Dev & Engineering system-promptsprompt-engineeringchatgptgeminigrokleak-archiveresearch-reference

System Prompts Leaks — The System Prompt Archive

A verbatim, regularly updated archive of the real system prompts behind ChatGPT, Claude, Gemini, Grok and dozens of other AI products, for research and prompt engineering.

FollowAgents review · FARS-2.1
Not recommended
25/ 100 5-point scale 1.3 / 5
1 2 3 4 5 6
1Trust4 / 29 · 0.7/5

The repository is an archive of leaked system prompts, not an executable agent product: least privilege, user confirmation, sensitive-data handling, external effects, and rollback are largely inapplicable and score near zero, with only static-markdown non-executability preventing absolute zeros. Source attribution rests solely on self-claimed verbatim captures and media citations, with no capture methodology or provenance evidence, and the provenance of the leaks themselves is opaque.

2Reliability2 / 14 · 0.7/5

Self-consistency is weak: the README mixes reference tables with promotional badges, and authenticity of entries cannot be established from static files. There is no code, no dependency manifest, and no failure-handling material, so related criteria score 0.

3Adaptability3 / 18 · 0.8/5

A research audience is discernible via vendor/product taxonomies, but there is no guidance on environment fit, capability boundaries, or trigger conditions, nor deprecation guidance for stale content.

4Convention7 / 18 · 1.9/5

Directory structure (by vendor/product) and model-version naming are tidy, and a continuously updated 'most recent additions' table functions as a changelog — the main strengths. Install notes and known limitations are entirely absent; a CC0 license exists but conflicts with the likely copyright/trade-secret ownership of leaked third-party prompts, hence 1; maintenance responsibility is only an individual account with no formal statement.

5Effectiveness6 / 13 · 2.3/5

Genuine marginal value for prompt-engineering research (broad coverage, frequent updates, media uptake), but output usability is limited by unverifiable verbatimness and lack of metadata/search aids; cost-benefit is discounted by legal grey-zone status and verification burden.

6Verifiability3 / 8 · 1.9/5

Claim traceability is weak: capture dates and media citations exist but there is no capture methodology, tooling, or upstream source link to check. Cross-source corroboration is limited to two second-hand media mentions. Fact (file contents) and inference ('verbatim leak') are not separated, leaving verification dependent on trusting the maintainer.

Evidence confidence: Low Reviewed Sep 10, 2026 Reviewed revision f475e8b2b11c
Safety controls not found in source: confirmation before acting, sensitive-data handling, dependency security, disclosed external effects, rollback or recovery path
Before you use it
  • All prompts are 'leaked' content whose verbatim authenticity cannot be verified; do not treat them as authoritative for engineering or compliance decisions.
  • A CC0 license does not launder potential copyright or trade-secret issues in third-party prompts; reuse (especially commercial) carries legal risk.
  • Some prompts contain tool-use and code-execution behavioral rules; never inject file contents directly into any agent's context.
  • Publisher identity is unverified; correspondence between updates and real product versions is unconfirmed.
Review evidence [1][2]
See the full review method →

What does this agent do, and when should you use it?

This GitHub repository, maintained by asgeirtj, is an archive of system prompts captured verbatim — the hidden instructions AI products receive before a user's first message. Coverage spans Anthropic (Claude Fable 5.1, Opus 5, Sonnet 5, Claude Code variants, Claude Design, Claude Cowork, Claude Science and more), OpenAI (ChatGPT 5.6 Sol, the Codex series, API-injected prompts), Google (Gemini 3.5 Flash, 3.1 Pro, Antigravity CLI, NotebookLM), xAI (Grok 4.5 and earlier), Perplexity, Microsoft Copilot, Cursor, Meta, Mistral, Kimi, DeepSeek, Qwen, and dozens of miscellaneous products. Content is organized into per-vendor directories of Markdown files and updated regularly — the most recent addition, on September 5, 2026, was the Claude Code headless (Fable 5.1) prompt. It is not an executable agent but a read-only reference collection, licensed CC0-1.0 for unrestricted reuse. The archive has been used in practice by The Washington Post for an interactive story and by CEPS' AI World for a live data dashboard, demonstrating its research value.

