skill-creator review: verdict on Anthropic's Claude skill creator

skill-creator is Anthropic's meta-skill that interviews you, drafts a new Agent Skill, and tests it against no-skill runs. We rate it Conditional, 57/100, at Moderate confidence, because its description optimiser reports false 0% recall.

Get the skill: skill-creator folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing

What is Anthropic's skill-creator?

Anthropic's skill-creator is a meta-skill for building Agent Skills. It interviews you and drafts a SKILL.md. It runs test prompts with and without the skill and shows results side by side.

An Agent Skill is a folder of instructions an agent loads on demand. Our guide explains what an Agent Skill is. The tool writes the main file, described in how SKILL.md files are structured.

It is two tools in one folder. The instruction file for authoring and review holds up. The Python scripts that tune the trigger description break at the reviewed commit.

The grader scores outputs against assertions. The comparator runs blind A/B checks, and the analyzer explains the winner.

Is skill-creator the same as SkillCreator.ai or other skill builders?

No, Anthropic's skill-creator is not SkillCreator.ai. SkillCreator.ai is an independent product that generates skills across several agents. Anthropic's skill-creator is a free, Apache 2.0 skill that runs inside Claude Code, Claude.ai and Cowork.

Three tools share the name.

Is skill-creator worth using?

Yes, skill-creator is worth using for drafting and reviewing a new skill. It scores 57/100, Conditional, at Moderate confidence. Its authoring loop works, but its description optimiser reports false 0% recall and can drain a plan's quota.

It fits 3 groups poorly:

• Anyone relying on run_loop.py or run_eval.py trigger numbers without checking them by hand.
• Windows users running the bundled Python scripts, which fail on pipe reads and UTF-8 text.
• Subscription users on Pro or tight usage limits, given the parallel claude -p fan-out.

See how caps and scores work for the scoring rules.

Compare it with Agent Skills ranked by verdict.

How many tokens and how much quota does skill-creator use?

skill-creator loads 8,292 tokens, our measured file size per load. Drafting one skill with three paired tests costs about 150,000 input tokens by our estimate. One optimiser run used a full Max plan in about three minutes.

Our typical task is one new skill with 3 paired test cases and one review round, once a week. Each test case spawns two subagents plus graders.

The description optimiser runs up to 5 iterations of 20 queries, 3 runs each. Each run is a separate claude -p session with your full skill and MCP context. The default is 10 parallel workers. Issue #1412 reports 121 calls that used a full Max plan in about 3 minutes. Issue #975 reports two Pro 5-hour limits drained with no result.

Is skill-creator safe to run?

Yes, skill-creator is safe to run with care; it carries safety grade C. Its eval scripts spawn claude -p sessions with your permissions. They write probe files into your project and serve a review page with incomplete escaping.

Eight of the nine things we found in the files bear on safety:

• Eval probes run as full agents and can act on imperative test queries (#1061).
• Probe command files land in .claude/commands and show up in other sessions (#1260).
• The review page renders grader data through innerHTML; 14 test payloads execute (#1394).
• generate_report.py loads Google Fonts from the web.
• skill_name from frontmatter goes into file paths without validation (#1631).
• quick_validate.py imports PyYAML, which no file declares (#1467).
• No telemetry or credential handling turned up.
• skills.sh audits show Gen Agent Trust Hub Pass, Socket Warn and Snyk Pass.

Network access comes from the local review server and the font request. Run the eval scripts in a throwaway project. Our guide on checking what a skill's scripts do explains the steps.

What does the community report about skill-creator?

Community reports about skill-creator describe a useful authoring loop and a broken optimiser. One user measured an 84% to 100% pass-rate lift. Several issues report 0% recall in trigger evals and whole plan quotas spent.

Our evidence holds 18 dated sources: 15 negative, 1 neutral and 2 positive.

Sogl's post is the only measured result from the six months before our review. Debbie O'Brien reported 81 to 97.5 on dev.to, better on 2 of 4 evals. Her post is from March, older than that, so the verdict does not rest on it.

What are skill-creator's known failures?

The known failures of skill-creator at commit 8a1541c are five main ones. Trigger evals report 0% recall, and crashed runs count as passes. Benchmarks come back empty, Windows scripts fail, and the optimiser drains usage quotas.

The worst is silent: the optimiser measures noise as 0% recall and rewrites a good description.

• Zero recall: the trigger detector stops at the first tool other than Skill or Read (#1721, #1352, #1169, #1412).
• Crashes as passes: crashed runs count as non-triggers, so negative queries pass (#1478).
• Empty benchmarks: the documented layout yields exit 0 and an inverted delta (#1383). Tokens are counted from characters (#1826).
• Windows breakage: scripts fail on pipe reads, PATHEXT and cp1252 text (#1061, #1692, #1827).
• Quota drain: one optimiser run used a Max plan; a Pro user lost two 5-hour limits (#1412, #975).
• Truncation: the view tool cuts SKILL.md at about 16K characters (#906).
• Ignored field: the harness ignores when_to_use (plugins-official #6327).

