skill-creator - A meta Skill launched by Anthropic, supporting Skill creation

What is skill-creator

skill-creator is a "meta Skill" officially launched by Anthropic, designed specifically for creating, iterating, and optimizing other Agent Skills. Through interactive guidance, skill-creator encapsulates the entire Skill development process—from requirement capture, SKILL.md writing, and test case generation to performance evaluation—into a reusable standard workflow, greatly lowering the barrier to custom Skills. It is currently hosted in the anthropics/skills official repository and is widely recognized as a fundamental must-install component in the Skills ecosystem.

Main features of skill-creator

• Create new Skills: Through natural language descriptions of requirements, automatically generate SKILL.md files and directory structures that comply with specifications, including YAML frontmatter, trigger conditions, and execution instructions.
• Modify and optimize: Supports iterative improvements to existing Skills, including rewriting description copy, adjusting trigger accuracy, and streamlining context loading strategies.
• Automated evaluation: Built-in eval framework can batch-run test cases and generate quantitative metrics, supporting qualitative review and variance analysis to ensure stable Skill output.
• Description optimization: Specialized tuning for the Skill's description field, using a "pushy" strategy to solve AI undertrigger issues and improve automatic invocation hit rates.
• Progressive loading: Follows the Progressive Disclosure principle, initially exposing only metadata to the Agent and loading full content only after determining relevance, saving context window space.

Technical principles of skill-creator

• Meta-Skill architecture: skill-creator is essentially a SKILL.md file of about 500+ lines, composed of YAML frontmatter (name, description) and Markdown body instructions. Through structured instructions, it teaches Claude how to act as a "Skill development coach," with built-in professional workflows, domain best practices, and automated toolchains, transforming a general assistant into a professional Skill architect.
• Progressive Disclosure loading system: Adopts a three-tier layered loading strategy to avoid context window overload:
  • L1 Metadata: name + description reside permanently in context (about 30–100 tokens), serving as a "dynamic catalog" for the Agent to globally perceive the existence of thousands of Skills
  • L2 Instruction body: SKILL.md is injected only upon triggering (recommended <500 lines), providing procedural knowledge for task execution
  • L3 Bundled resources: scripts/, references/, and assets/ are loaded on demand with no capacity limit; Python/Bash scripts in scripts/ can be executed directly without being loaded into context, achieving infinite expansion at the token level
• Description-as-Trigger mechanism: The description field is the core trigger of a Skill; Claude determines activation timing through semantic matching:
  • Adopts a "pushy" strategy, explicitly listing all potential trigger scenarios (including indirect needs where the user does not directly mention the Skill name), addressing the model's undertrigger tendency
  • Length limit of 1024 characters; recommended to keep around 100 words to balance recall rate and context overhead
  • Underneath, activation and execution of all Skills are uniformly scheduled through a Meta-Tool named Skill , which differs from the independent invocation method of traditional tools

How to use skill-creator

• Installation: In an Agent tool that supports Skills (such as Claude Code), run one of the following commands:
  • Command-line installation:npx skills add anthropics/skills --skill skill-creator
  • Natural language installation: simply tell the Agent to help me install this Skill.
• Create a Skill: After installation is complete, describe your requirements in natural language, for example "create a Skill that converts Word to PPT." skill-creator will automatically guide you to clarify trigger conditions, input/output formats, and edge cases, and generate SKILL.md along with supporting scripts.
• Evaluation and iteration: After generating the first draft, skill-creator will suggest writing test cases and running eval, and by using eval-viewer/generate_review.py to view results, optimize based on the feedback loop until it meets the standard .

Key information and usage requirements of skill-creator

• Runtime environment: Must be used with a client that supports Agent Skills, such as Claude Code, Codex, or OpenClaw, etc.
• Description field specification: description is the core trigger mechanism, and it needs to declare both the Skill's functionality and its trigger scenarios. It is recommended to write it specifically and slightly "pushy" to avoid the AI missing triggers
• Length limit: description is recommended to be kept within 1024 characters, and the main body of SKILL.md is recommended to stay within 500 lines. Content that is too long should be moved to references for on-demand loading

Core advantages of skill-creator

• Zero-code threshold: Fully natural language interaction, no need to hand-write complex configurations, and the AI guides you through Skill architecture design.
• Backed by official standards: Directly inherits Anthropic's official Skill specifications and best practices, and the outputs are compatible with Claude Code and mainstream Agent platforms .
• Closed-loop iteration: Built-in "design-test-evaluate-rewrite" loop, avoiding the problem of being difficult to maintain after one-time development .

Project address of skill-creator

• GitHub repository：https://github.com/anthropics/skills/tree/main/skills/skill-creator

Comparison of skill-creator with similar competitors

Application scenarios of skill-creator

• Vertical business encapsulation: Solidify professional knowledge such as internal enterprise SOPs, code review standards, and brand design guidelines into reusable Skills to reduce team knowledge transfer costs.
• Development toolchain extension: Create dedicated Skills for specific technology stacks (such as React, Python, and the MCP protocol) to achieve one-click generation of boilerplate code and automatic completion of best practices.
• Repetitive task automation: Encapsulate high-frequency mechanical operations such as daily report generation, PDF processing, data extraction, and format conversion as Skills, triggered with one click via natural language instructions.
• Agent capability orchestration: Combine multiple basic Skills to build complex workflows, such as full-chain automation of "first scrape web pages → extract key data → generate PPT → send email".
• Team collaboration standardization: Unify coding standards, document templates, and testing strategies within the team to ensure consistent output style and quality when different members use the Agent.