Anthropic’s release notes page now exposes the system prompts for Claude’s main chat experiences on the web, mobile, and desktop apps. These are the top-priority instructions prepended before a user’s message, and they now run to thousands of words. They cover basic identity and date handling, response style, crisis behavior, refusal policy, product facts about model tiers, and even explanations for cases where a user may have been silently routed from one model to another for safety reasons. That made the page interesting less as prompt craft and more as a plain look at the policy and product logic Anthropic is shipping in natural language.
The strongest reaction was that the prompts have become bloated. Early versions were a few hundred words. Newer ones are an order of magnitude larger, and several people argued that this kind of accretion usually means legal, safety, and edge-case requirements are being stacked into one giant document. The practical complaint was not just aesthetics. People think long, internally inconsistent prompts consume scarce context, dilute task-relevant instructions, and can make coding performance worse even when the added text is unrelated to the job at hand. A recurring example was Claude’s claimed preference for brief answers versus widespread user experience of verbose code comments,
PR descriptions, and explanations. That mismatch made the prompts look like policy aspirations more than reliable controls.
A second theme was what these prompts reveal about where model behavior really lives. Many commenters took the published text as evidence that labs still rely heavily on in-context steering for things that marketing often frames as intelligence or alignment. The prompts include post-training facts like election results, export-control notices, and model-
routing explanations precisely because those details changed after training and need a fast patch. Others pointed out that keeping behavior in prompts and external
guardrails is not just easier to update. It also lets Anthropic vary behavior across products, customers, and regulated use cases without retraining separate models. That framing landed more convincingly than the simpler complaint that Anthropic should just “bake it in.”
The comments also surfaced a clearer picture of hidden machinery around the prompts. Several people said the published prompts are only part of the story because Anthropic did not include
tool definitions, and those often matter more for understanding what Claude can actually do. Others noted that safety filtering for disallowed content almost certainly sits outside the prompt in cheaper classifiers and routing systems, with the prompt serving as defense in depth and tone control rather than the main lock on the door. That matters because it means prompt leaks tell you a lot about product behavior, but not the full control stack.
The most grounded take on the whole page was that these giant prompts are less a sign of elegant prompt engineering than of product governance. They read like living policy documents attached to a general model. That helps explain why they are messy, why they contain what looks like boilerplate, and why they keep growing even if that growth may degrade the thing power users care about most, which is clean task execution.
platform.claude.com