Claude Fable 5 Consult

Claude Fable 5 Consult

Use this skill when a user asks Codex to ask Claude Fable 5, Fable, or another Claude frontier model for a second opinion, review, or recommendation.

Core Rules

• Use the full model name claude-fable-5. Short aliases such as fable or fable-5 may fail in Claude Code CLI.
• Prefer prompt files over direct command arguments. This avoids shell quoting, long prompt, and Japanese/Unicode encoding failures.
• Set a cost cap for every real call with --max-budget-usd.
• Use --no-execute for smoke tests. Tiny paid Fable calls can still create Claude Code prompt/cache overhead before visible output.
• Use the prompt packer when context is large; Fable should receive a compact decision packet, not an unbounded workspace dump.
• Keep Fable read-only unless the user explicitly asks otherwise. Codex should own file edits, implementation, tests, commits, and final judgment.
• Do not send .env files, API keys, tokens, private keys, credentials, or other secrets.
• For production systems, separate review/recommendation from runtime change; do not change live systems without the user's approval boundary.
• This skill invokes the user's local claude CLI and account. It does not use the skill author's API key, subscription, server, or billing.

Standard Workflow

1. Create a compact UTF-8 prompt file with:

• objective and decision needed
• compressed facts, metrics, and relevant file paths
• safety and approval boundaries
• exact output format requested

1. Run the bundled script.
2. Read the saved response and let Codex verify it against local evidence.
3. Treat Fable output as a strong external review, not as automatic authority.
4. Implement only recommendations that pass local checks and the user's approval boundary.

Recommended Invocation

Cross-platform Python:

python scripts/invoke_fable5_consult.py \
  --prompt-file ./prompt.txt \
  --out-file ./fable_response.json \
  --max-budget-usd 0.75

Windows PowerShell wrapper:

powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\invoke-fable5-consult.ps1 `
  -PromptFile .\prompt.txt `
  -OutFile .\fable_response.json `
  -MaxBudgetUsd 0.75

No-cost preflight:

python scripts/invoke_fable5_consult.py \
  --prompt-file ./prompt.txt \
  --out-file ./fable_response.json \
  --no-execute

File-aware read-only 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

Prefer compressed-context prompts over broad file access unless Fable truly needs source files.

Token-Conscious Prompt Pack

When the user asks to save tokens or when the context has multiple files, build a compact prompt first:

python scripts/build_fable_prompt_pack.py \
  --goal "Review this plan and recommend 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 invoke Fable with --prompt-file ./prompt_pack.txt.

The packer strips excess whitespace, caps each file, marks truncation, and blocks obvious secret-like content. Codex still needs to choose the smallest useful file set.

Prompt Pattern

You are a senior reviewer for [domain].

Objective:
- [decision needed]

Context:
- [compressed facts, metrics, constraints]

Safety/approval boundaries:
- [what must not be changed automatically]

Ask:
- Provide [N] prioritized recommendations.
- For each: goal, why now, implementation/validation, promotion condition, kill condition.
- Name the one action you would choose this week.

Output:
- [language and format]
- Be concrete and falsifiable.
- Do not invent unavailable data.

Failure Handling

• If a model error mentions fable or fable-5, rerun with claude-fable-5.
• If the shell reports missing input, use --prompt-file instead of inline prompt arguments.
• If the command returns no visible text, use JSON output and inspect the saved .json plus generated .md sidecar.
• If error_max_budget_usd appears on a tiny prompt, it may be Claude Code prompt/cache overhead. Use --no-execute for checks and a realistic budget for actual work.
• If Fable gives a plausible but unverified recommendation, turn it into a local hypothesis and test it before acting.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kei99-web3
• Source: kei99-web3/codex-claude-fable-5-consult-skill
• License: MIT

Install and usage instructions live in the source repository linked above.