Skill Creator

Skill Creator is the skill Anthropic points to when you want to make your own. You describe a task you keep explaining to Claude, and it walks you through turning that into a skill folder, testing it on real prompts and improving it until it actually helps. Anthropic’s own “Create custom skills” guide recommends it for measuring whether a skill improves Claude’s output.

What it does (plain English)

A skill is a folder with a SKILL.md file: a short name and description at the top, then instructions. Optional scripts/, references/ and assets/ folders hold code, longer documents and templates. Skill Creator teaches Claude this layout and runs a loop:

1. Capture intent. Claude asks what the skill should do and when it should trigger.
2. Draft the SKILL.md. It follows rules like keeping the main file under about 500 lines and moving long material into reference files Claude opens only when needed.
3. Test. Claude writes a few realistic prompts and runs them with the skill. In Claude Code, which has subagents, it also runs the same prompts without the skill so you can compare.
4. Review. A bundled viewer shows outputs side by side so you can comment.
5. Improve and repeat.
6. Optimize the description. Claude writes about 20 test queries, a mix of ones that should and should not trigger the skill, and tunes the description, which is the only part Claude sees before deciding to load a skill.
7. Package. A script produces a file you can upload to claude.ai.

The skill is written for people of mixed technical background. It explicitly tells Claude to read cues about how familiar you are with terms like “JSON” before using them unexplained, and to drop formal evaluation if you just want to “vibe.”

Is it for you?

Yes if you have a process Claude should follow the same way every time: a weekly report format, your brand rules, a checklist for reviewing contracts, how your team names files. It is also the right tool when a skill you already have is not being picked up.

You do not need it for a one-off instruction. A project’s instructions or a saved prompt is simpler.

What you can do with it

1. Start from scratch. “Help me make a skill that writes our monthly investor update in our format. Interview me first.”
2. Improve an existing skill. “Here is my skill folder. Run three test prompts with and without it and show me the difference.”
3. Fix triggering. “My skill never loads when I ask for a status report. Optimize its description.”
4. Package for claude.ai. “Package this skill so I can upload it.” The output is a file you upload under Customize > Skills.

How to set it up

claude.ai or the Claude desktop app

1. In Settings > Capabilities, turn on Code execution and file creation.
2. Go to Customize > Skills and turn on skill-creator. Anthropic’s how-to page says it is listed there as from Anthropic.
3. Ask Claude to help you build or evaluate a skill.

Claude Code

1. Run /plugin install skill-creator@claude-plugins-official. This is the command Claude Code’s skills documentation gives.
2. Pick a scope (just you, the project, or this repo only).
3. To try a skill you made without uploading it, copy its folder to ~/.claude/skills/<skill-name>/ and start claude.

What changes by app

The skill adjusts itself to where it runs. This table comes from its own platform notes:

Step	Claude Code	claude.ai
Test runs	In parallel, with a no-skill baseline	One at a time, no baseline
Review	Browser viewer	Results shown in the chat
Benchmarks	Yes	Skipped
Description optimization	Yes (uses claude -p)	Skipped
Packaging	Yes	Yes, you download the file

So the full measure-and-tune loop needs Claude Code. On claude.ai you still get drafting, manual testing and packaging.

Watch out for

• Context and time. Claude Code estimates the skill at about 11k tokens when it loads, and a full test round runs several prompts. Expect a longer, more usage-heavy session than a normal chat.
• Upload format. For claude.ai, the ZIP must contain the skill folder itself at the top level, and the folder name must match the name field. A SKILL.md loose at the root of the ZIP is not recognized.
• No secrets in skills. Anthropic’s guide warns against putting API keys or passwords in a skill, because everyone you share it with gets its files.
• Reserved names. In Claude Code, don’t name a skill folder synced or anthropic-skills.
• Uploaded skills are not reviewed. Anthropic’s skills page says skills you upload or receive from others are not reviewed by Anthropic. Read any skill before you turn it on.

Alternatives

• Writing SKILL.md by hand, following Anthropic’s “Create custom skills” guide, if your skill is a few paragraphs.
• MCP builder when what you need is a new connection to an outside service, not new instructions.
• Knowledge work plugins if a ready-made role plugin already covers your job.