Agent Skills & Prompts

Relevant source files

The following files were used as context for generating this wiki page:

• code\_review\_graph/skills.py
• code\_review\_graph/prompts.py
• skills/review-pr/SKILL.md
• AGENTS.md
• code\_review\_graph/main.py
• CLAUDE.md

Agent Skills and Prompts form the orchestration layer of the code-review-graph system, enabling AI coding assistants to interact effectively with the underlying knowledge graph. This module provides structured guidance, automated hooks, and platform-specific configurations for tools like Claude Code, Cursor, and GitHub Copilot to ensure token-efficient codebase exploration and high-accuracy code reviews.

Sources: code_review_graph/skills.py:1-12, code_review_graph/prompts.py:1-10, CLAUDE.md:1-10

Agent Skill Architecture

Skills are defined as specialized instruction sets that guide an AI agent through complex workflows such as debugging, refactoring, or reviewing pull requests. The system automatically generates these skills in Markdown format, often including frontmatter metadata for the host AI platform.

Skill Categories

The project defines four primary internal skills used to structure agent behavior:

Skill Name	Description	Key Focus Area
Explore Codebase	Navigate structure via knowledge graph	Architectural overview and relationship tracing
Review Changes	Structured, risk-aware code review	Change detection and impact radius analysis
Debug Issue	Systematically trace and resolve bugs	Semantic search and execution flow tracing
Refactor Safely	Plan/execute safe refactoring	Dead code detection and rename previews

Sources: code_review_graph/skills.py:270-400

Skill Generation Flow

The generate_skills function manages the creation of these instructions, placing them in platform-specific directories like .claude/skills/ or .gemini/skills/.

MCP Prompt Templates

Prompt templates provide multi-turn conversation starters for AI agents, enforcing the "Graph-First" workflow. Every prompt follows the Token Efficiency Rules, which mandate using minimal context before escalating to deeper inspections.

Core Prompts

1. review_changes_prompt: A workflow for pre-commit reviews that starts with a risk overview and drills into high-risk functions.
2. architecture_map_prompt: Generates architecture documentation including Mermaid diagrams and community coupling summaries.
3. debug_issue_prompt: Uses semantic search keywords to find relevant nodes and trace call chains.
4. onboard_developer_prompt: Orients new developers using high-level stats and critical execution flows.
5. pre_merge_check_prompt: A GO/NO-GO assessment focused on test gaps and newly introduced dead code.

Sources: code_review_graph/prompts.py:44-123, code_review_graph/main.py:645-685

Prompt Data Flow

Prompts are registered within the FastMCP server and return a list of Message objects to the client.

Platform Hooks & Automation

To keep the knowledge graph synchronized with the source code, the system installs automated hooks into various IDEs and version control systems.

Git Pre-Commit Hook

The install_git_hook function creates a pre-commit script in the .git/hooks directory. This script runs detect-changes --brief to surface risk summaries to the developer before a commit is finalized.
Sources: code_review_graph/skills.py:539-585, AGENTS.md:59-66

Editor Hooks

The system supports specific hook configurations for different AI environments:

• Claude Code/Qoder: Uses PostToolUse and SessionStart matchers to trigger graph updates after file edits.
• Cursor: Installs shell scripts (e.g., crg-update.sh) into ~/.cursor/hooks/ for events like afterFileEdit and beforeShellExecution.
• OpenCode: Deploys a TypeScript plugin that uses the Bun shell API to monitor file.edited events.

Sources: code_review_graph/skills.py:435-517, code_review_graph/skills.py:801-850, code_review_graph/skills.py:910-970

Token Efficiency Rules

A critical component of all skills and prompts is the enforcement of token-saving protocols to minimize LLM overhead.

• Initial Call: Agents MUST start with get_minimal_context(task="<description>"), costing ~100 tokens.
• Detail Level: Use detail_level="minimal" by default; only escalate to "standard" for specific entities.
• Tool Targets: Aim for ≤5 tool calls per task and ≤800 total context tokens.
• Targeted Queries: Prefer query_graph_tool with a specific symbol over broad list commands.

Sources: CLAUDE.md:12-18, code_review_graph/prompts.py:16-30, skills/review-pr/SKILL.md:11-13

Summary

Agent Skills & Prompts provide the interface between the user's intent and the system's structural intelligence. By combining platform-specific skill files with standardized MCP prompt templates and automated execution hooks, the module ensures that AI agents interact with the codebase in a structurally aware, risk-scored, and token-efficient manner.

Sources: code_review_graph/skills.py:1-12, CLAUDE.md:162-178