跳到正文
cccskills
免费在 GitHub 公开

mcp-server-instructions

Write effective MCP server instructions for tool compaction support. Use when creating or updating the `instructions` parameter in MaaSProtectedFastMCP server initialization. Helps LLMs understand when to search for tools, workflow patterns, and cross-feature relationships.

中文简介正在准备中。当前显示原文说明。

集成工具GitLab
查看安装方法

包含的文件(1)

  • SKILL.md4.6 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

MCP Server Instructions

Write workflow-focused server instructions that help LLMs understand your server's capabilities.

What Instructions Are

The instructions field is an optional string in MCP's InitializeResult that gets injected into the LLM's system prompt. It helps models understand:

  • How to use your server's tools together
  • When to search for specific tools (with tool compaction)
  • Operational constraints and patterns

Evidence: In controlled testing, well-written instructions improved task success from 60% to 85% - a 25% improvement.

Template

instructions="""[Server name] for [domain] operations.

KEY WORKFLOWS:
- [Name]: [tool1] -> [tool2] -> [tool3]
- [Name]: [tool1] -> [tool2] (when [condition])

CONSTRAINTS:
- [Limit or behavior users need to know]
- [Input format quirks]
- [Auto-pagination or chunking behavior]"""

What to Include

Focus on what individual tool descriptions don't convey:

IncludeExample
Cross-tool relationships"Always call authenticate before any fetch_* tools"
Operational patterns"Search results expire after 10 minutes"
Constraints"Rate limited to 100 requests/minute"
Sequencing"Use get_metadata before fetching large content"
Format flexibility"IID accepts: #72, !72, '72', 72"

What to Avoid

AvoidWhyInstead
Repeating tool descriptionsAlready in tool docstringsShow how tools work together
Marketing language"comprehensive", "powerful" add no valueState capabilities factually
Lengthy manualsInstructions should be <500 charsBe concise and actionable
Listing all toolsTools are discoverableOnly mention in workflow context

Bad: "The search tool searches for files" Good: "Use search before read to validate paths. Results expire after 10 min."

Process

  1. Find tools - Read server.py, locate all @app.tool() decorators
  2. Group by workflow - Which tools are used together?
  3. Find dependencies - Which tools need output from others?
  4. Note constraints - Pagination, size limits, rate limits, formats
  5. Identify non-obvious behaviors - Auto-pagination, caching, chunking

Common Patterns

  • Search -> Details: search_X -> get_X
  • List -> Inspect -> Act: list_X -> get_X -> update_X
  • Metadata First: get_X_metadata -> get_X_content (for large content)
  • Two-Step Updates: get_editmeta -> update_X (when schema required)
  • Hierarchical Browse: list_parent -> get_parent -> list_children

Examples

Good

instructions="""GitLab server for repository and CI/CD operations.

KEY WORKFLOWS:
- Code Review: list_merge_requests -> get_merge_request -> get_merge_request_diffs
- Pipeline Debug: list_pipelines_jobs -> get_job_log_metadata -> get_job_log_paginated
- File Browse: search_repositories -> get_repo_tree -> get_file_contents

CONSTRAINTS:
- MR IID accepts: #72, !72, "72", 72
- Job logs >500KB auto-paginate to last 2000 lines
- Use get_job_log_metadata before fetching large logs"""

Bad

# Too vague - doesn't help LLM understand capabilities
instructions="GitLab MCP Server that validates tokens via Authorization Server introspection"

# Marketing fluff - no actionable information
instructions="""This is a comprehensive GitLab integration server that provides
powerful capabilities for working with GitLab repositories..."""

# Tool listing - just repeats what's already in tool descriptions
instructions="""Available tools: search_repos, list_projects, get_project,
get_file, list_commits, get_commit, list_issues, get_issue..."""

Limitations

Instructions are probabilistic guidance, not deterministic rules. They improve LLM behavior but cannot guarantee it. Don't rely on instructions for security-critical or privacy-sensitive enforcement.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

prd

免费

Generate Product Requirements Documents (PRDs) in Markdown format. Use when users want to create a PRD, spec out a feature, document requirements, or plan new functionality. Triggers on requests like "create a PRD", "write requirements for", "spec out this feature", or "document this feature idea".

中文简介正在准备中。当前显示原文说明。

adawalli/nexus222026年8月13日 更新

查看 adawalli 的全部技能

报告此技能的问题