mirror of
https://github.com/asgeirtj/system_prompts_leaks.git
synced 2026-09-25 12:00:41 +02:00
Ásgeir Thor Johnson
171d1db270
As shipped in the CC 2.1.259 bundle, which differs from the open-source
copy in anthropics/skills (skills/claude-api/):
- model names/ids resolved (the bundle ships them as {{FABLE_NAME}}-style
catalog templates resolved at runtime; resolution verified byte-identical
against upstream for 54/69 files)
- includes the eval suite not yet published upstream: build-eval /
hillclimb subcommands in SKILL.md + shared/evals/ (build-eval,
eval-hillclimb, cost-hillclimb, eval-audit, report scaffolding)
- reviewer-only grounding comments stripped (as processSkillMarkdown does
at runtime; the model never sees them)
- SKILL.md frontmatter (name + description) taken verbatim from the live
skill registry
76 files: SKILL.md + shared/ + per-language dirs (python, typescript,
java, go, ruby, csharp, php, curl) + shared/evals/
Anthropic — which file is which product?
The bare claude-<model>.md files in this folder are the claude.ai system prompts — the prompt served to that model in the Claude web/mobile app (claude.ai). They are not API prompts (the Claude API injects no system prompt) and not Claude Code prompts.
| File pattern | Product |
|---|---|
claude-fable-5.md, claude-opus-4.8.md, claude-sonnet-5.md, … |
claude.ai app system prompt for that model |
claude-*-no-tools.md |
claude.ai with tools disabled |
claude-code/ |
Claude Code (the CLI/agent harness) |
claude-design.md |
Claude Design |
claude-cowork/ |
Claude Cowork |
claude-for-excel.md, claude-for-word.md, claude-in-powerpoint.md |
Claude in Microsoft 365 |
claude-in-chrome.md |
Claude in Chrome extension |
claude-mobile-ios.md |
claude.ai iOS app |
anthropic_reminders.md, sonnet-4.6-reminders.md, research_instructions.md, visualize.md |
claude.ai injected fragments (reminders, research, artifacts) |
official/ |
Prompts Anthropic publishes themselves (release-notes versions — shorter than the real served prompts above) |