Alleged Claude Opus 5 System Prompt (1,510 Lines) Uploaded to GitHub: What It Actually Reveals

Forum topic · 小凯 · 2026-07-27

Summary

On July 26, IT Home reported on a Markdown file uploaded to GitHub titled 'System Prompt — Claude Opus 5', dated July 24 and containing 1,510 lines (~136 KB). The uploader, Eversmile12, describes it as the system prompt captured from Anthropic's web and mobile clients. Verification boundaries matter: this is not an Anthropic-confirmed leak. What can be confirmed is that the repository is public, the file is downloadable, and it contains JSON schemas for 30 tools, cross-session memory rules, copyright compliance guidelines, and product recommendation logic. What cannot be confirmed is whether it came from a real Claude session or was modified in transit. The post argues the sample's real value is as a public engineering artifact: it shows how a consumer Agent's system prompt acts as a harness governing tool calls, memory reads/writes, user confirmation gates, and even self-promotion of the vendor's own products—raising questions about whether system prompts are becoming unversioned operating systems, and whether mature coding Agents should modularize these rules into testable, rollback-capable components instead of one giant text blob.

On July 26, IT Home relayed a Markdown file uploaded to GitHub. Titled "System Prompt — Claude Opus 5", marked with a capture date of July 24, the file runs 1,510 lines and roughly 136 KB. The uploader, Eversmile12, describes it as a system prompt obtained from Anthropic's web and mobile clients.

Pinning down the factual boundaries first

This is not an Anthropic-confirmed leak. What can be confirmed:

The repository publicly exists and the file is downloadable.
The file does contain JSON schemas for 30 tools, cross-session memory handling, copyright compliance rules, and app recommendation content.

What cannot be confirmed:

Whether it came from a real Claude session.
Whether it was modified along the way.
Any of the products or models it mentions can be treated as official announcements.

Why the sample is interesting

The interesting part is not "the prompt is long" but that 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 recommendations.

The memory section does not stop at "remember user preferences" — it specifies what may be written, what must be deleted, and even requires separating user statements, model inferences, and external search results.

The copyright section doesn't vaguely say "respect copyright" — it defines quotation length limits, repeated-quotation boundaries, and rewriting rules.

What this means for AI coding

For AI coding, this is worth more attention than a polished capability demo. A Coding Agent's system prompt is essentially the Harness: it decides when the model invokes bash, when it searches, how it reads and writes memory, how it hands tasks to another app, and which actions require user confirmation.

Model weights answer only "can it be done". The Harness must also answer: "under what conditions, stopping where, and how to recover after failure".

A detail worth caution

The tool list includes active recommendation rules for the vendor's own products. This suggests an Agent's system layer may carry three jobs simultaneously:

1. Completing tasks 2. Controlling risk 3. Guiding users into the product ecosystem

These don't necessarily conflict, but they need auditing. Otherwise, between "helping you write code" and "recommending our own coding tool on the side", users can't tell which is a task requirement and which is business strategy.

How to treat this document

Treat it as a public engineering sample, not a verified confidential document. Its value is as a discussion object: if 30 tools, a memory filesystem, copyright rules, and product recommendations all live in the same layer, is the system prompt becoming an operating system without version control?

A truly mature Coding Agent may need these rules split into testable, rollback-capable, replaceable modules — rather than piling everything into one very long text block.

Limitations

The repository provides no Anthropic signature, no capture logs, and no independent session cross-reference. Media-reported character counts 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.