzly inverte a aposta dos produtos de IA para escrita: ao invés de gerar prosa, funciona como editor técnico com diagnósticos localizados, verificações de continuidade e propostas revisáveis. Abordagem sólida para autores que querem IA como ferramenta de refinamento, não substituição criativa.

README

A tradução para o português deste README ainda está na fila.

Grizzly

An open editor-in-a-skill for fiction writers — novelists, webnovel serial authors, anyone with a manuscript. Editors are expensive. AI is cheap but writes slop. Grizzly's bet: the value of AI for fiction is not generation, it is diagnosis, structure, and memory. The author keeps the pen.

Why another AI writing tool

Every AI writing product fails writers the same way: it produces prose, the prose has a voice, and the voice is the model's, not the author's. Grizzly inverts the design:

  • It never writes your book for you. Output is small, reviewable proposals: a few lines, side-by-side fixes, named findings. Never an unsolicited rewrite.
  • It diagnoses instead of generating. Findings with locations, severity tiers, and honest verdicts. No praise padding.
  • It remembers everything. An extractive card codex (chapters, arcs, characters, threads) so a 200-chapter serial costs the same to work on as a novella, and continuity errors get caught against recorded facts.
  • It learns your voice, not the model's. Setup derives a voice spec from your own passages; every suggestion anchors against it.

What it looks like in use

A surgical edit proposal (you accept, reject, or tweak each line):

Line 41 — emotion label.

- Mara felt the fear rise as the bell kept ringing.
+ The bell kept ringing. Mara counted the strokes.

A review finding:

[major] Ch 7, the cellar scene — buried landing. The chapter exists to deliver the brother's betrayal, and it arrives in the flattest sentence of the scene, mid-paragraph. The beat needs to land through a concrete moment, not a summary. Routing: structure (grizzly-plan) or line work (grizzly-edit).

A cold-read audit verdict:

I'd keep reading. The cellar scene pulls, but I skimmed the market chapter (no question raised for forty lines), and I cannot physically picture the innkeeper — she is a voice in white space. Highest-leverage change: give the market chapter one live thread.

How it works

flowchart LR
    M[Your manuscript] -->|grizzly-cards| C[Codex<br/>chapters, arcs,<br/>characters, threads]
    S[Your writing samples] -->|grizzly-init| V[VOICE.md<br/>your style spec]
    C --> X{Modes}
    V --> X
    X -->|edit, review, audit| F[Findings + side-by-side fixes]
    X -->|draft, plan, brainstorm| P[Small proposals, beat by beat]
    F --> A((You decide))
    P --> A
    A -->|accepted changes| M

Three separated quality layers run under everything, because they fail differently:

  1. Hygiene (the blacklist): catches AI tells — em dashes, "not X but Y", emotion labels, cinematic filler.
  2. Delivery (the six checks): catches the opposite failure, the spotless flat draft. Every scene needs the one place it lands.
  3. Architecture (the technique deck): 21 researched structural moves — false peace, perspective fracture, hook rotation, escalation — with mechanisms and failure modes.

Enforcement layer

The blacklist above is judgment. The enforcement layer is the part of it a machine can check without judgment, run as a deterministic gate.

tools/tell-scan.ps1 is a regex scanner for the mechanically-detectable subset of the blacklist: the em dash, the "too X to" construction, negative parallelism ("not X but Y", "nothing but", "without so much as"), and emotion or cognition report-frames ("she felt", "he realized", "Mara noticed"). Run it on a chapter:

powershell.exe -ExecutionPolicy Bypass -File tools/tell-scan.ps1 "chapters/ch07.md"

It accepts a single file or a *.md glob and prints one line per hit:

LEVEL  file:line  [rule]  | matched snippet

Two tiers:

  • HARD tells are unambiguous. Any HARD hit prints and the scanner exits 1 (so it can fail a commit hook or CI step).
  • REVIEW tells have a higher false-positive rate (rule-of-three, "seemed", cinematic-filler phrases). They print for human eyes but never fail the run.

