Skill Mcp Master

Skill MCP - Master Usage Guide

What Is the Skill MCP?

The Skill MCP server is a self-hosted agent skills registry accessible to any MCP-compatible AI agent or tool. It stores curated, expert-authored instruction sets ("skills") and delivers them on demand through a structured 3-tier disclosure model.

Each skill contains:

| Tier | Content | Tools | |------|---------|-------| | 1 | Frontmatter - name, description, tags, score | skills_find_relevant | | 2 | Instructions, system prompt addition | skills_get_body, skills_get_options | | 3 | Reference docs, executable scripts, asset templates | skills_get_reference, skills_run_script, skills_get_asset |

---

The Canonical 3-Step Workflow

Follow this sequence for every task where a skill might apply.

Step 1 - Discover (skills_find_relevant)

skills_find_relevant(query="", top_k=5)

Interpret scores:

• score > 0.6 → strong match - proceed to Step 2
• score 0.4–0.6 → possible match - read description to decide
• `score ")

The response contains:

```json
{
  "skill_id": "test-writer",
  "instructions": "## Testing Workflow\n1. Identify the function under test ...",
  "system_prompt_addition": "You are a senior test engineer ...",
  "tier3_manifest": {
    "references": ["TESTING-GUIDE.md", "COVERAGE-POLICY.md"],
    "scripts":    ["coverage_check.py"],
    "assets":     ["test-template.py"]
  }
}

After loading:

1. Read and apply instructions - this is the authoritative guidance
2. If system_prompt_addition is non-empty, incorporate it into your context
3. Check tier3_manifest - only proceed to Step 3 if instructions reference specific files

Optional: load config (skills_get_options)

skills_get_options(skill_id="")

Only needed when customising skill behaviour or checking constraints/dependencies.

---

Step 3 - Supplement (only when needed)

3a - Reference documents:

# First: get the manifest
skills_get_reference(skill_id="test-writer", filename="list")
# Then: fetch the specific file the instructions mentioned
skills_get_reference(skill_id="test-writer", filename="TESTING-GUIDE.md")

3b - Execute a helper script:

# First: see available scripts
skills_run_script(skill_id="test-writer", filename="list")
# Then: run with optional inputs
skills_run_script(skill_id="test-writer", filename="coverage_check.py",
                  input_data={"TARGET_DIR": "./src", "MIN_COVERAGE": "80"})
# Returns: exit_code, stdout, stderr - source is never exposed

3c - Fetch a template or asset:

# First: get the manifest
skills_get_asset(skill_id="test-writer", filename="list")
# Then: fetch the template
skills_get_asset(skill_id="test-writer", filename="test-template.py")
# Use as a starting template - adapt it to the specific task

> Rule: Only load Tier 3 resources that the Tier 2 instructions explicitly > reference. Do not load them speculatively - it wastes context and latency.

---

Complete Example

# 1. Discover
results = skills_find_relevant(
    query="write integration tests for a Python REST API using pytest"
)
# → test-writer: 0.84  ← strong match

# 2. Load instructions
body = skills_get_body(skill_id="test-writer")
# → instructions: "## Testing Workflow ..."
# → tier3_manifest.references: ["TESTING-GUIDE.md"]
# → tier3_manifest.assets: ["test-template.py"]

# 3. The instructions say "follow TESTING-GUIDE.md for project conventions"
guide = skills_get_reference(skill_id="test-writer", filename="TESTING-GUIDE.md")

# 4. The instructions say "use test-template.py as your file scaffold"
template = skills_get_asset(skill_id="test-writer", filename="test-template.py")

# 5. Apply instructions using the guide and template

---

Decision Rules

| Situation | Action | |-----------|--------| | Starting any non-trivial task | Always call skills_find_relevant first | | All scores 0.4 | Call skills_get_body to load instructions | | Instructions are complete | Apply them - stop, do not load Tier 3 | | Instructions mention a file | Load that specific file via Tier 3 | | User asks to customise skill | Call skills_get_options | | Task is trivial (.workers.dev/sse`

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Jignesh-Ponamwar
• Source: Jignesh-Ponamwar/skills-mcp
• License: Apache-2.0
• Homepage: https://skills-mcp-jignesh.vercel.app/

Install and usage instructions live in the source repository linked above.