Directory / skill-creator (example-skills)
skill-creator (example-skills)
StrongBuilds, evaluates, and iteratively improves Claude Code skills using an eval-loop: write → test → grade → rewrite → repeat.
Verdict
Skill-creator closes the loop on skill development: write a draft, run eval queries through a fresh Claude subprocess, grade pass/fail with a separate grader agent, auto-improve the description, repeat. The eval infrastructure is the real value — most skill authors never build this, and the gap shows in skills that trigger randomly. The only flag is a Google Fonts load in locally-rendered eval-viewer HTML; it is cosmetic and easy to block at the network level. The SKILL.md's explicit prohibition on building exfiltration or access-bypass skills is a small but meaningful trust signal.
Install
Source: github.com/anthropics/skills
Audit receipts
Every listing is verified against a specific commit. These are the receipts.
Ships as part of the example-skills bundle. run_eval.py and run_loop.py invoke `claude -p` via subprocess (list-form, no shell=True) to test skill triggering. eval-viewer spins a local HTTP server on 127.0.0.1 only.
generate_report.py embeds Google Fonts via <link rel="preconnect" href="https://fonts.googleapis.com"> in locally-generated HTML reports — a cosmetic font load that pings Google each time a user opens the eval-viewer. Not a security threat but an undisclosed third-party network contact. The SKILL.md explicitly prohibits creating skills for "unauthorized access, data exfiltration" — a positive safety signal.