Fix PRs #1298 and #1755 are still open. Our page on fixing a skill that does not trigger covers manual trigger checks.

Which agents does skill-creator work in?

skill-creator is verified to work in one agent, Claude Code on macOS and Linux. Anthropic claims Claude.ai and Cowork, and one user ran it in VS Code. Its scripts fail on Windows, and ChatGPT support is unknown.

The main eval path assumes subagents (#1942), and the fallback is described only for Claude.ai. Compare the best skills for Claude Code.

How do you install skill-creator?

Install skill-creator in Claude Code with /plugin install skill-creator@claude-plugins-official. You can use npx skills add or copy the folder instead. In Claude.ai, Anthropic's example skills are switched on in settings.

As a Claude Code plugin

Yes, a skill-creator plugin for Claude exists. Run this in Claude Code:

/plugin install skill-creator@claude-plugins-official

Update it later with claude plugin update skill-creator@claude-plugins-official. The plugin copy, synced on 2026-04-23, matches the reviewed SKILL.md byte for byte. Our guide to Claude Code plugin and skill setup covers plugins.

With skills.sh or by hand

Run npx skills add https://github.com/anthropics/skills --skill skill-creator. Or copy the folder to ~/.claude/skills/skill-creator/. The scripts need Python 3, the claude CLI and PyYAML. On skills.sh it had 395.0K installs at our review and 410.3K on 2026-10-20.

In Claude.ai and Cowork

Anthropic's official sources disagree on the path. The Academy tutorial uses Settings > Capabilities > Skills. The Help Center lists Customize > Skills for toggling skills. Plan access conflicts too: the Help Center says all plans, and the anthropics/skills README says paid plans. Our walkthrough shows how to enable skills in Claude.ai.

How do you use skill-creator to build and test a skill?

Use skill-creator by typing /skill-creator or asking Claude to make a skill. Answer the interview, review the draft, then compare test runs with and without the skill. Treat its trigger scores as untrusted until checked by hand.

1. Type /skill-creator, or ask Claude to make a skill.
2. Answer the interview about the task, inputs and outputs.
3. Review the drafted SKILL.md before any test runs.
4. Run each test prompt with and without the skill.
5. Grade the outputs against your assertions.
6. Review the paired outputs on the local review page.
7. Edit the skill and repeat until the outputs hold up.

Drafting and baseline benchmarks are the primary uses. Description tuning and .skill packaging are secondary, and tuning is unreliable. Blind A/B comparison is claimed by Anthropic only.

Test it yourself before you trust the trigger harness:

• Run run_eval.py with --num-workers 1.
• Use one obvious query that ought to trigger an installed skill.
• Check whether it reports a trigger.

What changed in skill-creator v2?

skill-creator v2 is the user name for Anthropic's March 2026 update. It added evals, benchmark mode, parallel agents, comparator agents and description optimisation. Anthropic's post never uses the label 2.0.

Anthropic announced the update on 2026-03-03 in "Improving skill-creator: Test, measure, and refine Agent Skills". Anthropic reported improved triggering on 5 of 6 built-in document skills, a publisher claim. The description optimiser is the broken part at commit 8a1541c.

What are the alternatives to skill-creator?

The closest alternatives to skill-creator are a hand-written SKILL.md, Obra's Superpowers writing-skills and Skill Seekers. Skill Seekers, from yusufkaraaslan, scrapes docs, repos and PDFs into SKILL.md packages. None of them fixes trigger testing.

Writing by hand costs no extra tokens and skips the broken scripts. See our tutorial to write a skill by hand.

Where is skill-creator on GitHub?

skill-creator is on GitHub in the anthropics/skills repository at skills/skill-creator. It holds 18 files, including ten Python scripts, under Apache 2.0. Its content last changed on 2026-03-06, and many fix PRs wait unmerged.

• Repository: anthropics/skills, folder skills/skill-creator.
• Reviewed commit: 8a1541c, dated 2026-09-28.
• Last folder change: b9e19e6 on 2026-04-20, a licence notice.
• Last content change: b0cbd3d on 2026-03-06, dropping the ANTHROPIC_API_KEY requirement.
• Copies: 13,312 exact copies in GitHub code search.
• Contributors: 15 to the repository.
• Open fixes: over 60 fix PRs target the eval harness, none merged.

Common questions

Does ChatGPT or Codex have a skill creator?

Yes for Codex: OpenAI ships its own skill-creator, which we have not evaluated. ChatGPT support for Anthropic's skill-creator is unknown.

Does skill-creator work in opencode?

Unknown; we found no evidence.

Is there an official guide to building skills?

Yes, Anthropic publishes a guide to building skills.

Should I trust skill-creator's trigger scores?

No. Check triggers by hand until PR #1755 or #1298 merges.