A Claude Code skill that generates deep, verified technical documentation from source code — zero hallucination, full traceability.
English | 中文
Asking Claude to "analyze my code" produces shallow, unreliable output. This skill solves that with:
- Anti-hallucination — Every claim requires
[VERIFY: file:line]code evidence. No exceptions. - Structured methodology — 9-phase progressive refinement (Phase 0–7, plus Phase 1.5 planning) ensures nothing is missed.
- Complete coverage — Phase 1.5 planning guarantees all modules/threads get analyzed.
- Automated verification — Built-in scripts validate every code reference after generation.
- Claude Code CLI installed and running
Option A: Claude Code Native (No plugins needed)
# Clone to Claude Code's global skills directory
# SKILL.md (uppercase) is auto-detected by Claude Code
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
~/.claude/skills/codebase-analysisRestart Claude Code. The skill is auto-triggered when you ask to analyze code.
Option B: With oh-my-claudecode (OMC)
# Provides /codebase-analysis slash command
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
~/.claude/plugins/marketplaces/omc/skills/codebase-analysisRestart Claude Code. Use /codebase-analysis <path> to trigger.
Option C: Project-Level (Per-Project)
# Only active within a specific project
mkdir -p your-project/.claude/skills
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
your-project/.claude/skills/codebase-analysisOption D: One-Time Use (No installation)
No clone needed. Just paste the content of skill.md as your prompt to Claude Code.
After installation, trigger the analysis:
# Option A/C (Native): Use natural language — Claude auto-detects the skill
> "Analyze the codebase at /path/to/project"
> "帮我分析这个项目的代码架构"
> "Document the algorithms in src/core/"
# Option B (OMC): Use the slash command
/codebase-analysis /path/to/project
# Any of these phrases trigger the skill:
# "analyze code" / "codebase analysis" / "document this code"
# "代码分析" / "代码梳理" / "数据结构分析" / "技术文档"Phase 0: Project Context Detection ← Auto-detect type, auto-/init if needed
Phase 1: Global Exploration ← Architecture, modules, threads
Phase 1.5: Analysis Planning ← Explicit plan for 100% coverage
Phase 2: Data Structure Analysis ← Struct/class deep dive
Phase 3: Data Flow Analysis ← Data movement tracing
Phase 4: Algorithm Deep Dive ← Math + code, 1500-3000 lines each
Phase 5: Key Function Analysis ← Line-by-line logic
Phase 6: Q&A Documentation ← Design rationale, configuration
Phase 7: Verification & Review ← Automated + manual validation
Every analysis passes through a mandatory 5-step verification pipeline:
Code Reading → Draft Generation → Code Verification → Fix Discrepancies → Final Output
Quality Gate: Zero unverified claims. Every [VERIFY: file:line] tag is validated.
The skill generates structured .md files in the target project directory:
| Document | Content | Typical Size |
|---|---|---|
ANALYSIS_00-SystemOverview.md |
Architecture, modules, threading, dependencies | 500-800 lines |
ANALYSIS_01-DataStructures.md |
Complete struct/class breakdowns with diagrams | 600-1,000 lines |
ALGORITHM_XX-Name.md |
Math derivations, step-by-step flows, code analysis | 1,500-3,000 lines each |
KEY_QUESTIONS-XX.md |
Design rationale, configuration guidance, pitfalls | 500-1,000 lines each |
All output is bilingual (English + Chinese headings) with ASCII architecture diagrams.
Generated for Voxel-SLAM (~10K LOC C++ SLAM system):
Total: 6,744 lines of verified technical documentation
Coverage: 4/4 threads (100%)
Quality: A+ (98/100)
Verification tags: 113 (98.9% valid)
Sample output excerpt
### Voxel Map Downsampling
Implementation: [VERIFY: VoxelSLAM/src/voxel_map.hpp:46-48]
pp.first.pnt = (pp.first.pnt * pp.second + pv.pnt) / (pp.second + 1);
This implements the running average formula:
μ_new = (μ_old · n + x_new) / (n + 1)Full file list generated
Voxel-SLAM/
├── ANALYSIS_00-SystemOverview.md (559 lines)
├── ANALYSIS_01-DataStructures.md (702 lines)
├── ALGORITHM_02-VoxelMap.md (2,219 lines)
├── ALGORITHM_03-IMUPreintegration.md (1,666 lines)
└── ALGORITHM_04-LiDARMeasurement.md (1,598 lines)
codebase-analysis-skill/
├── skill.md # Main skill definition (9-phase methodology)
├── WORKFLOW.md # Mandatory verification workflow
├── Gotchas.md # 15 common analysis pitfalls & prevention
├── CHANGELOG.md # Version history
├── CONTRIBUTING.md # Contribution guide
├── LICENSE # MIT License
│
├── templates/ # Document generation templates
│ ├── algorithm_flow.md # Deep algorithm analysis (1500-3000 lines)
│ ├── data_structures.md # Data structure breakdown
│ ├── key_questions.md # Q&A documentation
│ └── system_overview.md # Architecture overview
│
├── verification/ # Verification tools
│ ├── verify_all_refs.sh # Automated [VERIFY:] tag checker
│ ├── verify_analysis.sh # Full analysis verification
│ └── code_reference_checklist.md # Manual verification checklist
│
├── helpers/ # Helper scripts
│ ├── ascii_diagrams.sh # ASCII diagram utilities
│ └── code_ref_formatter.sh # Code reference formatting
│
├── docs/ # Detailed methodology docs
│ ├── ANALYSIS_FLOW.md # Progressive refinement explained
│ ├── INIT_GUIDE.md # Fresh clone handling
│ ├── PHASE1_ENHANCED.md # Phase 1 enhanced methodology
│ └── PHASE4_GUIDE.md # Algorithm deep dive guide
│
└── examples/ # Usage examples
└── example-output.md # Sample output from Voxel-SLAM
# Verify all [VERIFY:] tags in a generated document
./verification/verify_all_refs.sh ANALYSIS_00-SystemOverview.md
# Output:
# ✓ [VERIFY: src/voxel_map.hpp:46] - Valid
# ✓ [VERIFY: src/voxelslam.cpp:52] - Valid
# ❌ [VERIFY: BTC.h/BTC.cpp] - Malformed (combined path)
#
# Summary: 112/113 valid (98.9%)| Project | Language | Size | Output | Coverage |
|---|---|---|---|---|
| Voxel-SLAM | C++ (ROS) | ~10K LOC | 6,744 lines | 4/4 threads |
Tested on your project? Open a PR to add it here!
Q: Does this work with any programming language? A: Yes. The methodology is language-agnostic. Works for C++, Python, Rust, Go, Java, and more.
Q: How is this different from asking Claude to "analyze my code"?
A: Three things: (1) Structured 9-phase methodology ensures nothing is missed, (2) Mandatory [VERIFY:] tags prevent hallucination, (3) Automated verification tools validate every code reference.
Q: What if my project doesn't have CLAUDE.md?
A: The skill auto-detects project type from code structure. If /init is available, it will automatically generate CLAUDE.md for enhanced context.
Q: Can I use this without oh-my-claudecode?
A: Yes. The core methodology is in skill.md — use it as a prompt with any Claude Code session. OMC just provides the /codebase-analysis trigger.
See CONTRIBUTING.md for details. Quick summary:
- Fork → Branch → Commit → PR
- Areas we need help: new language templates, more gotchas, testing on diverse codebases
- Run
./verification/verify_all_refs.shto test your changes
- Skill methodology informed by Anthropic Claude Code team engineer Thariq's sharing: Lessons from Building Claude Code: How We Use Skills, including Gotchas anti-pattern, progressive disclosure, and trigger-based descriptions
- Inspired by the fastlivo2-tech-docs documentation style
- Tested on Voxel-SLAM codebase
- Built for Claude Code and oh-my-claudecode
一个 Claude Code 技能，从源代码生成深度、可验证的技术文档 — 零幻觉，全追溯。
- 防幻觉机制 — 每个论断都需要
[VERIFY: 文件:行号]代码证据 - 自动初始化 — 无 CLAUDE.md 时自动运行
/init获取最佳上下文 - 完整覆盖 — Phase 1.5 规划确保所有模块/线程都被分析
- 双语输出 — 英文 + 中文标题贯穿全文
- 可验证 — 自动化脚本验证每个代码引用
方式 A：Claude Code 原生（无需插件）
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
~/.claude/skills/codebase-analysis重启 Claude Code，用自然语言即可触发（如"分析这个项目的代码架构"）。
方式 B：oh-my-claudecode (OMC)
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
~/.claude/plugins/marketplaces/omc/skills/codebase-analysis重启 Claude Code，使用 /codebase-analysis <path> 触发。
方式 C：项目级别
mkdir -p your-project/.claude/skills
git clone https://github.com/Yangchengshuai/codebase-analysis-skill.git \
your-project/.claude/skills/codebase-analysis方式 D：一次性使用 — 直接将 skill.md 内容粘贴为提示词即可。
Phase 0: 项目上下文检测 ← 自动检测类型，缺少 CLAUDE.md 时自动 /init
Phase 1: 全局探索 ← 架构、模块、线程
Phase 1.5: 分析规划 ← 100% 覆盖率的明确计划
Phase 2: 数据结构分析 ← 结构体/类深入分析
Phase 3: 数据流分析 ← 数据流动追踪
Phase 4: 算法深度分析 ← 数学推导 + 代码，每篇 1500-3000 行
Phase 5: 关键函数分析 ← 逐行逻辑分析
Phase 6: 问题解答文档 ← 设计原理、配置指南
Phase 7: 验证与复核 ← 自动化 + 人工验证
MIT License — 详见 LICENSE 文件。