Skill Creator
Create, evaluate, and iteratively improve Agent Skills.
The document explicitly prohibits malware, exploit code, unauthorized access, and data exfiltration, and emphasizes avoiding user surprise. However, it does not systematically define permission boundaries, sensitive-data handling, data flows, user confirmation, dependency security, or rollback, so points are deducted.
The workflow, directory layout, commands, grading rules, and feedback schemas are fairly concrete, and an aggregation script is provided. However, execution depends on subagents, browsers, MCPs, supplied-but-not-verified references, and environmental assumptions; failure diagnosis is incomplete. Static calibration limits this dimension to 10 or below.
The name, trigger description, target audience, and create/improve/evaluate scenarios are clear, and the skill advises adapting terminology to user familiarity. Non-fit boundaries, trigger exclusions, Chinese-language support, and mainland-China environment compatibility are not specified, so points are deducted.
The skill provides metadata, progressive disclosure guidance, resource organization, examples, JSON schemas, iteration history, and script usage. The prompt supplies Apache-2.0 licensing and official Anthropic provenance, but the selected file lacks explicit versioning, changelog, installation guidance, and a clearly documented maintenance/update path, so it is not full marks.
The skill covers a complete intended workflow from intent capture and drafting through testing, benchmarking, review, and iteration, with reusable templates and scripts. However, the files contain no real execution results or third-party validation; under static calibration this dimension cannot exceed 7, and actual output still depends on environment and user feedback.
The SKILL.md, schemas, scripts, and agent specifications are auditable primary material, and some formats are statically checkable. There is no committed test suite, CI coverage, or third-party execution evidence, so static calibration limits this dimension to 5 or below.
- This is a static source review only; no scripts, subagents, evaluations, or browser viewer were executed, limiting reliability, effectiveness, and evidence scores.
- The workflow depends on external subagents, MCPs, a browser, and several referenced files not fully included in the supplied material; availability, permissions, and recovery should be verified before deployment.
- Although the documentation permits running tests and scripts, it does not define least-privilege access, sensitive-data redaction, external data-flow disclosure, or rollback confirmation procedures.
- Chinese-language interaction and compatibility of the core workflow with mainland-China network conditions are not explicitly documented.
What it does & when to use it
Skill Creator is for people building or refining Agent Skills. It covers intent capture, SKILL.md drafting, test-prompt design, skill-versus-baseline runs, quantitative evaluation, and iterative revision. It also includes a workflow for improving a skill description’s triggering accuracy. Much of the process is built around Claude Code, command-line scripts, and local files.
It helps define a skill’s purpose, trigger conditions, output format, and success criteria; drafts SKILL.md files with YAML frontmatter and Markdown instructions; creates test prompts in evals/evals.json; runs with-skill and baseline evaluations; stores outputs, timing, assertion grades, and benchmark data; aggregates results with aggregate_benchmark; presents qualitative and quantitative results through eval-viewer/generate_review.py; revises the skill from feedback; and can run a description-optimization loop.
- A developer wants to create an Agent Skill from scratch and needs help defining triggers, outputs, and test cases.
- A skill author has an SKILL.md draft and wants to compare its results with a baseline.
- A team needs quantitative evaluation using pass rates, timing, token counts, and variance.
- A maintainer wants to revise skill instructions repeatedly based on benchmark results and user feedback.
- An author wants to optimize the description field after completing the skill to improve triggering accuracy.
Pros & cons
- Covers the full skill-development loop from intent capture through testing, evaluation, and iteration.
- Combines qualitative review with quantitative benchmarks covering pass rate, time, tokens, and variance.
- Includes advanced workflows for expanding test sets, optimizing trigger descriptions, and optional blind comparisons.
- Documents SKILL.md anatomy, progressive disclosure, and bundled-resource organization.
- The workflow depends on Claude Code-related execution, command-line scripts, and local filesystem access, so portability may require adaptation.
- The sources do not document standalone installation or fixed versions for all dependencies.
- The evaluation loop can be substantial; simple or highly subjective skills may not benefit from the full process.
- The sources provide no standalone performance results or guarantees for this skill.
How to install
This skill is part of the internal plugin collection in anthropics/claude-plugins-official. The README documents collection installation with: /plugin install {plugin-name}@claude-plugins-official. In Claude Code, it can also be found through /plugin > Discover. A standalone installation procedure for this skill is not documented in the supplied sources.
How to use
After installing it in a compatible environment, give it a concrete request such as: “Help me create a skill for cleaning CSV files, design test prompts, and evaluate its performance.” For an existing skill, ask: “Here is my SKILL.md; run evaluations and improve it based on the results.” The documented workflow may create test prompts, output directories, benchmark files, and a review page.