Reverse Engineer

/reverse-engineer

Perform a complete reverse engineering analysis of any public GitHub repository. No API keys required — the Python script uses pure static heuristics, and Claude Code (you) provides the AI-powered narrative on top of the results.

Usage

/reverse-engineer <github-repo-url>

Examples:

/reverse-engineer https://github.com/nopSolutions/nopCommerce
/reverse-engineer https://github.com/spring-projects/spring-petclinic
/reverse-engineer https://github.com/django/django
/reverse-engineer https://github.com/expressjs/express

HOW TO SET UP THIS SKILL IN CLAUDE CODE

This skill lives at .claude/skills/reverse-engineer/SKILL.md.
Any SKILL.md inside .claude/skills/<name>/ automatically becomes a /name slash command in Claude Code.

Legacy path also works: .claude/commands/reverse-engineer.md
Both formats are supported. The skills/ folder is the current standard.

One-time setup (2 minutes)

1. Install Claude Code (if not already installed):

npm install -g @anthropic-ai/claude-code
claude --version   # verify

2a. Project-level install (skill available only in this folder):

your-project/
└── .claude/
    └── skills/
        └── reverse-engineer/
            └── SKILL.md          ← this file

Copy reverse-engineer/ into your project's .claude/skills/ directory.

2b. Global install (skill available in ALL your projects — recommended):

# macOS / Linux
mkdir -p ~/.claude/skills/reverse-engineer
cp SKILL.md ~/.claude/skills/reverse-engineer/SKILL.md

# Windows (PowerShell)
New-Item -ItemType Directory -Force "$env:USERPROFILE\.claude\skills\reverse-engineer"
Copy-Item SKILL.md "$env:USERPROFILE\.claude\skills\reverse-engineer\SKILL.md"

Tip: Global install means /reverse-engineer works in every project you open with Claude Code — you only set it up once.

3. No pip install needed — zero external dependencies:

# requirements.txt is intentionally empty
# Only Python stdlib + git are required
python --version   # Python 3.8+ required
git --version      # git required for cloning repos

4. Open Claude Code in your project:

claude

5. Verify the skill is registered: Type / in the Claude Code chat — reverse-engineer should appear.

6. Run your first analysis:

/reverse-engineer https://github.com/nopSolutions/nopCommerce

HOW THIS COMMAND WORKS

1. Python script runs full static analysis (no API key needed)
2. Five output files are generated in ./outputs/{repo_name}/
3. Claude Code IS the AI engine — reads the SDD JSON and provides:
   • Executive summary and architecture explanation
   • Business logic and domain analysis
   • Modernisation roadmap with reasoning
   • Block diagram walkthrough
4. Claude Code offers to write the AI content back into the report file

Instructions for Claude Code

When this command is invoked with a GitHub repository URL ($ARGUMENTS):

Step 1 — Validate input

Check that $ARGUMENTS is a valid GitHub URL (starts with https://github.com/). If not, ask: "Please provide a full GitHub URL (e.g. https://github.com/owner/repo)."

Step 2 — Check for the analysis script

python -c "import pathlib; print('ready' if pathlib.Path('reverse_engineer_skill.py').exists() else 'missing')"

Step 3a — Script found: Run static analysis

python reverse_engineer_skill.py $ARGUMENTS --heuristic

Do NOT run git clone before this. The script clones internally.

The --heuristic flag skips the interactive mode-selection prompt (since Claude Code IS the AI layer here). If the user passed --ai in $ARGUMENTS, the script will use Claude API for richer output file content instead.

The script uses layer-balanced static analysis — no API keys, no network calls beyond the initial git clone.

Pipeline stages:

1. Clone — shallow git clone --depth=1 to a temp directory
2. Load — discover source files (Python, Java, C#/.NET, JS/TS, up to 300 files)
3. Parse — extract classes, methods, routes, ORM entities (300-file layer-balanced cap)
4. Metrics — file/class/method counts, dependency graph
5. API & Dead Code — endpoints, database schema, tech stack, block diagram
6. Analysis — static heuristics (or Claude API if --ai flag passed)
7. Generate — SDD JSON, HTML dashboard, Markdown report, SVG diagrams
8. Evaluate — 100-point quality scorer
9. Cleanup — remove temp clone directory

Output files in ./outputs/{repo_name}/:

• {repo_name}_sdd.json — 14-section System Design Document
• {repo_name}_dashboard.html — Dashboard with block diagram & business logic
• {repo_name}_report.md — 12-section Markdown technical report
• {repo_name}_evaluation.md — 100-point quality evaluation
• {repo_name}_block_diagram.svg — Architecture block diagram (SVG)
• {repo_name}_dependency_graph.svg — Module dependency graph (SVG)
• manifest.json — Run record with file sizes

Step 4 — Read the analysis data (YOU are now the AI layer)

Read the generated SDD JSON — it contains all structured data you need:

outputs/<repo_name>/<repo_name>_sdd.json

Key sections to read:

• project — language, tech_stack, platform, layers
• codebase_metrics — file/class/method/endpoint counts, language distribution
• api_catalog — all API endpoints extracted
• data_architecture — ORM entities, relationships, microservice boundaries
• business_logic — heuristic values (you will enhance with AI reasoning)
  • block_diagram — the auto-generated Mermaid flowchart TB of the system
• dependency_analysis — external dependencies

Also read the Markdown report:

outputs/<repo_name>/<repo_name>_report.md

Step 5 — Provide AI Analysis (YOU are the AI engine)

You are a senior software architect. The Python script gave you structured facts. Now YOU interpret and narrate them with AI-quality reasoning.

