Skill Detail

Diagnose retrieved-but-unused agent skills through Ratel MCP

Compare skill retrievals with actual invocations through Ratel MCP, inspect the queries behind unused matches, and return an evidence-backed catalog diagnosis.

Monitoring & AlertsMCP
Monitoring & Alerts MCP Published
⭐ 561 GitHub stars
COPY SKILL INSTRUCTIONS (OPTIONAL)
npx skills add agentskillexchange/skills --skill diagnose-retrieved-but-unused-agent-skills-ratel-mcp Copy
Uses the third-party skills CLI, not an ASE-owned installer. Check your agent’s compatibility. This copies instructions; complete the upstream tool setup below separately.
At a glance
Tools required
MCP client; self-hosted Ratel with project retrieval/invocation telemetry; project API key; Docker Compose 2.24+ only if hosting a new instance
Install & setup
Use an existing Ratel instance with populated project telemetry. For a new local instance, follow the upstream README: git clone https://github.com/ratel-ai/ratel.git; cd ratel; docker compose up -d –build –wait. Create a project API key in Settings > API keys. Configure an MCP client for /api/mcp with Authorization: Bearer . Verify get_project_overview before running the read-only diagnosis. Docker Compose 2.24+ is required for local hosting; do not enable experimental features for this workflow.
Author
Ratel
Publisher
Organization
Last updated
Nov 14, 2026
Quick brief

Invoke this workflow when an operator needs to explain why an agent repeatedly retrieves a skill but never invokes it. Inputs are an existing self-hosted Ratel project, its recorded retrieval and invocation events, a project API key, and the exact runtime skill names. The output is a reviewable diagnosis linking unused matches to the user queries that retrieved them.

How it works

What this skill actually does

Connect an MCP client to POST <BASE_URL>/api/mcp with the project’s bearer key. Start with get_project_overview to confirm project identity and whether telemetry exists. Call get_improvement_signals to find retrieved-but-unused skills, then pass each exact skillName to list_search_queries_for_skill. Compare the returned queries with the intended skill purpose. Record mismatched intent, unknown catalog names, empty results, row limits and truncation as evidence gaps.

The scope boundary is read-only catalog diagnosis using those three MCP operations. Return suggested description changes for operator review; do not update skills, publish overrides, enable experimental evaluation features or claim that low invocation proves poor task quality. Generic trace dashboards cover latency and cost, while this workflow joins skill identity with retrieval intent and actual usage. An empty telemetry sample supports no quality conclusion. Query text may contain user data, so keep the evidence report within the operator’s approved project scope. The upstream worked example and operation definitions document this sequence.