Claude Code System Prompts (Piebald)
Prompt LibraryVersion-by-version teardown of Claude Code's system prompts and 27 built-in tool descriptions, updated with every release.
- system-prompt
- claude-code
- analysis
- prompt-engineering
Simon Willison's git-history mirror of Anthropic's published claude.ai system prompts — every revision a dated commit, every change summarized.
Anthropic publishes the system prompts behind claude.ai and the Claude mobile apps at platform.claude.com, one page per model with a dated section for every revision. Simon Willison turns those pages into a repository with real git history: every published revision is a commit dated the day Anthropic says it went live, so git log, git diff and GitHub's history views show exactly how the prompts evolved across the Fable, Opus, Sonnet and Haiku families — each change also gets a short LLM-written summary of what is new (for example, Fable 5 → 5.1 on 2026-09-01 tightened copyright refusal rules). Files are layered: claude-<family>.md carries the lineage across model generations, claude-<model>.md the current prompt per model, and claude-<model>-YYYY-MM-DD.md is an immutable permalink to a prompt as it stood on a given date. An Atom feed and CHANGELOG.md deliver revisions as they are published.
Prompt-engineering research, tracking how Anthropic's safety and behavior policy evolves release over release, and citing what the system prompt said on a specific date — source material rather than an installable tool.
# reference repo — no install needed: git clone https://github.com/simonw/claude-system-prompts.git # browse prompts/, or subscribe: https://simonw.github.io/claude-system-prompts/feed.atom
No license file at the time of writing — the prompts are Anthropic's published text, mirrored with attribution, so treat reuse accordingly. Revisions appear after Anthropic publishes them, not before.
Repository
simonw/claude-system-promptsCharted
Version-by-version teardown of Claude Code's system prompts and 27 built-in tool descriptions, updated with every release.
A single-file CLAUDE.md distilling Karpathy's coding observations into four working rules (211k stars).
The de-facto standard ChatGPT role-prompt collection (f.k.a. Awesome ChatGPT Prompts).