System Prompts Leaks AI Prompt Engineering Research Pipeline

The asgeirtj/system_prompts_leaks GitHub repository is the definitive public archive of extracted system prompts from virtually every major AI product on the market. As of July 2026, it contains 45+ models across 9 vendor folders: Anthropic (Claude Fable 5, Opus 4.8, Sonnet 5, Claude Code, Claude Design), OpenAI (GPT-5.5 Thinking/Instant/Codex, GPT-5.4, o4-mini, o3), Google (Gemini 3.5 Flash, 3.1 Pro, Gemini CLI, Antigravity CLI), xAI (Grok 4.3, 4.2, 4.1, Grok Build), Microsoft (Copilot, VS Code Agent), Cursor, Perplexity, Meta, and more. Licensed CC0-1.0 (public domain). 55,300+ stars. Trending #1 on GitHub (July 6-11, 2026). Covered by The Washington Post (May 2026) and the AWS Security Blog (July 8, 2026) as a critical resource for understanding AI transparency.

BUSINESS PROBLEM

According to the AWS Security Blog (July 2026), system prompt leakage is a fundamental limitation of current generative AI systems. The OWASP LLM Top 10 lists it as LLM07. A prompt engineer at a 50-person startup building AI agents has no reference for how leading labs structure production-grade prompts — safety rules, tool definitions, refusal templates, formatting constraints. They spend weeks iterating on agent prompts through trial and error. The system_prompts_leaks repo reduces this to hours: study how Anthropic handles refusals in Fable 5 (3,800 lines of XML), how OpenAI structures GPT-5.5's three-tier reasoning effort, and how safety patterns converge across vendors. The repo serves as both a transparency resource and a practical prompt engineering reference library.

WHO BENEFITS

For a prompt engineer at a 100-person AI startup. Situation: You are designing system prompts for a customer support agent. Refusal handling, tone control, and tool-use schemas are all trial and error. Payoff: Study how Anthropic structures Claude Code's refusals, how OpenAI formats GPT-5.5's tool schemas, and how Google structures Gemini's safety instructions. Use real patterns from production prompts. For an AI safety researcher at a university lab. Situation: You need to understand how different vendors constrain model behavior for safety evaluations. Payoff: The repo provides diffs between model versions (Opus 4.8 to Fable 5) showing exactly what safety instructions changed. Run your eval suite against documented constraints. For a developer building AI agent features at a mid-size SaaS company. Situation: You are adding AI features to your product and need to design the system prompt for your agent. Mitigation: The repo's folder structure per vendor is a reference architecture for organizing your own prompt library. Red-team your prompts per leaked patterns.

HOW IT WORKS

Step 1. Clone the repo (2 min). git clone https://github.com/asgeirtj/system_prompts_leaks.git. The repo is 45+ models across 9 vendors. Browse the folder structure to understand what is available. Step 2. Find your model (2 min). Navigate to the vendor folder for the model you care about. Anthropic/ contains Claude Fable 5, Opus 4.8, Sonnet 5, and Claude Code prompts. Each is a plain text Markdown file. Step 3. Study the diff (5 min). The repo provides diff files showing exactly what changed between model versions. For example, the Opus 4.8 to Fable 5 diff shows how Anthropic changed safety instructions, tool definitions, and behavioral constraints. Step 4. Extract patterns (15 min). Identify patterns across vendors: how they format tool-use schemas, how they structure refusal templates, how they define agentic behavior boundaries. Cross-reference these with your own prompts. Step 5. Apply to your prompts (ongoing). Use the patterns to improve your own agent system prompts. Add safety guardrails modeled after production deployments. Structure tool definitions following established conventions. Step 6. Stay updated (5 min/week). The repo is continuously updated. Star it and check weekly for new model captures. Recent additions include Claude Sonnet 5 (July 1), GPT-5.5 Codex (June 18), and Claude Design (June 26).

TOOL INTEGRATION

TOOL: system_prompts_leaks (CC0-1.0, 55K+ stars, trending #1 GitHub July 2026). Role: Public archive of extracted AI system prompts from 45+ models. Access: github.com/asgeirtj/system_prompts_leaks. Auth: None. Cost: Free (public domain). Gotcha: Prompts could be outdated minutes after being committed. Vendors A/B test different prompts for the same model version. A single extraction may be just one variant — do not treat any file as the single ground truth. TOOL: GitHub (platform). Role: Repository hosting and version history. Auth: GitHub account. Cost: Free. Gotcha: The repo is large (45+ models in Markdown). Clone with --depth 1 to save bandwidth if you only need current captures. TOOL: Diff tools (any). Role: Compare prompt versions across models and vendors. Tools: GitHub's built-in diff, git diff, or any Markdown diff tool. Auth: None. Cost: Free. Gotcha: Some prompt files are 3,800+ lines (Fable 5). Use line-filtered diff (git diff --unified=5) to focus on changed sections, not the entire file.

ROI METRICS

Metric Without repo With repo Source Prompt design time per agent Weeks 2-3 days Community estimates Refusal handling iterations Trial and error Study 45+ models Repo scope Safety pattern awareness Single vendor 9 vendors compared Repo vendor coverage Production prompt confidence Low (blind iteration) High (pattern reuse) Community feedback

The week-1 win: pick the model closest to your use case (e.g., Claude Code for coding agents). Read its system prompt file. Note one pattern you can apply to your prompts today. The strategic implication: system prompts are the new documentation. The teams that study how frontier labs structure agent behavior will build more reliable, safer, and more cost-effective AI agents.

CAVEATS

1. (significant risk) No verification guarantee: Prompts may be outdated, partially fabricated, or extracted from specific contexts that do not generalize. Vendors serve different prompts to different users for the same model. Mitigation: Cross-reference findings with official documentation. Treat the repo as one data point, not ground truth.
2. (moderate risk) Legal gray zone: The prompts are intellectual property of their respective vendors, even though the repo is CC0-licensed. Extracted through user interaction, not hacking. Mitigation: Use for research and education. Do not incorporate leaked prompts verbatim into commercial products without legal review.
3. (minor risk) Prompt bloat traps: Modern system prompts are 500-3,800+ lines. Copying this verbatim into your agent will increase latency and token costs. Mitigation: Extract patterns, not entire files. Production prompts should be lean — only the instructions needed for your specific use case.
4. (moderate risk) A/B testing blind spot: Vendors serve different prompts to different users. Any single extraction is one variant. Your own agent may get a different prompt from the same vendor. Mitigation: Build your own extraction workflow to capture the specific prompts your agents receive. Compare against the repo for changes.