Anthropic launched a new flagship model this month. Buried in the announcement was a warning most businesses running AI haven't heard yet: the instructions that worked fine on last year's model can now make this year's model perform worse.
The new models actively need fewer rules
Anthropic's engineering team just deleted over 80% of the instructions they give their own AI coding tool, and performance didn't just hold steady. In places, it got better. They didn't cut those instructions because the model got more powerful and could afford to ignore extra guidance. They cut them because the extra guidance was actively working against the new model. Anthropic's own writeup lays out the full case: The new rules of context engineering for Claude 5 generation models.
The instructions written for an earlier, less capable model assumed the tool needed to be stopped from doing something bizarre at every turn. That assumption gets baked into dozens of small rules over time. Feed those same defensive rules to a model that's already capable of good judgment, and you get contradictions. "Leave documentation as appropriate" sitting three lines above "never write more than one comment per function," both true once, now just fighting each other and confusing a model that would have handled the judgment call fine on its own.
The writeup names the reason directly: "context is used generally across many requests, so it cannot be as specific" as a one-off instruction can be. A rule written to patch one bad moment gets applied to every future case it was never meant for, whether or not the model still needs it.
Why this happened now, and not before
On July 24, Anthropic launched Claude Opus 5, delivering nearly all the intelligence of the company's top-tier Fable 5 model at half the price. It's now the default model on Claude Max. One developer on X, @TeksCreate, put it bluntly: Anthropic just rendered its own flagship model irrelevant, and did it while cutting the price in half.
The instruction cleanup shipped alongside that launch isn't a coincidence. A more capable model is exactly the point where an old, defensive rulebook stops being harmless and starts costing you accuracy. Their writeup names five specific reversals, and the pattern underneath all five is the same: stop writing instructions for a worker who might do something bizarre if you don't stop them, and start writing instructions for one competent enough to look around and act sensibly.
Picture handing a new cook two different sets of instructions for the same dish. The old version is a 40-step list covering every possible mistake: exact gram counts, exact seconds on the stove, a warning not to answer their phone mid-stir. The new version is a recipe card: here's the dish, here's what it should taste like, use your judgment. A cook who actually knows how to cook does better with the recipe card. The 40-step list just slows them down and, worse, starts contradicting itself once two different people have added warnings to it over the years.
That's what Anthropic did. Instead of listing out every rule for every situation up front, the new setup only hands over detail when a specific task actually calls for it, and trusts the model to ask or figure it out otherwise.
Check what you built for the old model
This applies specifically to Claude. If your team has any instructions, prompts, or skills written for an earlier Claude model, that's the thing worth re-checking now, not the model itself.
If your team runs Claude with its own editable instructions, whether that's a config file, a CLAUDE.md, or a similar setup, open Anthropic's writeup in a new tab and paste the link straight into that setup. Ask it directly:
“read this, then update your own instructions to match how the new models actually want to be talked to.”
It's a concrete, checkable fix, worth trying before you assume your existing setup still fits the model you're now running.
That's The Territory this month. If it's been a while since you looked at the instructions your team's AI tools are actually running on, that link above is the fastest way to find out what's due for an update.
— David