Skill Creator — A Complete Workflow for Creating, Improving, and Evaluating Claude Skills

Skill Overview

Skill Creator is a meta-skill that covers the entire lifecycle of a skill: helping you create a new Claude skill from scratch, improve an existing skill, or score a skill’s quality using a scorecard and provide a list of improvements.

Applicable Scenarios

1. Creating a new skill from scratch: When you have a process that you repeatedly perform manually and want to turn it into a reusable skill, Skill Creator first helps you clarify the requirements (what the task is, who will use it, which tools it depends on, and what the inputs and outputs are), then plans the architecture pattern (linear, routing, methodology, Widget, or API wrapper), and finally produces the SKILL.md and reference files.

2. Improving an existing skill with unsatisfactory results: When a skill is not triggering properly, its steps are too vague, or users report that “it doesn’t work right,” it reads the existing skill content, scores it across 10 dimensions, lists specific modifications in order of impact, and then applies the changes.

3. Evaluating and comparing skill quality: When you need to determine whether a skill meets publishable standards or choose among several versions, it outputs a complete scorecard, compares the skill against a benchmark skill, and provides a prioritized list of improvements.

Core Features

1. Automatic routing across three modes: It automatically enters Create, Improve, or Evaluate mode based on your intent, so you do not need to decide which path to take yourself. If your intent is unclear, it directly asks which one you want.

2. 10-dimensional quality scorecard: It scores the skill across ten dimensions: trigger quality, default-value coverage, step architecture, reference-file strategy, dynamic content, output templates, error handling, code and formula quality, SKILL.md conciseness, and domain accuracy. It outputs the total score and rating, along with a comparison against a benchmark skill.

3. Dynamic invocation design and detection workflow: This is the skill’s core rule: skills must be “dynamic” rather than “static.” It requires every skill involving external tools to begin with runtime checks (whether the CLI is installed, whether the user is logged in, which runtime has the relevant library, and whether real-time data is accessible), and then produce at least two method paths and fallback options based on the results, rather than hard-coding a single tool.

4. Layered writing for SKILL.md and reference files: Workflows, decision tables, default values, and output templates remain in SKILL.md (kept under 250 lines), while API documentation, code templates, formulas and edge cases, and troubleshooting tables are moved into the references/ directory for on-demand loading.

5. Pre-delivery quality self-check: It includes a 16-item checklist covering required frontmatter fields, the number of triggers, parameter defaults, step numbering and exit conditions, reference-file pointer formatting, and more. If any item fails, it fixes the issue before delivery.

Frequently Asked Questions

What does Skill Creator do?

It is a skill for “building skills,” covering the entire process from skill design and writing to self-checking and iterative evaluation. You can think of it as both scaffolding and a review system for skill authors: it provides structures and templates during creation, diagnostics and patches during improvement, and a scorecard and improvement priorities during evaluation.

How many lines should a qualified skill contain?

SKILL.md is recommended to stay under 250 lines, with a hard limit of approximately 300 lines. The recommended number of steps is 5–9; more than 9 steps indicates that the skill should be split up or use a routing pattern. The amount of content is not the standard—deferring complexity to reference files while keeping the main file precise and readable is the recommended approach.

How many triggers should be included in a skill description?

At least five different trigger phrases, including “indirect entry points”—that is, expressions where users do not directly mention the skill name but have the same intent. For example, in addition to “create a skill,” it should cover natural expressions such as “I keep doing X manually” and “turn this into a skill.”

Which dimensions does the skill scorecard assess?

There are 10 dimensions in total: trigger quality, default-value coverage, step architecture, reference-file strategy, dynamic content, output templates, error handling, code and formula quality, SKILL.md conciseness, and domain accuracy. Each dimension is scored from 1 to 10, and the scores are aggregated into a total score and quality rating.

Can it improve a skill I have already written?

Yes. In Improve mode, it first reads the existing SKILL.md and all reference files, provides dimension-by-dimension scores and issue explanations based on the scorecard, and then lists modification suggestions in descending order of impact. It only makes changes after you confirm, using either targeted patches or a complete rewrite.

Does using this skill require an internet connection or additional tools?

No. Creating and improving skills are purely writing and structural-design tasks and do not depend on external APIs or keys. Only when the skill you are creating itself needs to call external tools will Skill Creator require you to design runtime checks and fallback paths within that skill, rather than hard-coding a single invocation method.