Photo by www.advancedwebranking.com.

If you use LLMs regularly, you may have noticed something interesting.

Compare these two instructions: “Do not change the format.” and “DO NOT CHANGE THE FORMAT.”

They mean essentially the same thing, but an LLM may not process them in the same way.

Recent research suggests that letter casing can influence how models allocate attention. A 2026 study called “Attention is Case-Sensitive” tested multiple language and vision-language models and found that uppercase text placed inside lowercase context could attract more internal attention toward that part of the prompt.

In other words, capitalization can act as a kind of salience signal.

This may help explain why prompts containing words such as “MUST,” “ONLY,” “IMPORTANT,” or “DO NOT” sometimes feel more effective when those instructions are capitalized.

But there is an important catch.

More attention does not automatically mean a better answer.

The same research found that increasing attention through capitalization did not consistently improve task accuracy. In some cases, unusual casing could even hurt performance. Reasoning models also appeared to be less sensitive to this effect during their internal reasoning process.

This connects to a broader issue in prompt engineering: LLMs can be surprisingly sensitive to formatting. Changes in capitalization, punctuation, wording, spacing, or prompt structure can sometimes produce noticeably different outputs even when the underlying meaning stays almost the same.

So I would not recommend WRITING YOUR ENTIRE PROMPT LIKE THIS.

But strategically emphasizing a critical instruction may affect how a model interprets and prioritizes your prompt.

For me, the interesting lesson is bigger than capitalization itself.

When we communicate with an LLM, we communicate through more than meaning. Formatting is part of the input, too.

And sometimes, typography becomes part of the prompt.