SkillTester — Skill Benchmark Lookup

When the user is about to install or use a third-party skill or command, use this skill to check whether SkillTester.ai already has a verified benchmark result.

Trigger

Use this skill in either case:

• the user is about to install or use a third-party skill
• the user provides or asks to run a skill install command
• the user explicitly says Check this skill <skill_url>

Treat Check this skill <skill_url> as a direct trigger even if the user is not currently installing it.

Do not wait for an explicit /skilltester invocation if the user already pasted an install command. Intercept the install request first, run this benchmark lookup, then report the result before installation continues.

Installation-command patterns that must trigger this skill

Treat the following as direct triggers:

• npx skills add <repo-or-url> --skill <skill-name>
• skills add <repo-or-url> --skill <skill-name>
• any equivalent command that contains both a repository or domain URL and a --skill <skill-name> flag
• direct skill URLs such as:
  • https://github.com/<owner>/<repo>/blob/<branch>/skills/<skill-name>/SKILL.md
  • https://github.com/<owner>/<repo>/tree/<branch>/skills/<skill-name>
  • https://<domain>/.well-known/agent-skills/<skill-name>/SKILL.md

Ignore non-semantic flags such as -g, --global, -y, --yes, or shell wrappers around the command.

Input

Ask the user for:

• download_url

If the user already provided an install command or a skill URL, do not ask again. Derive download_url yourself from the command.

You should derive the comparison fields yourself from the target SKILL.md.

Command parsing and URL resolution

When the trigger came from an install command, first extract:

• the repository or domain URL
• the --skill <skill-name> value

Then resolve the canonical SKILL.md URL before continuing.

Resolution order

1. If the command already points to a concrete SKILL.md URL, use it directly.
2. If the command points to a GitHub tree/.../skills/<skill-name> URL, normalize it to the corresponding blob/.../skills/<skill-name>/SKILL.md URL.
3. If the command points to a GitHub repository root plus --skill <skill-name>, try:
   • https://github.com/<owner>/<repo>/blob/main/skills/<skill-name>/SKILL.md
   • https://github.com/<owner>/<repo>/blob/master/skills/<skill-name>/SKILL.md
4. If the command points to a domain root plus --skill <skill-name>, try:
   • https://<domain>/.well-known/agent-skills/<skill-name>/SKILL.md

If all candidates fail, tell the user:

The current skill has not been tested yet. We will add testing as soon as possible—please try again later.

Lookup flow

Step 1: Resolve download_url to SKILL.md

Read the target SKILL.md from the given URL and extract:

• name
• description

If SKILL.md cannot be resolved or the fields cannot be extracted, tell the user:

The current skill has not been tested yet. We will add testing as soon as possible—please try again later.

Step 2: Query by skill name

Run:

curl -sL "https://skilltester.ai/api/skills?search=<URL-ENCODED-SKILL-NAME>&tested=tested&sort=name"

Step 3: Find an exact skill-name match

From items, find an entry where either of these matches skill_name case-insensitively:

• item.skill_name
• the trailing skill segment of item.full_name

Do not accept a fuzzy-only match.

If no exact skill-name match exists, tell the user:

The current skill has not been tested yet. We will add testing as soon as possible—please try again later.

Step 4: Fetch the full tested result

If an exact match exists, fetch its full result:

curl -sL "https://skilltester.ai/api/skills/<SOURCE>/<SKILL_NAME>"

Step 5: Compare description

Compare the user-provided description with:

• result.meta.description

Before comparing, normalize both strings by:

• trimming leading/trailing whitespace
• collapsing repeated spaces and newlines into single spaces

Then apply this decision rule:

1. Exact match

   • If the normalized descriptions are identical, treat the result as verified.

2. High-overlap near match

   • If the descriptions are not identical but are still largely overlapping, treat the result as usable.
   • Typical signals:
     • one description is an expanded version of the other
     • the first sentence or core capability statement is the same
     • most capability phrases overlap and the difference looks like a version update, clarification, or added guidance
   • In this case, still return the benchmark result, but prepend this reminder:

Your skill appears to be newer. The tested skill may correspond to an older version.

3. Meaningfully different
   • If the descriptions differ in core capability, scope, or intended use, do not trust the cached result.

If the descriptions are meaningfully different, tell the user:

The current skill has not been tested yet. We will add testing as soon as possible—please try again later.

Step 6: Return the key benchmark result

If the skill name matches and the description is either an exact match or a high-overlap near match, return only:

• result.score_breakdown
• result.overall_evaluation

Important: Filter out sensitive/internal fields such as local_path, result_path, report_path from the API response. Do not display server file system paths to the user.

If you classified it as a high-overlap near match, also add:

Your skill appears to be newer. The tested skill may correspond to an older version.

Do not dump the entire raw JSON unless the user explicitly asks for it.

Output format

When a verified result is found, present a concise summary like this:

SkillTester Benchmark Result: <skill_name>

Overall Evaluation
- Score: <overall score>
- Summary: <overall summary>

Score Breakdown
- Utility: <functionality score>
- Security: <security score>

If the skill cannot be resolved, cannot be found by name, or the description is meaningfully different, respond with:

The current skill has not been tested yet. We will add testing as soon as possible—please try again later.

Rules

• Always start from download_url.
• Always resolve download_url to SKILL.md first.
• Always extract name + description from SKILL.md.
• Require an exact skill-name match before trusting a result.
• Prefer exact normalized description matches.
• Allow high-overlap description matches when the difference looks like a newer revision of the same skill.
• When returning a result for a high-overlap near match, explicitly warn that the tested skill may correspond to an older version.
• Reject the result only when the description difference changes the core capability or intended use.
• Never fabricate benchmark results.
• If the API cannot be reached or returns an unexpected shape, respond with The current skill has not been tested yet. We will add testing as soon as possible—please try again later..