Claude Code Internals

The full system prompts, all 58 tool definitions, and the token-accounting behavior of Claude Code — documented bilingually.

Chinese version · Detailed analysis

---

What's inside

Path	Contents
SUMMARY.md	Detailed analysis of every prompt section and all 58 tools
groups/	The prompt/tool content, one folder per unique combination

Each group folder contains the originals in English plus Chinese translations (*.zh-CN.md), and a meta.json with basic statistics.

Key findings

1. Every request carries ~11.4K chars of system prompt + ~124K chars of tool definitions (58 tools) before a single word of conversation.

2. ~60% of requests are not conversations at all — they are token-counting probes. Claude Code sends max_tokens=1 requests to weigh every context part (core tools / extended tools / MCP tools / Skill / text fragments separately), feeding the /context breakdown and compaction decisions.

3. The 58 tools decompose exactly into 11 core + 19 extended + 28 MCP + 1 Skill — the same grouping the probes measure.

4. The desktop app generates conversation titles with a separate Agent SDK process that loads its own 27K-char system prompt (with the full memory system) and no tools.

5. Reasoning effort is sent as output_config.effort with thinking: {"type": "adaptive"} — the newer qualitative replacement for the fixed budget_tokens thinking budget.

6. Different client builds send byte-identical prompts — only the build stamp (cc_version) and model-ID string differ.

Why is everything in Anthropic's format?

Claude Code is built by Anthropic on the Anthropic Messages API. Even when you point it at a different LLM endpoint, the client still constructs Anthropic-format requests — Anthropic system-prompt structure, Anthropic tool definitions, Anthropic thinking fields — and relies on the endpoint (or a gateway in front of it) to translate or accept that format. So what you see in this repository is determined by the client, not by whichever model ultimately answers: the prompts and tool definitions in groups/ are inherently Anthropic-flavored, regardless of backend.

Repository structure

├── README.md / README.zh-CN.md     ← this file (EN) / Chinese version
├── SUMMARY.md / SUMMARY.zh-CN.md   ← detailed analysis (EN) / Chinese version
├── LICENSE                         ← CC-BY-4.0
└── groups/                         ← prompt & tool content
    ├── title-generation-prompt/    ← 27K general-agent prompt (title-gen process)
    ├── main-prompt-and-tools/      ← ★ main prompt + 58 tools + skills
    ├── main-prompt-build-*/        ← build variants of the main prompt
    └── probe-*/                    ← token-counting probes (skill/extended/mcp/core/fragments)

Disclaimer & license

• The system prompts and tool definitions are Anthropic's intellectual property, presented here for educational research only. They change across versions (from cc_version=2.1.209). If Anthropic objects to this repository's presence, it will be taken down on request.
• The analysis, summaries, and translations are © the repository author, licensed under CC-BY-4.0.
• Personal paths and identifiers have been replaced with <placeholders>.