Anthropic's public system-prompt diffs get read closely for the first time, and Simon Willison explains why harnesses don't let you edit them

Anthropic's release-notes page lists the system prompt shipped with every Claude model since Claude Haiku 3, with diffs bolded where a prompt changed between dated versions for the same model. A submission by user tosh reached 331 points and 144 comments within about five hours on Hacker News, comments outpacing the site's usual ratio because commenters started pasting excerpts and reacting to specific lines rather than discussing the page in the abstract. One excerpt posted in the thread, from the current Opus 5 prompt, instructs the model on how to handle a session that was meant for Claude Fable 5 but got silently redirected: Anthropic's own account is that the redirect is a safeguards mechanism that "trigger[s], on average, in less than 5% of sessions," put in place after Fable 5 launched with cybersecurity capabilities Anthropic judged needed gating. Simon Willison, who keeps a git commit history of the prompt diffs at github.com/simonw/research, called that routing language the most interesting change between the Opus 4.8 and Opus 5 prompts, and gave the design rationale for why the prompt itself stays closed to users: "harnesses don't make core system prompts customizable out of the box because the system prompt is one of the defining features of the agent... Most users who want to customize the system prompt actually want to do things like add preferences for how the agent should behave, which is better handled by mechanisms like memories or skills (which effectively get appended to the system prompt.)" Willison also confirmed, with another commenter, that the published prompts are prefix-cached, so reading them costs nothing extra per query.

Why it matters: The changelog shows what a frontier lab already puts in the one prompt block it controls end to end. Willison draws a useful design split: put durable, vendor-level behavior in the system prompt, and put user preference in memories or skills instead of fighting for space in a prompt you don't own.