SkillsMP Research by AIW

Need to explore a topic or build the right skill stack for a task? This skill is for you.

Marketplace metadata can find candidates. SkillsMP Research reads the source, removes duplicates, compares the methods, and tells you what is actually useful.

SkillsMP Research is an independent project that uses SkillsMP; it is not affiliated with or endorsed by SkillsMP.

What it does

• Turns your need into short, focused SkillsMP queries.
• Reads each finalist's complete SKILL.md instead of trusting the listing.
• Looks for a useful method delta, not a polished description of familiar advice.
• Groups duplicates, forks, and stale sources before scoring.
• Explains why the winner fits, where it falls short, and when to use an alternative.
• Never runs code from candidate repositories during evaluation.

Install

The easiest route works across the Agent Skills ecosystem:

npx skills add FixAdmin/skillsmp-search

The installer detects supported agents and asks where to place the skill. You can target one explicitly:

npx skills add FixAdmin/skillsmp-search --agent codex
npx skills add FixAdmin/skillsmp-search --agent claude-code
npx skills add FixAdmin/skillsmp-search --agent cursor

Project installation is the default. Add --global only when you deliberately want the skill in every project.

The shared SKILL.md format is portable. The search runtime requires Node.js 18 or newer and network access to SkillsMP and GitHub.

Try it

Ask your agent naturally:

Find me skills for writing prompts and developer documentation.

Or ask for the deeper pass explicitly:

Use heavy mode to find skills for evaluating production AI agents.

Heavy mode never activates just because a topic looks difficult. You have to ask for heavy, deep search, глубокий поиск, or equivalent maximum-depth wording.

Heavy searches keep a local checkpoint as they run. If the agent's context is compacted or the process is interrupted, it can continue from saved API responses and source capsules instead of starting over.

Standard and heavy mode

	Standard	Heavy
Trigger	Normal skill-search request	Explicit request only
Queries	3-5	10-12
Sort passes	stars	stars + recent
API requests	3-5	20-24
Candidate pool	Up to 40	Up to 250 after merging
Fully inspected finalists	6-8	20-25
Recommendations	3-5	5-8

See a standard search example and a heavy search example.

How recommendations earn their place

Each inspected finalist receives a content score out of 100:

Signal	Weight	Question
Task fit	30	Does it solve the requested problem in this environment?
Useful method delta	30	Does it change decisions or execution beyond baseline model knowledge?
Current AI-native alignment	20	Does it use practical, maintained agent methods?
Actionability	10	Can an agent follow the workflow with available tools?
Validation	10	Does it include checks, feedback, or recovery?

A polished skill with no meaningful method delta cannot score above 60.

Run the search script directly

Most people should invoke the skill through their agent. For debugging or integration, the bundled CLI prints JSON:

node skills/skillsmp-search/scripts/search-skillsmp.mjs \
  --query '"prompt engineering"' \
  --query 'developer documentation' \
  --limit-per-query 20 \
  --max-candidates 40

Windows users can keep the familiar PowerShell entry point. It calls the same Node implementation:

& skills/skillsmp-search/scripts/search-skillsmp.ps1 `
  -Query @('"prompt engineering"', 'developer documentation') `
  -LimitPerQuery 20 `
  -MaxCandidates 40

Set SKILLSMP_API_KEY in your environment for authenticated limits. The script never writes or prints the key.

Inspect a shortlist in one pass

Put the finalists you selected from the search output into finalists.json:

{
  "candidates": [
    {
      "name": "Example skill",
      "githubUrl": "https://github.com/owner/repository/tree/main/skills/example"
    }
  ]
}

Then fetch and inspect every selected SKILL.md together:

node skills/skillsmp-search/scripts/inspect-skillsmp.mjs \
  --input finalists.json \
  --output inspection.json \
  --run-id my-search \
  --resume

On PowerShell, the equivalent command is:

& skills/skillsmp-search/scripts/inspect-skillsmp.ps1 `
  -InputPath finalists.json `
  -OutputPath inspection.json `
  -RunId my-search `
  -Resume

The first run saves immutable source bytes, extracted facts, compact evidence capsules, one checkpoint per candidate, and inspection.review-index.json. An agent reads the smaller review index first and opens full capsules only for plausible or uncertain candidates. Repeating the command with the same run ID resumes from those files and does not download completed sources again.

Heavy-mode automation uses the same Node runtime for its search state, candidate index, and progress checkpoint. The complete agent-facing workflow lives in references/heavy-mode.md.

Cache objects and active checkpoints are durable. Rendered inspection and review-index files can be regenerated after completion. The tool never deletes them automatically.

The run is complete only when:

successful capsules + terminal failures = canonical selected candidates

Candidate repositories are treated as untrusted input. The inspector reads SKILL.md as text and never executes its commands, scripts, or installers.

API budget

SkillsMP currently documents these limits:

• Anonymous: 10 requests per minute and 50 per day.
• Authenticated: 30 requests per minute and 500 per day.

One maximum heavy search uses 24 requests. That is 4.8% of the authenticated daily allowance.

Compatibility

The repository follows the Agent Skills specification. The skills CLI can place it for Codex, Claude Code, Cursor, GitHub Copilot, OpenCode, and many other supported agents.

Basic skill instructions transfer well. Agent-specific tools do not. This project therefore promises a portable format and a cross-platform Node runtime, not identical behavior in every agent.

The release checks Node on Windows, macOS, and Linux. The PowerShell wrapper is an optional Windows convenience.

Contributing

Found a stale source rule, ranking edge case, or clearer way to explain a recommendation? Contributions are welcome. Start with CONTRIBUTING.md.

For security problems, please follow SECURITY.md instead of opening a public issue.

License

SkillsMP Research is available under the MIT License.