Anthropic Cut Claude Code's Instructions by 80 Percent. Builders Are Writing Them Back In.

A model’s system prompt is the standing instructions it gets before you ever type a word, the house rules that shape how it behaves by default. Anthropic quietly rewrote Claude Code’s, and one product coach says that change is why the same model can feel smarter and harder to work with at once.

Paweł Huryn, in a post on his newsletter The Product Compass, says he measured the difference directly:

“Anthropic removed 80% of Claude Code’s system prompt and was proud of it. I measured the cut: 2,686 words down to 514.”

Who Huryn is

Huryn is the founder of The Product Compass, a product management newsletter with more than 100,000 subscribers. He describes himself as an ex-Chief Product Officer with 15 years in product roles, and now writes and teaches about how AI changes product work. He is not affiliated with Anthropic.

What he gets right, and where it’s incomplete

Huryn’s measurement is specific and checkable in principle: a system prompt cut from 2,686 words to 514 is a real, sizable change to the instructions a model runs on before it does anything else. He connects that cut to a pattern he noticed in daily use: Opus 5 acts more independently and asks for confirmation less often than earlier Claude Code versions did. His explanation is that Anthropic trained the model to complete long jobs without checking in, using reinforcement learning (a training method that rewards a model for outcomes rather than for following fixed steps), which he summarizes as “it was trained to work without you, so it stopped working with you.”

That is Huryn’s own theory connecting a measured prompt change to an observed behavior shift. He does not have access to Anthropic’s training process, so the causal link between the two is his interpretation, not a confirmed engineering fact. He also cites reactions he says he has seen on his social media timeline, including calls to “bring back 4.6,” as evidence of wider frustration. Those are secondhand and anecdotal, not a survey or Anthropic’s own data, so they show that some users are unhappy without establishing how common that reaction is. Huryn also cites an Anthropic evaluation table from July 24, 2026, saying Opus 5 scored about three times the next-best model on ARC-AGI-3, a benchmark for solving problems the model hasn’t seen before. That figure comes from Anthropic’s own published results, not Huryn’s independent testing.

What he adds beyond the complaint is a fix. Rather than switching models, he wrote explicit rules back into his own custom instructions to restore the behavior he wanted. Two of the three rules he shares publicly (five more are behind his newsletter’s paywall) are concrete and specific: “Act. Don’t Ask,” which tells the model to complete reversible tasks without pausing for permission, and “A Question Is a Question,” which stops the model from acting on a suggestion when he only asked an evaluative question like “how could we improve this?”

Why it’s notable

This lands the same week Claude Code is set to make auto mode, which lets the model run commands without asking first, the default permission setting for most users starting August 14. Huryn’s account of Opus 5 acting first and asking less is the same shift in a different form: Anthropic has been steadily moving Claude Code’s defaults toward acting independently rather than checking in. Whether that trade favors the builder or the model’s convenience is exactly what Huryn is arguing about.

What it means for builders

Anthropic’s default instructions are tuned for Anthropic’s idea of a typical user, not for your workflow. If Opus 5 (or any model update) suddenly acts more independently than you expect, don’t assume you have to accept the new behavior or switch models. Write down the specific failure, the exact way it went wrong, and turn it into an explicit instruction in your own system prompt or custom instructions.

Huryn’s approach is worth copying even if you never read his full rule list: treat every repeated frustration with a coding agent as a missing instruction, not a fixed personality trait of the model. The fix is usually smaller and faster to write than switching tools or waiting for the vendor to change it back.

End of article