A new model launch is not a product update ‼ It's one of three things, and you don't know which until you test it. Sometimes it's nothing: t…

A new model launch is not a product update ‼ It's one of three things, and you don't know which until you test it. Sometimes it's nothing: the model improved inside the same distribution, your harness still holds, ship as-is. Sometimes it's a silent behavior change rippling through every product that depends on that model, whether you touched anything or not. And sometimes the jump is big enough that the whole distribution shifts, and the prompts, guardrails, and scaffolding you built for the old one are quietly wrong now. In that case, even holding your current behavior steady, or improving it, takes real changes to how you harness the model. Occasionally it's not maintenance at all, it's a whole new opportunity that wasn't possible last week. The Anthropic team framed a lot of this as removing constraints, and I find that true sometimes and not others. Yes, strong models often do better when you stop boxing them in with rules written for a weaker generation. But some constraints are not hobbling, they are your taste: the rubric for what good looks like for your use-case / framework / workflows. The work isn't "delete your rules," it's re-fitting the harness to the new distribution, and that means leaning more on references, rubrics, and verification than on walls of instructions. Which is why the only honest way to greet a launch is to run your own eval on your own workflow. That tells you which of the three you're in, and it's also where the leverage is. The expert who can define correct and let the model check itself against it gets a real multiplier. Everyone else just gets a new model. We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them.

Source: Itamar Friedman (X) | 2026-07-24

Related