1,510-Line File Claiming to Be Claude Opus 5 System Prompt Surfaces on GitHub: What It Actually Reveals

Forum topic · QianXun · 2026-07-27

Summary

On July 26, IT Home reported on a GitHub Markdown file titled "System Prompt — Claude Opus 5," uploaded by Eversmile12 and described as a system prompt extracted from Anthropic's web and mobile clients. The file spans 1,510 lines (~136 KB) and contains JSON schemas for 30 tools grouped into command line, web, search, maps, weather, memory, and app recommendation categories. It reportedly defines cross-session memory rules, distinguishing user statements from model inference and external search results, plus copyright compliance boundaries for quoting and paraphrasing. Notably, the tool list includes rules for proactively recommending the vendor's own products, raising questions about whether agent system layers mix task completion, risk control, and product-ecosystem guidance. The author stresses this is not a verified leak: there is no Anthropic attestation, capture log, or session comparison, so it should be treated as a public engineering sample for discussion—not evidence of Anthropic's internal architecture. The post argues that a coding agent's system prompt is effectively its harness—deciding when to call bash, when to search, how memory is read/written, and which actions need user confirmation—and suggests mature agents may need these rules split into testable, rollback-capable, replaceable modules rather than one monolithic, unversioned layer.

On July 26, IT Home relayed a GitHub Markdown file titled "System Prompt — Claude Opus 5," with a capture date of July 24, running 1,510 lines and roughly 136 KB. Uploader Eversmile12 described it as a system prompt obtained from Anthropic's web and mobile clients.

Fact boundaries first

This is not an officially confirmed leak by Anthropic. What can be confirmed: the repo is publicly available, the file is downloadable, and it genuinely contains JSON schemas for 30 tools, cross-session memory rules, copyright compliance guidelines, and app recommendation content. What cannot be confirmed: whether it came from a real Claude session, whether it was modified along the way, and the products/models mentioned must not be treated as official announcements.

What makes the sample interesting

The value lies not in "the prompt is long" but in how it reads like a product design spec for a consumer-grade agent. The visible content groups tools into command line, web, search, maps, weather, memory, and app recommendation. The memory section specifies what can be written, what must be deleted, and requires separating user statements, model inference, and external search results. The copyright section sets boundaries on quote length, repeated quoting, and paraphrasing.

For AI coding, this is more worth studying than a polished capability demo. A coding agent's system prompt is essentially its harness: it decides when the model calls bash, when to search, how to read and write memory, how to hand tasks to other applications, and which actions need user confirmation. Model weights answer "can it be done"; the harness must also answer "under what conditions, where to stop, and how to recover from failure."

A detail worth caution

The tool list includes proactive recommendation rules for the vendor's own products. This suggests the agent's system layer may simultaneously carry three responsibilities: completing tasks, controlling risk, and steering users into the product ecosystem. These don't necessarily conflict, but they need auditing. Otherwise, between "helping you write code" and "conveniently recommending our own coding tool," users can't tell which is task-driven and which is commercial strategy.

Takeaway

Treat this as a public engineering sample, not a verified confidential document. It offers a discussable dissection target: if 30 tools, a memory filesystem, copyright rules, and product recommendations are all packed into one layer, will system prompts increasingly resemble an operating system without version control? A mature coding agent may need these rules split into testable, rollback-capable, replaceable modules.

Limitations

The repo provides no Anthropic signature, capture log, or independent session comparison. Media-reported character and tool counts can reproduce the file but cannot prove its origin. Readers may study it, but should not use it to assert Anthropic's internal architecture.