The author's STANDARDS.md may relax or extend individual entries; the scanner encodes only the defaults.

It is the floor, not the verdict

The scanner is the deterministic floor of the hygiene layer. It is paired with two things regex cannot do: a separate-reader pass (re-read the draft cold, against the full blacklist) and the delivery pass (decks/delivery-pass.md), for everything mechanical detection cannot judge.

Regex matches form, not meaning. "He felt it pass" is physical sensation, not emotion-labeling, but the report-frame rule flags any felt. That is by design: the scanner errs toward flagging, and a human or an independent reader adjudicates. A clean scan means no obvious tells, not good prose. Clean is the floor, never the goal.

Optional: run it automatically (Claude Code hook)

tools/tell-scan-hook.ps1 lets a Claude Code PostToolUse hook run the scanner every time a .md file is written or edited. If the file picks up a HARD tell, the hook surfaces it back to the model so it self-corrects on the next turn. It is opt-in: nothing fires unless you add the snippet below to your settings.json under hooks.PostToolUse.

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit|MultiEdit",
        "hooks": [
          {
            "type": "command",
            "command": "powershell.exe -NoProfile -ExecutionPolicy Bypass -File \"C:\\Users\\you\\Grizzly\\tools\\tell-scan-hook.ps1\""
          }
        ]
      }
    ]
  }
}

Use the correct absolute path to tell-scan-hook.ps1 on your machine; the path above is an example. The hook reads the tool payload on stdin, scans only .md files, and exits 2 (feeding the tells back to the model) only on a HARD hit. It fails open: if the payload is malformed, the file is not markdown, or anything inside the hook breaks, it exits 0 silently. A broken hook never blocks writing.

What's inside

PieceWhat it does
skills/grizzlyCore: orientation, mode routing, the non-negotiable rails
skills/grizzly-initSetup wizard: interview + voice-spec derivation
skills/grizzly-cardsThe codex maker; bootstraps existing manuscripts
skills/grizzly-editSurgical line editing, diagnosis-first
skills/grizzly-draftBeat-by-beat drafting on strict rails
skills/grizzly-reviewFive-pass review: continuity, pacing, voice drift, hygiene, delivery
skills/grizzly-auditCold-read simulation: dread map, visualization check, belief check
skills/grizzly-planArc and chapter structure, technique-deck integration
skills/grizzly-brainstormCosted directions anchored in your canon
skills/grizzly-voiceVoice spec derivation, updates, drift reports
decks/Blacklist, delivery pass, dread toolkit, technique deck
packs/Genre packs: progression fantasy, webnovel serial, romance
docs/SPEC.mdThe tool-agnostic format spec
docs/example-codex/A working codex built from Treasure Island

Install

Option A — as a Claude Code plugin (recommended):

/plugin marketplace add HarishDvs/Grizzly
/plugin install grizzly@grizzly

This installs the skills and ships the decks, templates, and packs alongside them.

Option B — manual:

  1. Grab the latest archive from the Releases page (or clone the repo) and put it next to your manuscript.
  2. Copy the folders under skills/ into your project's .claude/skills/ (or your user-level skills directory).

Then, either way:

  1. Say "set up grizzly for my novel". The init wizard interviews you, derives your voice spec from your own writing samples, and scaffolds the codex.
  2. Have chapters already? Ask for a codex bootstrap; it cards your manuscript chapter by chapter.

The data layer (state files, codex, decks, packs) is plain markdown with zero harness-specific syntax; see docs/SPEC.md to adapt Grizzly to other agent runtimes.

Status

Pre-release: the audit mode has survived contact with a real manuscript; the remaining modes are being dogfooded now. See sprints/TODO.md for the live queue.

Built dogfood-first against a private reference manuscript that never enters this repo. Every example here is invented or public-domain, enforced by a sanitization gate run on every commit.

MIT licensed. Contributions welcome — genre packs and deck entries especially: see CONTRIBUTING.md.

Ver o original no GitHub

Bom para

Relacionados