AGENT DEVELOPMENT / Anthropic
Anthropic Skill Creator
Develops skills from realistic task fixtures and baseline comparisons, with separate trigger-description experiments and human output review.
“Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.”
01 / THE REASONING
Why this made the selection.
- Separates activation experiments from task-output quality, preserving both a baseline and human feedback.
- Provides concrete task, grading and benchmark schemas and an inspectable trigger evaluator rather than treating a well-formed SKILL.md as proof of usefulness.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- skills/skill-creator/SKILL.md
- skills/skill-creator/references/schemas.md
- skills/skill-creator/scripts/quick_validate.py
- skills/skill-creator/scripts/run_eval.py
- skills/skill-creator/scripts/run_loop.py
- skills/skill-creator/scripts/improve_description.py
- skills/skill-creator/scripts/utils.py
- skills/skill-creator/LICENSE.txt
- README.md and pinned Skills installer README source-format sections
Our findings
- Task inputs are a skill directory, realistic prompts, expected outputs, example files and assertions; trigger inputs are query/should_trigger pairs. Task handoffs include iteration outputs, grading.json expectations with text/passed/evidence, timing.json, benchmark.json and human feedback. Trigger output contains trigger_rate, triggers, runs and pass per query.
- The evaluator creates a uniquely named .claude/commands Markdown stub near the current project, invokes claude -p, kills it after detecting the first relevant event or timing out, and removes the stub in finally. This is not an isolated trial workspace.
- The optimizer splits positives/negatives, blinds test fields from description-improvement prompts, and chooses best_description by test_passed. improve_description.py sends the full skill text and failed training queries through Claude CLI and optionally logs those prompts/responses.
- The same pin remains necessary for the catalogue's existing root-discovery receipt; neither that receipt nor install command was changed or upgraded.
Not established by this review
- No Claude subprocess, subagent, metadata validator, viewer, installer or evaluation was run.
- No runtime schema compatibility, trigger accuracy, skill effectiveness, packaging safety or dependency-advisory scan. PyYAML, Claude CLI and the Skills installer implementation are not exhaustively audited.
- Not executed: it was not installed or run.
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Add a skill to your workflow.
DOCUMENTED COMMAND
npx skills add anthropics/skills --skill skill-creator Run in your terminal to choose your AI tools, project or global scope, and installation method. This short command follows the current upstream version; the review covers the revision shown above.
Optional: install the reviewed revision
Keeps the original source and installer version where recorded, while leaving your AI tools and installation scope for you to choose.
npx skills add https://github.com/anthropics/skills/tree/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator - Install the complete skill directory with its references and helpers; a copied SKILL.md alone may be incomplete.
- Choose your AI tools and project or global scope in the installer, then review its permissions before using the skill.
Before you start
- Node.js/npm for this installer
- An agent that supports the Agent Skills format
- Python/PyYAML for metadata validation
- Claude CLI and supported subagents for full evaluation
04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/SKILL.md
Supports: summary, upstreamDescription, whySelected, bestFor, review
- Task and review schemas ↗Checked
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/references/schemas.md
Supports: review, limitations
- Trigger evaluator ↗Checked
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/scripts/run_eval.py
Supports: review, limitations, access
- Description selection loop ↗Checked
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/scripts/run_loop.py
Supports: review, limitations
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/scripts/improve_description.py
Supports: review, access
- Metadata checker ↗Checked
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/scripts/quick_validate.py
Supports: limitations, review
- Skill licence ↗Checked
https://github.com/anthropics/skills/blob/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/LICENSE.txt
Supports: license
- Installer source formats ↗Checked
https://github.com/vercel-labs/skills/blob/7407f3893ad4dceab546ac002c3ef806e4000c73/README.md
Supports: install
https://raw.githubusercontent.com/anthropics/skills/34040c9c568585f6929bedeaad110ad08f079624/skills/skill-creator/SKILL.md
Supports: upstreamDescription, review.notTested, review.reviewedAt