Post
JA EN

Using Skills Without Writing Skills—Maximizing AI Delegation with the Official Meta-Skill

Posted
By
8 min read
Using Skills Without Writing Skills—Maximizing AI Delegation with the Official Meta-Skill
  • Target readers: Developers using AI coding tools like Claude Code, Cursor, etc.
  • Prerequisites: Basic experience with Claude Code
  • Reading time: 12 minutes

Overview

When it comes to leveraging Claude Code’s skill functionality, the common approach is to “write skills yourself.” However, you don’t need to write skills yourself.

Anthropic has published a meta-skill called skill-creator in their official repository. Using this skill, you can simply request “create a skill that does X,” and a skill following best practices will be generated.

This article introduces a workflow that uses the official skill-creator as a base, with project-specific rules added for operation.

Why “Not Write Skills”?

Problems with the Traditional Approach

Skill functionality tutorials typically present the following steps:

  1. Learn the SKILL.md structure
  2. Understand how to write frontmatter
  3. Design with Progressive Disclosure in mind
  4. Write SKILL.md yourself

This approach has problems:

  • Learning cost: You need to understand skill design principles
  • Lack of consistency: Quality varies between skills when written by humans
  • Maintenance burden: You need to recall design principles with every update

Paradigm Shift: Let AI Write Skills

As discussed in The Review Paradigm Has Changed—The Essential Difference Between Human Teams and AI Collaboration, in AI collaboration the human role has shifted from “executor” to “intent guarantor”.

The same applies to skill creation. Humans communicate “what kind of skill they want,” and AI handles “how to implement it.”

flowchart TB
    subgraph Traditional["Traditional Approach"]
        direction TB
        T1["Learn skill design principles"]
        T2["Write SKILL.md yourself"]
        T3["Discover problems"]
        T4["Fix yourself"]
        T1 --> T2 --> T3 --> T4
        T4 -.-> T2
    end

    subgraph Meta["Meta-Skill Approach"]
        direction TB
        M1["Introduce official skill-creator"]
        M2["Request 'create a skill that does X'"]
        M3["AI generates skill"]
        M4["Review and approve"]
        M1 --> M2 --> M3 --> M4
        M4 -.->|"Next skill"| M2
    end

What is the Official skill-creator?

Overview

Anthropic has published a meta-skill called skill-creator in the anthropics/skills repository1.

“skill-creator is used when users want to create a new skill (or update an existing skill)”

This skill includes:

  • Skill structure and best practices
  • Progressive Disclosure design principles
  • How to write frontmatter
  • Usage of bundled resources (scripts/, references/, assets/)
  • init_skill.py (skill initialization script)
  • package_skill.py (skill packaging script)

The skill functionality itself is positioned in Claude Code’s best practices2 as “a mechanism to provide domain-specific knowledge to Claude.” skill-creator is a meta-skill for delegating this skill creation process to AI.

Design Principles of skill-creator

The official skill-creator contains the following important principles1:

1. Conciseness is Priority

“Claude is already highly capable, so only add context Claude doesn’t already have. Consider whether each piece of information justifies its token cost.”

2. Setting Appropriate Degrees of Freedom

FreedomUse CaseExample
HighText-based instructionsWhen multiple approaches are valid
MediumPseudocode/scriptsWhen patterns exist with possible variation
LowSpecific scriptsWhen operations are fragile and consistency is needed

3. Progressive Disclosure

LevelLoading TimingSize Guide
MetadataAlways~100 tokens
SKILL.md bodyWhen skill triggers<5,000 tokens
Bundled resourcesAs neededUnlimited

Workflow: Operating with Official skill-creator as Base

Step 1: Introduce the Official skill-creator

First, obtain skill-creator from the official repository.

1
2
3
4
5

# Clone the repository
git clone https://github.com/anthropics/skills.git

# Copy skill-creator to your project
cp -r skills/skills/skill-creator .claude/skills/

Or, ask Claude Code directly:

1
2
3

Read the contents of
https://github.com/anthropics/skills/tree/main/skills/skill-creator
and place them in .claude/skills/skill-creator/

Step 2: Add Project-Specific Skill Design Rules (Optional)

The official skill-creator works as-is, but you can add project-specific rules regarding skill design itself if needed.

1
2
3
4
5

Add the following rules to skill-creator:

- When creating/updating skills, review against official documentation
  - https://github.com/anthropics/skills
  - https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices

Step 3: Create Other Skills Using skill-creator

Once skill-creator is introduced, you can create other skills like this:

1
2
3
4
5
6
7

Create a skill for writing articles.

Requirements:
- Evidence-based claims
- Clear references (with reliability levels)
- Search for similar articles first
- Use TB direction for Mermaid diagrams

The skill-creator includes an init_skill.py script, and skill names are given in kebab-case (e.g., article-writer). You don’t need to explicitly specify the save location—the AI determines an appropriate name based on the skill’s purpose and places it in .claude/skills/.

