How to Use Claude's Skill-Creator Skill to Build and Test New Skills

Developers can use Claude's skill-creator skill to generate, evaluate, and optimize custom skills by invoking the skill through API calls that trigger template-based generation and automated testing harnesses.

The asgeirtj/system_prompts_leaks repository reveals how Anthropic implements extensible capabilities through a skill system. One powerful meta-skill, skill-creator, enables developers to programmatically build and test new skills using Claude's API, loading templates from protected example directories and outputting validated skill definitions.

What Is the Skill-Creator Skill?

According to the source code in Anthropic/claude-opus-4.6.md (lines 473-475), the skill-creator is designed to "Create new skills, improve existing skills, and measure skill performance." The system prompt specifies that developers should invoke this skill when they want to:

• Create a skill from scratch
• Update or optimize an existing skill
• Run evals to test a skill
• Benchmark skill performance with variance analysis
• Optimize a skill's description for better triggering accuracy

Where the Skill-Creator Template Is Located

The skill references a canonical template stored at /mnt/skills/examples/skill-creator/SKILL.md (as noted in Anthropic/claude-opus-4.6.md, line 475). This template contains placeholders for:

• Goal description
• Input schema
• Tool calls (e.g., web_search, api_call)
• Evaluation criteria

The Anthropic/claude-cowork.md file (lines 491-497) also lists skill-creator in the skill catalogue with a "Path to skill-creator" reference, confirming its availability in the runtime environment.

How to Invoke the Skill-Creator Skill

Developers interact with skill-creator through Claude's API by specifying the skill name in the tools array. The repository shows two primary invocation patterns: creating a new skill and evaluating an existing one.

Creating a New Skill

To generate a new skill, send a request with the skill-creator tool and a natural language description of the desired functionality:

{
  "model": "claude-sonnet-4-20250514",
  "max_tokens": 2000,
  "messages": [
    {
      "role": "user",
      "content": "Create a new skill called *price‑fetcher* that calls our internal `/api/price` endpoint with a product ID and returns the price as JSON."
    }
  ],
  "tools": [
    {
      "type": "skill",
      "name": "skill-creator",
      "input": {}
    }
  ]
}

Claude processes this request by loading the template from /mnt/skills/examples/skill-creator/SKILL.md, filling in the placeholders with your specifications, and generating a complete skill file in the /mnt/skills/custom/ namespace.

Testing and Evaluating Skills

After creation, use skill-creator to run evaluations against sample inputs:

{
  "model": "claude-sonnet-4-20250514",
  "messages": [
    {
      "role": "assistant",
      "content": "Here is the `price-fetcher` skill you asked for…"
    },
    {
      "role": "user",
      "content": "Run a quick eval on `price-fetcher` with product IDs 101, 102, 103."
    }
  ],
  "tools": [
    {
      "type": "skill",
      "name": "skill-creator",
      "input": { "action": "eval", "skill": "price-fetcher" }
    }
  ]
}

Claude executes the skill against the provided test cases, computes success rates and latency metrics, and returns a variance analysis report (e.g., "All three calls returned valid JSON, latency < 120 ms").

The Skill Development Workflow

The complete workflow follows six steps based on the architecture described in the source files:

1. Trigger Detection – Claude's prompt engine scans requests for the skill-creator trigger defined in the skills list block of Anthropic/claude-opus-4.6.md.
2. Template Loading – The system loads the read-only template from /mnt/skills/examples/skill-creator/SKILL.md.
3. Skill Generation – Claude fills template placeholders (goal, input schema, tool calls, evaluation criteria) with user specifications and writes to /mnt/skills/custom/.
4. Testing – The evaluation harness runs the skill against sample inputs to capture accuracy and latency metrics.
5. Benchmarking – Variance analysis identifies performance inconsistencies across test cases.
6. Optimization – Developers request description updates or parameter adjustments, triggering iterative refinement through steps 3-5.

Safety and Sandboxing

The repository's filesystem configuration in Anthropic/claude-opus-4.6.md (lines 887-894) marks /mnt/skills/examples as read-only. This prevents accidental mutation of canonical templates. When skill-creator executes, Claude copies the template to a writable workspace before modification, ensuring the original SKILL.md remains pristine for future use.

Summary

• The skill-creator skill in Claude's ecosystem enables programmatic creation, testing, and optimization of custom skills through API invocation.
• Developers trigger skill generation by specifying the skill-creator tool in API requests, which loads templates from /mnt/skills/examples/skill-creator/SKILL.md.
• The workflow includes automated evaluation harnesses that benchmark performance with variance analysis and latency metrics.
• Filesystem sandboxing protects canonical templates by marking /mnt/skills/examples as read-only, forcing copies to writable workspaces before modification.

Frequently Asked Questions

What is the difference between skill-creator and regular Claude tools?

Regular Claude tools are built-in functions like web search or code execution. The skill-creator is a meta-skill that generates new skills by filling templates with custom logic. While standard tools perform tasks, skill-creator creates the infrastructure for new capabilities according to the template defined in /mnt/skills/examples/skill-creator/SKILL.md.

Can I modify the skill-creator template itself?

No. The repository configuration in Anthropic/claude-opus-4.6.md explicitly marks /mnt/skills/examples as read-only to prevent modification of canonical templates. If you need custom templates, create new skills in the /mnt/skills/custom/ namespace using skill-creator, or copy the template to a writable location and modify the copy.

How does skill-creator handle error handling in generated skills?

The template at /mnt/skills/examples/skill-creator/SKILL.md includes placeholders for evaluation criteria and error handling logic. When you invoke skill-creator with an evaluation request, Claude runs the generated skill against test cases and reports failure modes, latency spikes, and variance analysis. You can then request optimizations to improve error resilience based on these metrics.

Is skill-creator available in all Claude models?

According to the source files, skill-creator is defined in the system prompts for Claude Opus 4.6 (Anthropic/claude-opus-4.6.md) and appears in the Claude cowork configuration (Anthropic/claude-cowork.md). Availability depends on whether the specific model deployment includes the skills list containing skill-creator in its system prompt. The API examples use claude-sonnet-4-20250514, suggesting modern Sonnet models also support this capability.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}