AgentSkillsCN

knowledge

Project Cortex——面向大语言模型的结构化知识注入

SKILL.md
--- frontmatter
name: "knowledge"
version: "1.0.0"
description: "Project Cortex - Structural Knowledge Injection for LLM"
routing_keywords:
  [
    "knowledge",
    "context",
    "rules",
    "standards",
    "documentation",
    "how to",
    "explain",
    "what is",
    "guidelines",
    "project rules",
    "conventions",
    "workflow",
  ]
authors: ["omni-dev-fusion"]
intents:
  - "Consult project rules"
  - "Look up coding standards"
  - "Check architecture decisions"
  - "Review workflow guidelines"

Knowledge Skill Policy

Router Logic

Scenario 1: User asks to "commit", "save", or "finish"

  1. Observe: Call get_development_context()
  2. Analyze: Extract commit types, scopes, and active guardrails
  3. Execute: Use git skill to commit (if Client lacks git)
  4. Verify: Ensure message follows format

Scenario 2: User asks "how do I...", "what is...", "explain..."

  1. Analyze: Identify topic from user question
  2. Search: Call consult_architecture_doc(topic)
  3. Synthesize: Summarize relevant information
  4. Respond: Provide concise answer with source references

Scenario 3: User asks to write documentation

  1. Prepare: Call get_writing_memory() FIRST
  2. Follow: Apply writing style rules
  3. Polish: Use writer skill for linting

Scenario 4: User asks to write code (Python, Nix, etc.)

  1. Standards: Call get_language_standards(lang)
  2. Context: Call get_development_context()
  3. Execute: Use appropriate skill (python, terminal, etc.)

Workflow: Commit Flow

code
User: commit

Claude:
  1. get_development_context() → {scopes: ["mcp", "core"], guardrails: ["vale"]}
  2. Generate message following format: "type(scope): description"
  3. Warn user about guardrails if applicable
  4. Execute git_commit

Workflow: Documentation Flow

code
User: write docs for feature X

Claude:
  1. get_writing_memory() → Load writing rules
  2. consult_architecture_doc("documentation") → Get doc structure
  3. Write document following rules
  4. Run vale check if available

Anti-Patterns

  • Don't use knowledge tools as execution shortcuts
  • Don't skip loading writing memory before documentation
  • Don't ignore guardrails returned by get_development_context()