Discovery Ranking

For each capability, generate:

• direct query: domain or object + required action + Skill;
• operation query: input + transformation + output;
• supporting query: desired quality or reduced risk + operation;
• integration query: relevant system + CLI, MCP, API, connector, or browser automation;
• Chinese and English forms;
• synonyms, abbreviations, and desired artifact names;
• a GitHub query targeting SKILL.md when supported.

Do not reuse queries from a different domain. Humanizer-style helpers are found through queries about natural writing, tone, rewriting, or style quality rather than the main domain name.

Search order

1. Current project profile and local Skill index.
2. Installed and archived Skills not yet indexed.
3. Registries such as skills.sh and agentskill.sh.
4. GitHub repository and file search.
5. OpenCLI or general web search for broader recall.
6. Platform-specific search only when current platform evidence is needed.

Search snippets discover candidates; they do not verify them. Verify from the canonical repository.

Run browser-backed OpenCLI searches sequentially. Retry one rejected navigation with an explicit profile and trace, then fall back to another read-only source.

Match by capability

Compare each candidate against:

• input compatibility;
• operation performed;
• expected output;
• domain constraints;
• read/write boundary;
• environment and dependencies;
• evidence from full instructions and scripts.

Do not rank on title similarity alone.

Internal trust record

Keep this record internally. In plain-language mode, translate it to 安全检查通过, 安全试跑通过, and 最近确认可用.

Hard gates

Do not recommend installation while any of these remains unresolved:

• source or exact version cannot be identified;
• no readable installable Skill structure exists;
• full contents do not support the claimed capability;
• mandatory runtime, tool, account, or operating system is incompatible;
• critical security behavior is unexplained;
• the only test would mutate a real external system;
• intended use creates material license or terms uncertainty.

Weighted score

Score only after the hard gates.

Break the 25 community points down as guidance:

• installation or adoption count: up to 10;
• repository stars/forks relative to age and niche: up to 5;
• ratings and written feedback with enough volume: up to 5;
• independent examples, integrations, or repeated use: up to 5.

Avoid double-counting a monorepo’s stars for every small Skill. Normalize numbers by source, age, and niche where possible. A recently released niche Skill may be labeled promising rather than treated as bad, but it should not outrank a similarly fitting and well-proven alternative without evidence.

Use confidence labels:

• Confirmed: hard gates pass, full check complete, safe trial passes.
• Promising: fit looks good but trial or dependency verification is incomplete.
• Unconfirmed: insufficient evidence; exclude from one-click installation.
• Blocked: hard gate failed or critical risk remains.

Minimal stack selection

Choose the fewest non-overlapping Skills that cover all required success conditions.

Prefer:

1. an existing confirmed local Skill;
2. one well-routed Skill covering several necessary capabilities;
3. a primary Skill plus narrowly scoped helpers;
4. a new installation only for a real gap.

Do not recommend two primary Skills for the same step unless the user wants alternatives.