The AI uses skill-creator’s knowledge to generate a skill that follows best practices.

Step 4: Review the Generated Skill

Review the skill generated by AI. Points to check:

  • Is the intended functionality implemented?
  • Has unnecessary complexity been added?
  • Is Progressive Disclosure applied?

If there are issues, request corrections; if not, approve.

Step 5: Operation and Maintenance

When you discover problems while working with a skill, request skill updates on the spot.

1

Citation numbers are inconsistent. Update the skill so they're sequential.

Because you’re in the context of the work, the AI can determine which skill to update and what rules to add. If you fix it later in a different session, you’d need to explain the problem in detail, but on the spot, minimal instructions suffice.

Since Claude Code 2.1.0, automatic hot-reload of skills is supported3. Updated skills are immediately reflected without restarting the session, making this “discover→update→continue” cycle seamless.

Advanced User Practices

The Superpowers Case

Jesse Vincent has been practicing a similar approach with a workflow called “Superpowers”4:

“One of the first skills I taught was ‘how to create skills.’ This means when I want to add functionality like a git worktree workflow, I just explain how it should work, and Claude assembles it and adds notes to existing skills.”

Self-Improvement Loop

skill-creator is itself just another skill. Like other skills, you discover problems while using it and update it.

flowchart TB
    subgraph Loop["Self-Improvement Loop"]
        direction TB
        L1["Create skill"]
        L2["Work using skill"]
        L3["Discover problem"]
        L4["Update skill"]
        L1 --> L2 --> L3 --> L4
        L4 -.-> L2
    end

In Vincent’s Superpowers, the following self-improvements are made4:

  1. Learning from failures: “Pressure testing” skills using subagents
  2. Gradual instruction improvement: When failure patterns are discovered, update skills

Practical Considerations

Don’t Skip Reviews

Even when delegating skill creation to AI, review is essential. Points to verify:

  • Is the intended functionality implemented?
  • Has unnecessary complexity been added?
  • Does it comply with official best practices?

Prevent Skill Bloat

skill-creator includes the following constraints:

  • Keep SKILL.md under 500 lines
  • Separate details into references/
  • Don’t write the same information in multiple places

Maintain these constraints when adding project-specific rules.

Practice on This Blog

This blog adopts the following workflow:

  1. Initial setup: Introduce official skill-creator and add project-specific rules
  2. Article writing skill: Create article-writer skill using skill-creator
  3. Continuous improvement: Reflect problems discovered during writing into skills

For example, the article-writer skill has accumulated the following rules:

  • “Use TB direction for Mermaid diagrams” (added due to mobile compatibility issues)
  • “Sequential citation number rules” (added for reference traceability)
  • “Dark mode compatible style rules” (added due to visibility issues)

All of these are rules added by requesting AI updates when problems were discovered during work.

Summary

“Using skills without writing skills”—the essence of this approach is thorough AI delegation.

  1. Introduce official skill-creator: Leverage the meta-skill published by Anthropic
  2. Add project-specific rules: Customize as needed
  3. Delegate skill creation to AI: Only communicate “what kind of skill you want”
  4. Ensure quality through review: Verify and approve generated skills
  5. Delegate maintenance to AI too: Request updates when problems are discovered

With this workflow, humans can focus on the role of “intent guarantor.” There’s no need to memorize skill design principles. The knowledge accumulated in the official skill-creator is consistently applied to all skill creation.

The “externalization of learning” discussed in The Review Paradigm Has Changed is further accelerated by meta-skills. What humans need to learn is “what to create,” and “how to create it” can be left to AI.

See other articles related to this theme:

References

References corresponding to citation numbers in the text are listed in numerical order.

Additional References (Not Cited by Number in Text)

On Citation Accuracy: The references cited in this article are primarily composed of official documentation and official repositories. Claude Code’s skill functionality is a relatively new feature released after October 2025, and the information is current as of the writing date (January 2026). Please refer to official documentation for the latest specifications.

  1. skill-creator - Anthropic (2025). [Reliability: High] ↩︎ ↩︎2

  2. Claude Code Best Practices - Anthropic (2025). Best practices for Claude Code including skill functionality. [Reliability: High] ↩︎

  3. Claude Code Changelog - Anthropic (2025). Automatic hot-reload feature added in version 2.1.0. [Reliability: High] ↩︎

  4. Superpowers: How I’m using coding agents in October 2025 - Vincent, J. (2025). [Reliability: Medium-High] ↩︎ ↩︎2

AI・Technology
AI Claude Code Skills Automation Delegation Meta-Skill Workflow
This post is licensed under CC BY 4.0 by the author.