Stop Yelling at the Robot

I used to write prompts the way I wrote change-window emails to the NOC: in capital letters, with the word CRITICAL in them. It worked about as well on the model as it did on the NOC.

The difference is that the models got better at listening. Shout at a current one and it overreacts to whatever you shouted about, which Anthropic’s own docs now call overtriggering and tell you to tone down. Most of what I knew about prompting in 2023 has aged the same way. The chart below is eight of those habits, what each one does now, and what to do instead.

Three of them cost me the most. “If in doubt, use the search tool” sounds harmless until a model that takes you at your word finds doubt in every question, including hello. “Do not use markdown” names one thing to avoid and nothing to produce, so you get whatever it picks as the replacement. And prefilling the reply with a brace to force JSON, which a good share of the guides online still recommend, returns a 400 error on Claude 4.6 and later. The guides haven’t noticed.

In fairness, most of the chart comes straight from Anthropic’s docs, but one row is mine: contradictory rules, where “be concise” and “explain every step” fight it out and the model picks a different winner every run. That’s a thing I’ve watched happen, not a thing I’ve measured. And if you’re still pinned to an older model, some of these habits are still harmless there. Check before you unlearn.

The common thread is the one habit under all eight: say what you want, say why, and stop hitting caps lock. The models were never the problem with my change-window emails either.

What’s the worst thing you’ve put in capital letters in a prompt? Mine’s on the chart.