Claude Code System Prompts — Summary

claude-code-system-prompts by Piebald AI is a reverse-engineering reference repository that extracts and publishes all of Claude Code's internal system prompts, tool descriptions, agent prompts, and system reminders — updated within minutes of each Claude Code release, tracking 187+ versions since v2.0.14. At Claude Code v2.1.150 (May 22, 2026), the repo documents 220+ named prompt strings grouped into: Agent Prompts (general-purpose, plan mode, explore, code review 8-part series), Data files (API references, managed agents docs), Skills (35+ skill definitions), System Prompts (60+ behavioral instructions), System Reminders (40+ context-injection messages), and Tool Descriptions (75+ tool parameter specs). The companion product is tweakcc — a tool that patches users' local Claude Code installations with custom versions of individual system prompt strings. The project is published by Piebald AI, which also ships a commercial IDE product (Piebald).

Differs from seeds: This is a unique archetype not represented in the 11 seeds — a system-prompt observatory rather than a user-facing framework. It ships zero runnable Claude Code primitives (no commands, skills, hooks, MCP servers) and has no user workflow. Its value is entirely informational: showing what Claude Code's internal instructions look like, enabling security research, prompt tuning, and the tweakcc patching tool. Closest architectural analog might be ccmemory (infrastructure-level tool) but the purpose is documentation not augmentation.

Claude Code System Prompts — Overview

Origin

Published by Piebald AI (https://piebald.ai/), a commercial IDE company built on top of Claude Code. 10,516 stars, MIT license. Updated within minutes of each Claude Code release since v2.0.14 (187+ versions tracked).

Mission (verbatim)

"This repository contains an up-to-date list of all Claude Code's various system prompts and their associated token counts as of Claude Code v2.1.150 (May 22nd, 2026). It also contains a CHANGELOG.md for the system prompts across 187 versions since v2.0.14."

"Claude Code doesn't just have one single string for its system prompt. Instead, there are: Large portions conditionally added depending on the environment and various configs. Descriptions for builtin tools like Write, Bash, and TodoWrite... Separate system prompts for builtin agents like Explore and Plan. Numerous AI-powered utility functions... The result — 110+ strings that are constantly changing and moving within a very large minified JS file."

Companion product: tweakcc

"Want to modify a particular piece of the system prompt in your own Claude Code installation? Use tweakcc. It lets you customize the individual pieces of the system prompt as markdown files, then patches your npm-based or native (binary) Claude Code installation with them, and also provides diffing and conflict management for when both you and Anthropic have conflicting modifications to the same prompt file."

Extraction method

"This repository contains the system prompts extracted using a script from the latest npm version of Claude Code. As they're extracted directly from Claude Code's compiled source code, they're guaranteed to be exactly what Claude Code uses."

Strategic value

• Security research into Claude Code's internal behavioral constraints
• Understanding what Anthropic's instructions look like at production scale
• Prompt engineering reference: real production instructions from Anthropic
• Identifying changes between CC versions (CHANGELOG.md tracks all prompt diffs)

Claude Code System Prompts — Architecture

Distribution

• GitHub repository — pure reference documentation
• No npm package, no CLI tool (tweakcc is a separate repository)

Directory structure

.
├── README.md
├── CHANGELOG.md          (187+ version prompt diffs)
├── CLAUDE.md
├── LICENSE               (MIT)
├── system-prompts/       (220+ extracted prompt files)
│   ├── agent-prompt-*.md         (agent system prompts)
│   ├── data-*.md                 (reference data bundled in CC)
│   ├── skill-*.md                (skill definitions)
│   ├── system-prompt-*.md        (core behavioral instructions)
│   ├── system-reminder-*.md      (context-injection messages)
│   ├── tool-description-*.md     (tool parameter descriptions)
│   └── tool-parameter-*.md       (tool parameter schemas)
└── tools/                (extraction utilities — not public)

Naming convention

Each file is named <category>-<kebab-description>.md with embedded HTML comment metadata:

<!--
name: 'Agent Prompt: General purpose'
description: ...
ccVersion: 2.1.86
agentMetadata:
  agentType: 'general-purpose'
  tools: [*]
  whenToUse: >
    General-purpose agent for researching...
-->

Target audience

• Security researchers
• Claude Code power users
• Prompt engineers
• Piebald IDE users
• tweakcc users wanting to customize CC behavior

Runtime requirements

None — read-only reference.

Claude Code System Prompts — Components

No runnable primitives

0 commands, 0 skills, 0 agents, 0 hooks, 0 MCP servers.

Extracted prompt categories (220+ files)

Agent Prompts (30+)

Data Files (20+)

Reference documentation bundled into Claude Code (Claude API references in 8 languages, managed agents docs, tool use concepts, etc.)

Skills (35+)

System Prompts (60+)

Core behavioral instructions: communication style, memory instructions, security, subagent delegation, tool usage policy, auto mode, minimal mode, harness instructions, etc.

System Reminders (40+)

Context-injection messages: session continuation, plan mode active, memory file contents, hook events, compact file reference, etc.

Tool Descriptions (75+)

Parameter schemas and usage instructions for: Bash, Edit, Write, Read, Grep, Task, TodoWrite, WebFetch, WebSearch, MCP tools, etc.

Claude Code System Prompts — Uniqueness

Differs from seeds

No seed framework matches this archetype. This is the only system-prompt observatory in the corpus — a reverse-engineering reference repository, not a user workflow framework. The closest structural analog is ccmemory (infrastructure-level tool), but ccmemory augments CC behavior while claude-code-system-prompts only documents it. The 8-part code review pipeline extraction reveals that Anthropic has already built a sophisticated multi-stage review system internally — which many community frameworks are unknowingly re-implementing. The CHANGELOG.md tracking 187 versions of prompt evolution is unprecedented in this batch.

Positioning

A forensic tool for advanced Claude Code users, security researchers, and framework builders who want to understand what Anthropic actually ships. The companion tweakcc product enables surgical modification of CC's internal behavior — more powerful than any CLAUDE.md convention.

Strategic insight for the catalog

This repo reveals that Anthropic's own internal CC system uses:

• Sequential multi-agent code review (8-stage pipeline)
• Dream memory consolidation/pruning
• Dynamic pacing for loop execution
• Structured vs plain-text subagent output variants
• Security monitoring subagents (2-part)
• Permission allowlist generation from transcripts

These are capabilities that community frameworks attempt to approximate but Anthropic has already formalized internally.

Observable failure modes

1. Legal gray area: Extracting and publishing proprietary compiled code's internal strings. Anthropic has not objected publicly but could.
2. Stale cache risk: 187 tracked versions suggests prompts change every few days; any downstream user of the prompts must re-fetch frequently.
3. Commercial framing: The prominent Piebald advertisement may signal that the reference repo is primarily a lead-generation tool.
4. Template literal interpolation: The ${} syntax in extracted prompts shows these are code-interpolated strings, not static text — the published versions may lack runtime-injected context variables.