Claude Code /claude-code-internals skill

skill.md

name

description

claude-code-internals

Reverse-engineer and explain Claude Code's internal architecture. Extract prompts, commands, data structures from the binary. Use when user asks about how Claude Code works internally, what prompts it uses, how features are implemented, or wants to extract/backup internal logic. Triggers: 'claude inside', 'how does claude code work', 'extract prompt', 'extract prompt', 'claude architecture', 'internals', 'binary'.

Claude Code Internals Explorer

Reverse-engineer Claude Code's architecture by extracting and analyzing its bundled JavaScript from the binary. Save findings to ~/obsidian-vault/reference/ for the user's review.

Principles

1. Binary is the source of truth — Claude Code ships as a single binary with bundled JS. Use strings to extract.
2. Cache findings — Always save extracted prompts/structures to ~/obsidian-vault/reference/ so they survive across sessions.
3. Deobfuscate — The JS is minified. Variable names are mangled. Read the logic, not the names.
4. Version-stamp — Always note the Claude Code version in extracted docs.

Binary Location & Version

CLAUDE_BIN=$(which claude)
CLAUDE_VERSION=$(claude --version 2>/dev/null || strings "$CLAUDE_BIN" | grep '"VERSION"' | head -1)

Extraction Techniques

1. Find a specific feature/prompt

# Search for keywords related to the feature
strings "$CLAUDE_BIN" | grep -i "keyword" | head -20

# Get surrounding context (the prompt text)
strings "$CLAUDE_BIN" | grep -B 5 -A 50 "exact phrase from feature"

# For longer extractions
strings "$CLAUDE_BIN" | grep -B 2 -A 200 "start of prompt text" | head -300

2. Find all slash commands

# Commands are registered with type:"prompt" and name:"commandname"
strings "$CLAUDE_BIN" | grep 'type:"prompt",name:"'
strings "$CLAUDE_BIN" | grep 'description:".*".*progressMessage'

3. Find system prompts

# Main system prompt function
strings "$CLAUDE_BIN" | grep -A 100 'You are an interactive CLI tool'

# Tool descriptions
strings "$CLAUDE_BIN" | grep -A 20 'description:.*tool'

4. Find configuration schemas

# Settings, hooks, permissions
strings "$CLAUDE_BIN" | grep -A 30 'settings.json'
strings "$CLAUDE_BIN" | grep -A 30 'hooks'

5. Find API interaction patterns

# How Claude Code talks to the API
strings "$CLAUDE_BIN" | grep -A 20 'anthropic.com'
strings "$CLAUDE_BIN" | grep -A 10 'model.*claude'

Known Internal Architecture

File Layout

Path	Purpose
~/.claude/settings.json	User settings (permissions, hooks, model preferences)
~/.claude/projects/{encoded-path}/	Per-project session data
~/.claude/projects/{path}/*.jsonl	Session transcripts (JSONL format)
~/.claude/projects/{path}/{session-id}/	Session artifacts (tool results, subagent logs)
~/.claude/usage-data/facets/{session-id}.json	AI-analyzed session facets (from /insights)
~/.claude/usage-data/session-meta/{session-id}.json	Session metadata (duration, tools, tokens)
~/.claude/usage-data/report.html	Generated insights HTML report
~/.claude/stats-cache.json	Aggregate usage statistics
~/.claude/telemetry/	Telemetry data
~/.claude/history.jsonl	Command history across all sessions
~/.claude/skills/	User-defined skills

Session JSONL Structure

Each line is a JSON object with a type field:

type: "user"                  — User message (content: string | array of content blocks)
type: "assistant"             — Claude response (content: array with text/tool_use/thinking blocks)
type: "file-history-snapshot"  — File state snapshot (has timestamp)
type: "progress"              — Progress indicator
type: "queue-operation"       — Internal queue management
type: "summary"               — Conversation summary (from context compression)

Facet Schema (from /insights)

{
  "underlying_goal": "string",
  "goal_categories": {"category": count},
  "outcome": "fully_achieved|mostly_achieved|partially_achieved|not_achieved",
  "user_satisfaction_counts": {"level": count},
  "claude_helpfulness": "unhelpful|slightly_helpful|moderately_helpful|very_helpful|essential",
  "session_type": "single_task|multi_task|iterative_refinement|exploration|quick_question",
  "friction_counts": {"type": count},
  "friction_detail": "string",
  "primary_success": "none|fast_accurate_search|correct_code_edits|...",
  "brief_summary": "string",
  "session_id": "uuid"
}

Known Built-in Commands

Command	Description	Source
/insights	Session usage analysis + HTML report	builtin
/help	Help menu	builtin
/clear	Clear conversation	builtin
/compact	Compress context	builtin
/config	Open settings	builtin
/cost	Show token usage	builtin
/doctor	Diagnose issues	builtin
/fast	Toggle fast mode	builtin
/init	Initialize project CLAUDE.md	builtin
/login	Authenticate	builtin
/logout	Sign out	builtin
/mcp	MCP server management	builtin
/memory	Edit CLAUDE.md	builtin
/model	Switch model	builtin
/permissions	Manage tool permissions	builtin
/pr-review	Review a pull request	builtin
/review	Code review	builtin
/status	Show session info	builtin
/terminal-setup	Configure terminal integration	builtin
/vim	Toggle vim mode	builtin

Output Format

Include:

1. Version — Claude Code version at extraction time
2. Raw extraction — The actual extracted text/code
3. Deobfuscated explanation — What each part does in plain language
4. Usage notes — How this knowledge can be applied

Common Questions to Explore

• "How does context compression work?" → Search for summary, compact, compress
• "What's the system prompt?" → Search for You are an interactive, SV8
• "How do hooks work?" → Search for hooks, lifecycle, PreToolUse
• "How does /review work?" → Search for pr-review, review, diff
• "How are MCP tools loaded?" → Search for mcp, McpServer, tool_use
• "What models are available?" → Search for claude-opus, claude-sonnet, model
• "How does permission system work?" → Search for allowedTools, permission
• "How does Task agent work?" → Search for subagent, Task, agent