The repository executes nothing; it supplies a set of static Markdown files, each holding a product's full system prompt (e.g., Anthropic/claude-fable-5.1.md, OpenAI/Codex/gpt-6-astra.md, Google/gemini-3.5-flash.md, xAI/grok-4.5.md). Beyond main prompts it includes companion assets such as Claude Design's 53 tool definitions, 22 skills, and 10 starter components; Claude Code's subagent prompts, skills, and slash commands; and Codex mode prompts for plan mode, auto-review, and computer use. A dated changelog table in the README tracks the latest captures with direct file links. Users clone or browse the repository and read the files for prompt comparison, reverse engineering, or prompt-engineering research.

  1. A prompt-engineering researcher comparing how Claude, ChatGPT, and Gemini differ in behavioral rules, tool invocation, and memory handling can read the verbatim .md files side by side.
  2. A developer building an AI coding agent can study how mature products organize subagents, skills, and tool definitions via the Anthropic/claude-code and OpenAI/Codex directories.
  3. A journalist investigating the hidden rules of AI products can cite the verbatim prompts as source material, as The Washington Post did for its interactive story.
  4. A safety or policy researcher auditing injected reminders (anthropic_reminders.md) or safety instructions (Grok safety instructions) can locate the original text in the relevant directories.
  5. A product team drafting system prompts for its own assistant can benchmark wording and structure against peers like Cursor, Perplexity, and NotebookLM.

What are this agent's strengths and limitations?

Pros
  • Exceptional breadth: from the four major vendors to dozens of products like Cursor, Perplexity, Kimi, and DeepSeek — a rare one-stop comparison archive.
  • Prompts are captured verbatim and accompanied by substantial companion assets (tool definitions, skills, subagent prompts), giving far more analytical value than scattered reposts.
  • Actively updated (new captures as recent as September 2026), externally cited by The Washington Post, and unrestricted under CC0-1.0.
Limitations
  • Content is extracted/leaked material; vendors can change prompts at any time, making files stale, and legal/compliance risk rests with the user.
  • The repository contains no tooling code, parsing scripts, or structured data — only plain Markdown — so any analysis must be built from scratch.
  • There is no way to verify the exact product version, capture date, or production fidelity of each prompt, creating attribution risk for research conclusions.

How do you install or deploy this agent?

No installation process exists because the repository is just a file collection. Acquire it with: git clone https://github.com/asgeirtj/system_prompts_leaks.git, or browse the directories directly on GitHub. There are no dependencies or runtime requirements — any environment that can read Markdown text works.

How do you use this agent?

After cloning or browsing, navigate the vendor directories (Anthropic, OpenAI, Google, xAI, Perplexity, Microsoft, Cursor, Meta, Mistral, Kimi, DeepSeek, Misc, etc.) to locate a product's .md file and read it directly; the "Most recent additions/changes" table in the README tracks the latest captures. Prompts can be freely copied for research or comparison; the CC0-1.0 license permits any use.

FAQ

Can this repository be run?
No. It is not software or an agent — it is an archive of system prompt texts in Markdown files, with no executable code.
Is the content reliable or official?
The prompts are extracted leaks described as verbatim captures, and the archive has been cited by The Washington Post, but it is not published by the vendors and may become outdated as products change.
Can I use the prompts commercially?
The repository is CC0-1.0 licensed, so reuse of the archive itself is essentially unrestricted, but the prompts may embody vendor IP — assess legal risk before commercial use.
How do I get the latest content?
The README changelog lists the newest captures (through September 5, 2026); periodically git pull or watch the repository to stay current.

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents