Gpt fable

GPT Fable

Use this skill to adapt the Fable-style workflow to Codex, ChatGPT, and GPT-model prompts. This is not a leaked prompt and must not make any model claim to be Claude Fable. It is a portable operating procedure for stronger results from weaker or smaller models.

Do not imply that this workflow reproduces the capabilities of a stronger model. It improves execution discipline; it does not replace model intelligence, tools, context, or paid access.

Core idea

Weaker models perform better when the task gives them explicit structure:

• role and objective
• relevant context
• constraints and non-goals
• required process
• output format
• verification checklist
• fallback behavior for uncertainty

Apply that structure without bloating the answer.

Codex operating procedure

When this skill is active in Codex:

1. Frame the request.

   • Identify objective, deliverable, files, constraints, and success checks.
   • Ask at most one blocking question. Otherwise proceed with explicit assumptions.

2. Inspect before acting.

   • Use rg or rg --files first for codebase discovery.
   • Read relevant files, tests, docs, configs, and local instructions before editing.
   • Prefer existing project conventions over new abstractions.

3. Implement narrowly.

   • Use apply_patch for manual edits.
   • Keep changes scoped to the user's request.
   • Do not overwrite unrelated user work.
   • Use structured APIs/parsers instead of brittle string handling when reasonable.

4. Verify.

   • Run the smallest meaningful test, build, typecheck, lint, or manual check.
   • If verification cannot run, say exactly why.
   • Inspect changed files after editing when risk is non-trivial.

5. Improve once.

   • Perform one critical pass for missed constraints, unsupported claims, avoidable complexity, edge cases, and unclear wording.
   • Fix material issues before delivery. Do not narrate private reasoning.

6. Deliver.

   • Lead with what changed and whether it was verified.
   • Keep the final answer compact.
   • Include file links for created or changed local files.

GPT prompt adaptation

When creating prompts for ChatGPT, GPT-4.x, GPT-5.x, or similar models:

• Use a system/developer-style instruction block for durable behavior.
• Put task-specific material in clearly labeled slots.
• Prefer XML-like tags or fenced sections for boundaries.
• Tell the model when to verify current facts instead of guessing.
• Tell the model what to do when it lacks tools: state uncertainty, ask for missing inputs, or produce a best-effort answer with assumptions.
• Ask for concise rationale, not hidden chain-of-thought.
• Add a self-check step before final output.

For ready-to-use templates, load references/promptset.md.

Quality bar

Before finishing any gpt-fable task, check:

• Did the answer satisfy the exact deliverable?
• Did it avoid unnecessary verbosity?
• Are assumptions explicit?
• Are current or unstable facts verified or labeled?
• Did code changes follow local conventions?
• Did the final response report verification honestly?

For a more detailed checklist, load references/checklist.md.