Think through: What does this system actually do? Who uses it? How do the layers connect? What are the most important business workflows? What risks exist?

5a — Executive Summary

5b — Business Logic & Domain

1. Business Domain — e.g. "E-Commerce Platform", "HR Management System"
2. What It Does — 2–3 paragraphs plain-English from the user's perspective
3. Core Workflows — 3–5 key workflows, each with name, steps, and endpoints
4. User Roles — Who uses this system? (Customer, Admin, Vendor, etc.)
5. Key Business Rules — Rules the system enforces (infer from entities + routes)
6. Entity Glossary — For each data entity, explain what it represents in plain English
7. Block Diagram Walkthrough — Walk through the Mermaid block diagram in business_logic.block_diagram: explain each layer and how data flows top to bottom

5c — Architectural Block Diagram Explanation

The dashboard's "How It Works" section shows an auto-generated Mermaid flowchart TB. Explain it in detail:

• What each subgraph represents (Client → API → Business Logic → Data Access → DB)
• How a typical request flows through the layers
• Which controllers handle which workflows
• What the database entities represent in business terms

Step 6 — Output AI Analysis in Chat

Present your complete AI analysis clearly using markdown formatting:

Executive Summary

Purpose: [your analysis]

Architecture: [pattern detected]

Modernization Priority: [HIGH/MEDIUM/LOW] — [reasoning]

Top Tech Debt Concerns:

• [concern 1]
• [concern 2]
• [concern 3]

Business Logic & Domain

Business Domain: [domain name]

What It Does: [your 2–3 paragraph plain-English explanation]

Core Workflows:

1. [Workflow Name] — [description + key steps + endpoints involved]
2. ...

User Roles: [role 1], [role 2], ...

Key Business Rules:

• [rule 1]
• [rule 2]

Entity Glossary:

How the Architecture Works (Block Diagram Walkthrough)

Layer-by-layer explanation:

1. Client / User — [who makes requests]
2. API / Presentation Layer — [what controllers/handlers do]
3. Business Logic / Service Layer — [what services/managers do]
4. Data Access / Repository Layer — [how data is accessed]
5. Database — [what entities are stored and why]
6. External Services — [integrations detected]

Request flow: [describe how a typical request flows through the layers]

Step 7 — Write AI content directly into the report (no prompt needed)

Do NOT ask the user — automatically update the report file immediately after presenting the analysis in chat.

Edit outputs/<repo_name>/<repo_name>_report.md:

• Replace Section 1 (Executive Summary) with your AI-generated executive summary
• Replace Section 2 (Business Logic) with your full AI-generated business logic analysis
• Replace (or append to) Section 4 (Architecture Overview) with your block diagram walkthrough

Print a one-line confirmation after writing:

  [ok] AI analysis written to outputs/<repo_name>/<repo_name>_report.md

Step 8 — Report completion

Reverse engineering complete for: {repo_name}

Output files (./outputs/{repo_name}/):
  {repo_name}_sdd.json         — System Design Document (14 sections)
  {repo_name}_dashboard.html   — Dashboard (block diagram + business logic)
  {repo_name}_report.md        — Technical Report (12 sections, AI-enhanced and written automatically)
  {repo_name}_evaluation.md    — Quality Evaluation
  manifest.json                — Run Manifest

Analysis (static engine + Claude Code AI):
  Files analyzed    : N
  Classes found     : N
  Methods found     : N
  API endpoints     : N
  Data entities     : N
  Dead code files   : N
  Primary language  : LANG
  Business domain   : DOMAIN (AI-enhanced by Claude Code)
  AI engine         : Claude Code (no API key required)

Step 3b — Script missing (manual fallback path)

If reverse_engineer_skill.py is not found, perform the analysis manually:

Clone:

git clone --depth=1 $ARGUMENTS ./temp_analysis_repo

Discover source files: .py .java .cs .ts .tsx .js .jsx Skip: node_modules/ bin/ obj/ .git/ dist/ build/ migrations/ __pycache__/

For each file extract:

• Classes and interfaces defined
• Methods/functions defined (exclude language keywords)
• Import/using/require statements
• API routes: [HttpGet]/[Route] (C#), @GetMapping (Java), @app.route (Python), app.get (JS/TS)
• ORM entities: DbSet<X> (C#), @Entity (Java), class X(models.Model) (Python)

Generate 5 output files using templates in templates/ as schema guides:

• templates/sdd_template.json → {repo}_sdd.json (14 sections)
• templates/dashboard_template.html → {repo}_dashboard.html
• templates/report_template.md → {repo}_report.md (12 sections)
• Evaluate your own output → {repo}_evaluation.md
• manifest.json

Then proceed with Steps 4–8 above.

Notes

• Zero API keys required by default — Claude Code (you) IS the AI engine; heuristic mode needs no keys
• Optional --ai flag — pass --ai to use Claude API for richer content written directly into output files (ANTHROPIC_API_KEY env var required; pip install anthropic)
• Zero Python dependencies — only stdlib + git needed (stdlib only for heuristic mode)
• GitHub Copilot users: See .github/prompts/reverse-engineer.prompt.md
• Templates in templates/ define exact schemas
• The HTML dashboard is fully self-contained — open in any browser without a server
• Supports: Python, Java, C#/.NET, JavaScript, TypeScript (+JSX/TSX)
• SVG diagrams use hierarchical column layout (dependency graph) and tinted layer colors (block diagram)

Key project files