Emphasis in Prompts Is Attention Engineering

People treat prompt emphasis like folk medicine. Bold this. CAPS that. XML tags. It works. Most people stop there.

It works because emphasis changes your tokens. Not in the way it signals importance to a human reader. It changes the token representation in ways that make those tokens more retrievable during the attention lookup.

The Mechanism

Every token in context gets projected into three vectors: Query, Key, and Value. When the model generates the next token, its Query computes similarity against every Key. High similarity means high attention weight.

When you write **critical instruction**, the model sees tokens: **, critical, instruction, **. Those delimiter tokens are unusual. They don’t appear often in normal prose, but they appear consistently in training data co-located with content humans marked as important.

CAPS works the same way. IMPORTANT is a different token than important. It’s rarer in the training distribution, co-occurs with human-flagged urgency, and creates a distinct Key vector.

Three Things Happen

Token distinctiveness. The ** markers create a token pattern that stands out. The Key vectors have a different signature than ordinary text.

Attention head specialization. Some transformer attention heads specialize during training. Some track syntax. Some track instruction boundaries and formatting. Emphasis markers activate the heads that learned to attend to human-flagged importance.

Retrieval boost. When the model generates output and its Query looks for “what was I told to do,” emphasized tokens surface more reliably. The markers act as index anchors.

The Caveat

Emphasize everything and you’ve emphasized nothing. The effect depends on the markers standing out from surrounding context. If your whole prompt is bolded, you’ve changed the baseline, not created emphasis.

Use it sparingly. Put critical instructions at the start and end. Emphasize non-negotiable constraints. Break instructions into independently retrievable units rather than burying them in paragraphs.