Anthropic's Claude runs on 1500-line operational instructions, not just personality tweaks
Original: Spent time today reading through one of Anthropic’s massive Claude prompts after sniffing the network traffic. A 1,500 line prompt!
Source: x.com ↗
Who: Posted by @jmwind (Jean-Michel Lemieux, former Shopify CTO), sharing his own first-hand findings after inspecting network traffic from Anthropic's Claude products.
What's new: Lemieux discovered that 's production for Claude runs to roughly 1,500 lines of text — far longer than industry observers typically assume. He obtained it by sniffing the between a Claude client application and Anthropic's servers.
How it works: The prompt is not primarily personality or tone guidance. Most of its bulk consists of operational logic: full API documentation outlines, code samples, , formatting protocols, and JavaScript snippets with explicit dos and don'ts. Lemieux's framing is precise: this is less a "prompt" in the everyday sense and more a encoded in natural language and embedded directly into the model's context window on every single request.
Why it matters: The finding reframes what means at production scale. If a leading AI company is effectively embedding program logic — branching rules, documentation, code — inside a text prompt, then prompt engineering is converging with software architecture. Developers building on top of models like Claude may need to think of the hidden system prompt as an invisible middleware layer, not a simple personality file.