Fable Frame

A universal skill for Claude Code, Minimax Code, and any AI agent that accepts markdown skill files. Fable Brain behavior + Karpathy workflow + 50 production-ready prompt templates.

What this is

A single markdown file (SKILL.md) that gives any AI agent three things in one paste:

1. Fable Brain, 8 behavior modules that fix how the agent responds (no stalling, no padding, leads with the answer, only claims what the evidence supports).
2. Karpathy principles, mindset and workflow for orchestrating AI agents (program.md protocol, AutoResearch loop, parallel tracks, feature-level delegation).
3. 50 mega-prompt templates, production-ready prompts across 10 categories, invokable with use template [NAME].

Drop SKILL.md in as system instructions, a project's CLAUDE.md / AGENTS.md, or the first message in any chat. The agent gets a working behavior, a working workflow, and 50 ready prompts.

Installation

Claude Code

Clone directly into Claude Code's skills directory:

mkdir -p ~/.claude/skills
git clone https://github.com/AlessandroFare/fable-frame.git ~/.claude/skills/fable-frame

Or copy the skill file manually if you already have this repo cloned:

mkdir -p ~/.claude/skills/fable-frame
cp SKILL.md ~/.claude/skills/fable-frame/

Then run it with:

/fable-frame

Minimax Code

Add SKILL.md as a skill file in your Minimax Code config:

mkdir -p ~/.minimax/skills/fable-frame
cp SKILL.md ~/.minimax/skills/fable-frame/

Reference it in your session with load skill fable-frame, or paste SKILL.md directly into your project's agent instructions file.

OpenCode

Clone directly into OpenCode's skills directory:

mkdir -p ~/.config/opencode/skills
git clone https://github.com/AlessandroFare/fable-frame.git ~/.config/opencode/skills/fable-frame

Or copy the skill file manually if you already have this repo cloned:

mkdir -p ~/.config/opencode/skills/fable-frame
cp SKILL.md ~/.config/opencode/skills/fable-frame/

OpenCode also scans ~/.claude/skills/, so a single clone there is enough if you use both tools.

Any agent / manual

For any agent that accepts system instructions, paste the entire SKILL.md as your first message, as the system prompt, or in the custom instructions field:

• ChatGPT → Settings → Personalization → Custom instructions → paste the full SKILL.md
• Gemini → Gems → create a new Gem → paste the full SKILL.md as the Gem's instructions
• Cursor → Settings → Rules for AI → paste the full SKILL.md
• Windsurf → Customizations → paste the full SKILL.md
• Any API → pass the entire SKILL.md content as the system message

The file is self-contained. No external dependencies, no extra configuration.

Usage

Core behavior (always on)

Once the skill is loaded, the Fable Brain 8 modules apply to every response automatically. No trigger needed. The agent will lead with the answer, ground every claim, and only pause for destructive or irreversible actions.

Karpathy workflow (multi-step tasks)

For any task that is multi-step, agentic, or has an objective metric, the agent switches into the Karpathy workflow: program.md protocol for complex briefs, AutoResearch loop for measurable experiments, parallel tracks when you ask for several things at once.

To use the program.md protocol, paste a structured brief with these sections:

## Objective
[what success looks like]

## CAN change
- [in-scope item 1]
- [in-scope item 2]

## CANNOT change
- [hard boundary 1]
- [hard boundary 2]

## Process
1. [step 1]
2. [step 2]
3. [step 3]

The agent will follow the brief exactly until you tell it otherwise.

Template invocation

To use a template, type use template [NAME]. The agent will identify the missing placeholders, ask for them in a single message, then execute the template.

Example:

use template CODE-REVIEW

The agent will respond with something like:

Code Review template needs: LANGUAGE/FRAMEWORK, CODE PURPOSE, CODE. Paste them and I'll run the security and performance audit.

You then provide the placeholders:

LANGUAGE/FRAMEWORK: TypeScript / Next.js
CODE PURPOSE: Server action that creates a Stripe checkout session
CODE:
[paste code]

And the agent returns the full audit.

program.md example

For an AutoResearch-style task, you might paste:

## Objective
Reduce the API p95 latency below 200ms. Success = 7-day rolling average.

## CAN change
- Caching layer (TTL, key strategy, in-memory vs Redis)
- Query batching and N+1 cleanup
- Response payload size

## CANNOT change
- The public API contract
- The database engine
- Authentication flow

## Process
1. Run baseline. Record p50, p95, p99.
2. Make ONE change per iteration.
3. Log result in experiments.md.
4. If p95 improved, keep. If worse, revert.
5. Stop at 20 iterations or when p95 < 200ms.
6. Report the winning configuration.

The agent runs the loop, logs each iteration, and reports back.

What's inside

50 templates, 10 categories, paste-ready. The full list is at the bottom of SKILL.md (Quick Reference). Categories:

Sources

Built from three public sources:

• Anthropic Fable Brain, the official prompting guide distilled into the 8-module behavior system. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5
• Andrej Karpathy, No Priors Podcast (March 2026), the source of the mindset shifts, program.md protocol, and AutoResearch loop.
• Hyper Automation Labs 50 Mega-Prompts, where the 50 templates come from. https://hyperautomationlabs.co

Version history

• 1.0.0: Initial release. Fable Brain (8 modules) + Karpathy principles (4 mindset + 6 workflow) + 50 templates across 10 categories.

License

MIT