Sign InOpen Brain
GitHubGitHub RepoNeeds Review

asgeirtj/system_prompts_leaks

A maintained archive of 100+ products’ system prompts exposes how coding-agent vendors structure tools, skills, memory, safety, and behavior—useful reference material for harness design.

GitHub · Trending today
Open Source Open MarkdownOpen JSON
Source Summary

The repository maintains extracted system prompts from **100+ AI products**, including Claude Code, Codex, Cursor, Gemini, Copilot, and several research agents. It also archives subagent prompts, skills, tool instructions, memory rules, and reasoning variants.

Practical Implication

Use the captures as implementation references when designing agent harnesses or debugging behavior your own prompt does not explain. Claude Design’s bundle of **53 tools**, **22 skills**, and **10 starter components** is a particularly concrete example of how a product divides instructions from capabilities.

Agent-Ready Context
The repository maintains extracted system prompts from **100+ AI products**, including Claude Code, Codex, Cursor, Gemini, Copilot, and several research agents. It also archives subagent prompts, skills, tool instructions, memory rules, and reasoning variants.

Use the captures as implementation references when designing agent harnesses or debugging behavior your own prompt does not explain. Claude Design’s bundle of **53 tools**, **22 skills**, and **10 starter components** is a particularly concrete example of how a product divides instructions from capabilities.

The archive mixes recent captures with old models and deprecated features, so version and product context matter. Extracted prompts reveal vendor instruction patterns, but the material alone does not establish which patterns improve agent outcomes.
Connected Context · Feed7 Judgment

This supplies a comparative evidence base for how real products partition prompts, tools, skills, memory, and subagents, making unexplained harness behavior easier to investigate. It does not show that copied patterns work: captures must be tied to their product and model version, then tested against an explicit behavioral contract rather than treated as current best practice.

AISPA: User-Centric System Prompt Auditing for Large Language Model ApplicationsAISPA provides a structured way to audit directives found in the archive, turning prompt captures into review inputs while still requiring runtime evaluation.Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, AnthropicAnthropic’s account that one model’s harness fix harmed another directly reinforces the archive’s version-context warning and argues against copying captured scaffolding unchanged.The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest & Isaac MillerDSPy supplies the missing evaluation discipline: extracted prompts can be treated as replaceable implementations while a stable task contract measures whether any borrowed pattern helps.Codex, Behind the Harness — Dominik Kundel, OpenAIThe Codex harness discussion offers documented architectural context for mechanisms visible in prompt captures, while also confirming that those mechanisms remain model- and API-dependent.
Context Map
contextcoding#prompting#context-engineering#harness-engineering
Uncertainty
The archive mixes recent captures with old models and deprecated features, so version and product context matter. Extracted prompts reveal vendor instruction patterns, but the material alone does not establish which patterns improve agent outcomes.