domains / agent-tooling / create-agent-skill / skill-creator

skill-creator

Skill card

combined

The four copies overlap on the core loop (capture intent, interview, write SKILL.md, test/iterate) but each contains material steps or constraints the others lack. The Anthropic copy is the only one with a full quantitative eval/benchmark loop and a description-improver step. The OpenAI copy is the only one with a 'degrees of freedom' design principle and a specific agents/openai.yaml UI metadata generation step. The OpenChatCut copy is the only one with environment-specific constraints (slug whitelist, body size limits, manage_skill tool, security rules). The finance-skills copy is the only one with a structural pattern taxonomy (Linear, Router, etc.) and a runtime detection flow. Crowning one would hide these real alternatives.

Repositories

Other copies

Combined skill

---
name: skill-creator
description: Create, improve, and evaluate agent skills. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, or turn a repeated workflow into a skill.
---

# Skill Creator

A skill for creating new skills and iteratively improving them.

## Process

1. **Capture Intent**: Understand what the skill should do, when it should trigger, and the expected output format. Extract workflow details from conversation history if available.
2. **Interview & Research**: Ask about edge cases, input/output formats, and dependencies. Research available tools or MCPs if needed.
3. **Plan Architecture**: Choose a structural pattern (Linear, Router, Methodology, etc.) and plan the outline. Decide what goes in SKILL.md vs references/.
4. **Write SKILL.md**:
   - **Frontmatter**: `name` (kebab-case) and `description` (pushy, specific trigger contexts).
   - **Body**: Imperative instructions. Keep concise. Use examples.
   - **Resources**: Add `scripts/` for deterministic tasks, `references/` for detailed docs, `assets/` for output files.
5. **Test & Iterate**: Create test prompts. Run them. Evaluate results qualitatively and quantitatively. Rewrite the skill based on feedback.
6. **Optimize Description**: Use a description improver script to optimize triggering accuracy.

## Design Principles

- **Concise is Key**: Only add context the model doesn't already have.
- **Degrees of Freedom**: Match specificity to task fragility. High freedom for heuristic tasks, low freedom for fragile/consistency-critical tasks.
- **Runtime Adaptation**: Detect installed tools and auth state at runtime. Offer fallback paths.

## Environment Constraints

- **Slug**: Kebab-case, alphanumeric and hyphens only.
- **Body Size**: Keep under ~200-250 lines. Split detail into references/ if larger.
- **Security**: Do not include malicious content or prompt-injection payloads.

## Conflicts

- **openai/skills**: Requires generating `agents/openai.yaml` UI metadata.
- **0xsline/OpenChatCut**: Requires using `manage_skill` tool and adhering to specific slug/size limits.
- **anthropics/skills**: Requires running quantitative evals with `eval-viewer/generate_review.py`.
- **himself65/finance-skills**: Requires choosing a structural pattern and planning a runtime detection flow.