feat: add mcp documentation skill for spec-aware development (#609)

Add a Claude Code skill (`.claude/skills/mcp-docs/`) that provides live
MCP documentation and specification context

## Motivation and Context
The modelcontextprotocol.io site provides both an [MCP
server](https://modelcontextprotocol.io/mcp) and an
[llms.txt](https://modelcontextprotocol.io/llms.txt) index for
machine-readable access to the documentation. This skill leverages both
to give Claude live access to the current spec during development.

## How Has This Been Tested?
The skill was evaluated on 3 realistic scenarios (implementation, code
review, spec question) comparing with-skill vs without-skill baseline:
- **Pass rate**: 100% vs 67% — skill consistently cites spec sources and
uses RFC 2119 language
- **Token efficiency**: 41% fewer tokens on average (44k vs 75k) — goes
directly to spec instead of reverse-engineering from codebase
- **Accuracy**: with-skill discovered URL mode elicitation (new in spec
2025-11-25) that baseline missed entirely
- **Description triggering**: optimized through 5 iterations of
automated trigger evaluation against 20 test prompts

## Breaking Changes
None

## Types of changes
- [ ] Bug fix (non-breaking change which fixes an issue)
- [ ] New feature (non-breaking change which adds functionality)
- [ ] Breaking change (fix or feature that would cause existing
functionality to change)
- [x] Documentation update

## Checklist
- [x] I have read the [MCP
Documentation](https://modelcontextprotocol.io)
- [x] My code follows the repository's style guidelines
- [x] New and existing tests pass locally
- [x] I have added appropriate error handling
- [x] I have added or updated documentation as needed
This commit is contained in:
Pavel Gorgulov 2026-03-18 17:53:26 +01:00 • committed by GitHub
commit 2542c9cae1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 77 additions and 1 deletions

View file

@ -0,0 +1,76 @@
---
name: mcp-docs
description: >
Fetch live MCP specification and docs from modelcontextprotocol.io. Training data may be
outdated — default to using this skill when the query touches MCP protocol in any way.
Trigger for: implementing or debugging MCP features (sampling, elicitation, completion, tools,
resources, prompts, transports); SEPs or spec requirements; MCP transport behavior (SSE,
Streamable HTTP, reconnection, retry); authorization/OAuth; conformance test failures;
capability negotiation; message schemas. Skip only for pure Kotlin, build system, or
refactoring tasks with zero MCP protocol dependency. When in doubt, trigger.
user_invocable: true
---
# MCP Documentation Skill
Provides access to the official MCP specification and docs from modelcontextprotocol.io. The spec evolves faster than training data — always consult live documentation for protocol specifics like message schemas, required fields, capability negotiation, error handling, and transport requirements.
## When to use
**Proactively** (without being asked):
- Implementing a feature defined in the MCP spec
- Reviewing code that implements MCP protocol behavior
- Fixing conformance test failures
**On demand** (user asks or invokes `/mcp-docs`):
- Questions about how a specific MCP feature works
- Clarifying spec ambiguities
- Looking up SEPs (Specification Enhancement Proposals)
## How to fetch documentation
### Strategy: MCP tool first, llms.txt fallback
**Step 1 — Try the MCP search tool.** If `search_model_context_protocol` is available, use it:
```
search_model_context_protocol(query: "<your search query>")
```
One call, semantically ranked results. Fastest path.
**Step 2 — Fallback to llms.txt + WebFetch.** If the MCP tool is not available:
1. Fetch the documentation index:
```
WebFetch("https://modelcontextprotocol.io/llms.txt", "Return the full content as-is")
```
2. Identify the relevant page URL(s) from the index. Key sections:
- **Specification**: `specification/...` — protocol details, message formats, transports, lifecycle
- **SEPs**: `seps/...` — specification enhancement proposals
- **Learning**: `docs/learn/...` — architecture, concepts
- **Development**: `docs/develop/...` — building clients/servers
3. Fetch the specific page:
```
WebFetch("<page-url>", "Return the full content of this documentation page")
```
### Version handling
Always use the latest spec version by default. Only use a specific version or draft when the user explicitly requests it (e.g., "check the 2025-03-26 spec" or "what does the draft say").
### Search guidance
- Protocol message formats, required fields, capability negotiation → **Specification** section
- Feature motivation, design rationale → relevant **SEP**
- High-level concepts and architecture → **Learning** resources
- Use your judgment on query count. A field-level question may need one page; implementing a feature may need 2-3.
## Presenting information
- **Always cite sources.** Include links to spec pages used: `(source: <url>)`
- **Spec over training data** for specific protocol details (schemas, fields, error codes, capabilities). Training data is fine for general MCP concepts.
- **Preserve RFC 2119 language.** The spec uses MUST/SHOULD/MAY — keep this distinction when explaining requirements.
- **Be concise.** Focus on the details relevant to the user's task. Avoid dumping entire spec pages — extract what matters for the specific question or implementation at hand.

2
.gitignore vendored
View file

@ -56,7 +56,7 @@ node_modules
dist
### SWE agents ###
.claude/
.claude/settings.local.json
.junie/
### Conformance test results ###