AINEWS
Claude Code's hidden system instructions shrank 80%, the 'don't' lists went with them, and the hand review line is the part any working builder can copy.
The Claude Code team at Anthropic shrunk its system prompt, the hidden instruction list shipped with the model, by roughly 80% in a recent rewrite, according to Cat and Thariq in a fireside chat posted Tuesday on Simon Willison's blog. Adding examples to the prompt is no longer best practice for the company's current models, they said, and writing "don't do X" lists now actively makes their outputs worse.
That is an inversion of the prompt-engineering playbook most working developers are still writing to. The inversion comes from the team shipping the model itself, on a fresh enough signal that the change has not yet been absorbed by external guides.
The two moves travel in the same direction. Older models treated examples in a system prompt as calibration: show the model three good responses and a bad one, and it would imitate the shape. The latest Claude models, the team said, do not need that scaffolding. Extra examples read as noise that drags down the response, and a list of prohibitions starts to compete with the user's actual instruction. The current models already know how to apologize less, when to use bullets, and when to hedge, and writing it all down for them starts to cost attention the prompt could spend on the real task.
The Claude Code team has drawn a clean line on what to automate and what to keep human. The critical changes to the product, the parts that touch how the model behaves or how user data flows, still go through manual review. The "outer layers," which the team uses as shorthand for UI shells, peripheral tools, and the connective tissue between the model and the product, increasingly get reviewed by automated agents. The boundary is not "ship fast" versus "ship slow." It is "ship with a human eye on the part that can hurt a user, and let the agents grade their own work where the cost of a miss is low." That line is the most portable piece of the chat for a builder outside Anthropic.
"Ant fooding," the team's term for internal dogfooding, carries a sharper weight than the usual version. Every Claude Code feature ships to Anthropic employees first, and the team only promotes features that show user retention in that internal cohort. The dogfooding group is the de facto product filter: if a feature does not stick inside the company that builds the model, it does not ship to outside customers.
Anthropic's Claude Tag announcement describes the agent as a Slack-resident collaborator that takes issues and lands pull requests, and Cat said in the chat that Claude Tag now lands 65% of the product engineering pull requests for the Claude Code team. The number is high enough that the team's working assumption is that most routine code changes in the product will, at some point, be agent-generated. The 65% is a current snapshot, not a target, and it sits next to a stricter rule: the internal retention check is what decides whether any of those PRs make it into the product.
That is the part a builder outside Anthropic should treat with care. The ant-fooding pattern only works because the dogfooding cohort is also the company that trains the model, and the retention signal is real because the cohort has both technical depth and a high cost of switching to a different internal tool. Most engineering teams cannot reproduce the feedback loop. What they can borrow is the underlying idea: ship to the people who will catch the regression first, and let promotion depend on whether the people who know the product best actually use it.
Auto mode, the team's framing for an agent that picks its own tool and decides its own plan within a stated goal, sits underneath Claude Tag. Anthropic sees auto mode as the enabling layer for the kind of autonomous Slack work Claude Tag does, and the team framed it as a precondition rather than a feature. For a working builder, the take is narrower: an agent that picks tools on its own is what makes a 65% PR rate plausible, and the 80% system-prompt shrinkage is what keeps the agent's instructions from getting in its own way.
Whether the inversion holds outside Anthropic is the open test. The team's 80% prompt shrinkage and 65% agent-PR rate came from one team using its own model on its own product, and the next concrete check is whether the same moves on a different model land the same retention numbers. The prompt to ship next is a shorter one.