Skip to content
Coding

Skill Creator

Draft, evaluate, and refine a reusable skill with task examples and feedback.

Install with Skills CLI
npx skills add anthropics/skills --skill skill-creator

What it does

A skill can describe a useful workflow yet trigger at the wrong time or fail on realistic tasks. Skill Creator combines drafting with examples, baseline comparisons, and iterative evaluation.

What changes after install

  • Clarifies the skill's intended tasks, inputs, outputs, and triggering conditions.
  • Uses test prompts and comparison runs to examine behavior rather than only wording.
  • Provides evaluation review tools and a loop for refining instructions and descriptions.

Install

Run this command with Node.js 22.20 or later and select your agent in the Skills CLI.

Install with Skills CLI
npx skills add anthropics/skills --skill skill-creator

Before using Skill Creator

  • Keep the skill’s scripts, evaluation viewer and reference files together. The documented benchmark aggregation and review-generation commands need Python.
  • Prepare representative test prompts and access to the tools the proposed skill actually needs; evaluation workflows depend on the host’s available agent capabilities.

Read the reviewed setup instructions for the selected workflow.

When to use it

  • Turning a recurring workflow into a reusable SKILL.md package.
  • Improving an existing skill whose trigger description or task performance is inconsistent.

When not to use it

  • Assuming a skill is reliable because its instructions sound complete.

Begin with a repeated task

A useful brief is: “Create a skill for turning support tickets into a weekly product feedback report.” Provide a few representative tickets, the expected report structure, and examples of requests that should not activate the skill. This gives the authoring process a concrete boundary.

The upstream workflow starts with intent and research, then drafts the instructions and tests them on representative prompts. Where the environment supports it, it compares runs with the skill against baseline behavior and presents outputs for human review. It also addresses the skill description, which influences when an agent chooses to load the instructions.

Evaluate the actual deliverable

The environment needs a way to execute the relevant agent tasks and inspect their outputs. Some helper scripts and advanced evaluation paths depend on the surrounding Claude environment. Copying the instruction folder into another client does not guarantee that every evaluation helper will work unchanged.

Choose assertions that matter to the result: whether a report cites the right ticket, preserves uncertainty, or follows the requested structure. Avoid measuring only cosmetic features that mirror the instructions. Review failures and expand the examples when a revision fixes one case but breaks another.

Use Skill Creator for instructions that should be reused. Use MCP Builder when the missing capability is an external-service tool connection instead. This page describes the upstream authoring process; it does not claim that a newly written skill has been benchmarked or that any specific model quality improvement was measured.

FAQ

Can it improve a skill that already exists?
Yes. The source explicitly supports entering the evaluation and iteration loop with an existing draft.
Does it require the same evaluation process in every environment?
No. The upstream guidance includes environment-specific adaptations. Available model execution and review tools affect which evaluation steps can run.