convez.ai
SkillDesenvolvimento WebAvançado
zakarya526

Fable Skill

zakarya526/fable-skill

Audita código e gera planos precisos para agentes executarem

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que integra Claude (Fable 5) para auditar codebases em profundidade e gerar planos executáveis por qualquer agente de IA. Funciona com 70+ agentes de código e divide o trabalho inteligentemente: Fable 5 faz a análise e especificação, enquanto executores como Claude Opus implementam de forma isolada.

Instalação

$npx convez-ai zakarya526/fable-skill

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
v1.0.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Auditoria automática de codebases antes de refatorações
  • Geração de especificações técnicas detalhadas para implementação por agentes
  • Identificação priorizada de problemas de performance, segurança e testes
  • Planejamento de novas funcionalidades baseado na análise do repositório

Veredito

Ferramenta sofisticada que resolve um problema real: separar julgamento (Fable 5) de execução (agentes), evitando alucinações em análises complexas. A abordagem de 'never edit source' é sensata. Requer setup em ecossistema de skills compatíveis e CLI específico.

README

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

fable-skill

Works in 70+ AI coding agents

The goal: spend your best model's judgment once. Fable 5 reads a codebase deeply and writes plans precise enough that a capable executor (Opus 4.8) — or any agent, or a human — can build them safely. The skill never edits your code itself.

you          →  /fable-skill                (Fable 5 audits & advises)
plans/       →  001-fix-n-plus-one.md        (self-contained specs)
other agent  →  implements, tests, ships     (Opus 4.8 executes)

Install

npx skills add zakarya526/fable-skill                  # pick your agents (space to select)
npx skills add zakarya526/fable-skill -a claude-code   # straight into Claude Code

Usage

/fable-skill                  full audit → ranked findings → plans
/fable-skill quick | deep     cheaper / more exhaustive pass
/fable-skill security         focus one category (also: perf, tests, bugs…)
/fable-skill branch           audit only what your branch changes
/fable-skill next             feature ideas grounded in the repo
/fable-skill plan <desc>      spec one thing, skip the audit
/fable-skill execute <plan>   dispatch an executor, review its diff
/fable-skill reconcile        refresh the backlog
/fable-skill … --issues       also publish plans as GitHub issues

How it works

  1. Recon — maps the stack and the exact build/test/lint commands (they become verification gates).
  2. Audit — parallel subagents across 9 categories; every finding cites file:line.
  3. Vet — re-reads each finding to drop false positives before you see them.
  4. Prioritize — ranks by leverage (impact ÷ effort, weighted by confidence); you pick what to plan.
  5. Plan — one self-contained file per finding in plans/, with steps, commands, and STOP conditions (example).

Why the model split

Fable 5 — the frontier planner — does the judgment that compounds (understanding, vetting, specifying). A capable executor (Opus 4.8; Sonnet 4.6 for lighter work) implements the spec in an isolated worktree, and Fable 5 reviews the diff. You get frontier judgment where it matters and a dedicated builder where the work is already specified.

Guarantees

  • Never edits your source — the only writes go to plans/; executors work in disposable worktrees, and merging is always yours.
  • Never mutates your working tree — read-only analysis only.
  • Never prints secrets — locations and rotation advice only.

Credits & License

A Fable 5 rebrand/improved version of improve by shadcn — original concept and architecture are shadcn's, used under MIT.

MIT © Muhammad Zakarya. Original work © shadcn — see LICENSE.md.

Ver o original no GitHub

Bom para

Relacionados