MCP server that dynamically loads skills from Markdown files for AI assistants
The server defines a single dynamic tool 'dynamic_skill_tools' with critical quality issues. The tool has an empty input schema (no parameters), a generic description that does not follow LLM-optimized patterns, and no documented output schema. The tool name itself is vague and does not start with a clear action verb. The description is ~180 characters and explains what the tool does in general terms, but lacks WHEN to use it, WHAT it returns structurally, and any prerequisites. Per the rubric baseline, A+ tools have descriptions of 50-200 chars (this is 180, acceptable length) but must be LLM-optimized with clear intent signaling, this one reads like internal documentation rather than a prompt-optimized guide. Most critically, the empty input schema ({}) means the tool accepts no parameters, which severely limits its utility and forces the LLM to invoke it without control, a red flag for composition and reusability. The tool's behavior is file-system driven (scanning ~/.ai-skills-hub/skills/) rather than parameterized, making it impossible for the LLM to specify which skill to load or control execution context. No error handling guidance, no structured output specification, no security considerations documented.
Dynamically loaded skill tools from Markdown files in ~/.ai-skills-hub/skills directory. Each SKILL.md file in a skill subdirectory becomes a callable tool with the skill name converted to a tool identifier (hyphens converted to underscores). When invoked, returns the skill content with execution context (skill name, directory path) and available resources index.
Empty input schema, tool accepts zero parameters and thus cannot be controlled by the LLM. Schema is {"type":"object","properties":{},"required":[]}, which means the LLM has no way to specify which skill to load, control behavior, or pass context.
Tool name 'dynamic_skill_tools' is non-specific and does not start with a clear action verb. 'dynamic_skill_tools' sounds like a category or module, not an action. The LLM cannot infer intent from the name alone. Suggested renames: 'get_skill' (if accepts skill_name param), 'list_skills' (if returns available skills), or 'load_skill_content' (if fetches markdown).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 25 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Description does not follow LLM-optimized pattern. It reads like internal documentation: 'Dynamically loaded skill tools from Markdown files in ~/.ai-skills-hub/skills directory...' It lacks explicit WHAT, WHEN, and WHAT IT RETURNS structure. State WHAT the tool does, WHEN to use it, and any prerequisites.' Current description: ~180 chars, acceptable length (p10=34, p90=392), but content is vague. LLM cannot determine: (1) When to call this vs other tools? (2) What data does it return structurally (list of skills, single skill content, metadata)? (3) What are the prerequisites (must have ~/.ai-skills-hub/ directory)? (4) What happens on error (no 404 guidance).
No documented output schema. The tool returns skill content as a text string with embedded execution context and resource index (visible in src/index.ts call_tool handler), but LLM has no formal schema to understand structure. LLMs need to know what fields to expect so they can plan downstream tool calls and extract the right data.' The response type is content: [{type: 'text', text: ...}], but the LLM has no schema declaring what fields are in that text or how to parse the execution context metadata.
No error handling guidance in tool description. Try search_users() with a partial name." A raw error code or stack trace gives the agent nothing to act on.' Code shows generic error throws (e.g., 'Invalid tool name: {name}', 'Failed to read skill file: {filename}'), which do not guide recovery. If a skill file is malformed or missing, what should the LLM do? Call list_skills? Check filesystem? No guidance.
Tool behavior is file-system driven with no LLM control. The tool scans ~/.ai-skills-hub/skills/ on every invocation and returns dynamic results based on what files exist locally. The LLM cannot specify which skill to load (no 'skill_name' parameter) or filter/limit results. This violates composition pattern: 'Each tool should do exactly one thing.' The tool conflates discovery (list all skills), loading (fetch a skill), and enrichment (add execution context + resource index) in a single tool with no parameters. A better design: separate tools for (1) list_available_skills(), (2) get_skill(skill_name), (3) get_skill_resources(skill_name).
No security considerations documented. "read:email", "write:calendar"). This enables least-privilege agent configurations and clear audit trails.' The tool reads arbitrary files from ~/.ai-skills-hub/skills/, including user-created Markdown files. No mention of: (1) File read permissions required, (2) Path traversal safeguards, (3) What happens if a SKILL.md file contains malicious content (command injection, prompt injection), (4) Rate limits on filesystem scans. The code shows basic error handling (try/catch) but no input validation against path traversal (e.g., skill_name='../../../etc/passwd').