Sonnet 5 turned 3 old prompt tricks into one 400 error.

Sonnet 5 turned 3 old prompt tricks into one 400 error.

If you use Claude like a chat buddy and coding helper, this is the part that can actually hurt. I came in expecting a stronger model and left realizing the real risk was my own vague wording.

Under all the Introducing Claude Sonnet 5 hype, the shift is simple: it reads you more literally. The docs say that when you keep its thinking level low, it will not fill in the blanks for you, so fuzzy requests can go off target fast.[S002]

⚠️ Before, I could get away with 1 messy sentence plus 3 little control dials. Now, changing those 3 old dials away from their normal setting can trigger a 400 error, basically the app saying no.[S003]

Ngl, that is the part people will argue about: not that it got stronger, but that the boundaries got tighter first. This model does not reward mystery. It rewards clear boxes: what you want, what shape you want it in, and what it should not do.

📌 I only checked this in Claude's direct setup screen and the notes for switching older setups to Sonnet 5, not inside a huge real-world app, so your setup may differ. Save this for your next setup, or share it with the friend who still thinks the model should just magically get it?