Claude Fable 5 Consult Skill
kei99-web3/codex-claude-fable-5-consult-skill
Consulte Fable 5 localmente com packing inteligente de context e caps de custo
Ver no GitHubSkill para Codex que consulta Claude Fable 5 via Claude Code CLI local, com empacotamento de prompts consciente de tokens, bloqueio de segredos e controle de custos. Ideal para trabalhos de revisão técnica: reviews de estratégia, implementação, arquitetura e risco.
Instalação
npx convez-ai kei99-web3/codex-claude-fable-5-consult-skill1 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
Casos de uso
- Reviews de arquitetura e implementação técnica
- Verificação de risco e segurança em decisões de código
- Segunda opinião em estratégia de desenvolvimento com otimização de tokens
Veredito

Skill bem estruturado para quem usa Claude Code CLI e quer consultas determinísticas em Fable 5 sem overhead de tokens. Solução madura para review loops, mas exige setup local robusto (CLI autenticado, acesso a Fable 5, Python 3.9+).
README
A tradução para o português deste README ainda está na fila.
Claude Fable 5 Consult Skill for Codex
A Codex skill for asking Claude Fable 5 for a second opinion through your own Claude Code CLI, with token-conscious prompt packing, reliable model naming, UTF-8 prompt files, cost caps, and saved outputs.
It is designed for review-style work: strategy reviews, implementation reviews, risk reviews, architecture reviews, and other high-stakes reasoning passes where Codex should ask Fable, then verify and apply the result locally.
If this saves you time or tokens, starring the repo helps other Codex users find it.
Important: Whose Claude/Fable Account Is Used?
This repo does not proxy through the maintainer's API, key, subscription, or server.
It runs the claude command on your own machine. That means:
- you must have Claude Code CLI installed and authenticated locally;
- you must have access to
claude-fable-5; - any usage limits or costs are charged to your own Claude/Anthropic account or provider setup;
- your prompts are sent by your local Claude Code CLI, not by this repository.
What It Solves
- Normalizes unreliable aliases like
fableandfable-5toclaude-fable-5. - Sends UTF-8 prompt files through stdin instead of fragile inline shell prompts.
- Builds compressed prompt packs from selected files so Fable sees the important context, not your whole workspace.
- Adds explicit
--max-budget-usdcost caps. - Saves raw JSON output and extracts the
resultfield into a Markdown sidecar. - Disables Claude tools by default, with optional read-only file review.
- Blocks obvious secret-like values before sending the prompt.
Requirements
- Codex with local skills support and shell access.
- Claude Code CLI installed in the same environment Codex can run.
- Claude Code CLI authenticated with your own account.
- Access to
claude-fable-5from that Claude Code CLI account/provider. - Python 3.9+ for the cross-platform script.
This repository was verified with Claude Code CLI 2.1.152. Claude CLI flags can change, so rerun the no-cost preflight after CLI upgrades.
Claude model availability and pricing can change. Check Anthropic's current model and pricing docs before heavy use.
Install
macOS/Linux:
git clone https://github.com/kei99-web3/codex-claude-fable-5-consult-skill.git \
"$HOME/.codex/skills/claude-fable-5-consult"
Windows PowerShell:
git clone https://github.com/kei99-web3/codex-claude-fable-5-consult-skill.git `
"$env:USERPROFILE\.codex\skills\claude-fable-5-consult"
Then ask Codex:
Use $claude-fable-5-consult to ask Claude Fable 5 for a second opinion on this plan.
Token-Conscious Prompt Packing
The skill works best when Codex sends Fable a compact decision packet instead of a broad directory dump. Use the included packer to assemble a smaller prompt from only the files that matter:
python scripts/build_fable_prompt_pack.py \
--goal "Review this implementation plan and identify the next best action." \
--context-file ./README.md \
--context-file ./SKILL.md \
--out-file ./prompt_pack.txt \
--max-chars-total 12000 \
--max-chars-per-file 4000
Then send the packed prompt:
python scripts/invoke_fable5_consult.py \
--prompt-file ./prompt_pack.txt \
--out-file ./fable_response.json \
--max-budget-usd 0.75
This is deterministic context compression: it strips excess whitespace, caps each file, records what was included/truncated, and blocks obvious secret-like values before building the prompt. It does not replace careful judgment; choose the smallest useful context set.
Manual Usage
Create a prompt file:
cp examples/fable_prompt_template.txt prompt.txt
Run a no-cost preflight:
python scripts/invoke_fable5_consult.py \
--prompt-file ./prompt.txt \
--out-file ./fable_response.json \
--no-execute
Run a real Fable call:
python scripts/invoke_fable5_consult.py \
--prompt-file ./prompt.txt \
--out-file ./fable_response.json \
--max-budget-usd 0.75
Read-only file review:
python scripts/invoke_fable5_consult.py \
--prompt-file ./prompt.txt \
--out-file ./fable_response.json \
--add-dir /path/to/project \
--allowed-tools Read \
--max-budget-usd 1.00
The script accepts the friendly --allowed-tools flag and forwards it to Claude Code CLI's --allowedTools option.
Windows wrapper:
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\invoke-fable5-consult.ps1 `
-PromptFile .\prompt.txt `
-OutFile .\fable_response.json `
-MaxBudgetUsd 0.75
Safety Notes
Fable output is a review artifact, not an automatic instruction to mutate systems. Let Codex verify the recommendation against local evidence before editing code, trading, deploying, posting, sending, or changing production settings.
Do not put secrets in the prompt file. The script blocks common secret-looking patterns, but it cannot guarantee complete data-loss prevention.
Star-Friendly Use Cases
- Ask Fable to review a Codex implementation plan before editing.
- Compress a few local reports into a small Fable review packet.
- Get a second opinion on architecture, tests, risk, or launch readiness.
- Keep the expensive model focused on the decision instead of raw workspace noise.
License
MIT