Stop Letting AI Write Your Prompts

More precisely, stop shipping AI-generated context without compression

2026-04-18

Stop Letting AI Write Your Prompts

An AI read my USPS tracking number and confidently reported it as a nine-sextillion-digit number.

I saw a package.

Context matters. And the way most teams add context is making things worse.

Stop letting AI write your prompts. More precisely, stop shipping AI-generated context without compression.

A recent paper on agent design (Gloaguen et al., arXiv:2602.11988) shows a counterintuitive pattern: more tokens create more constraints, and more constraints often lead to worse performance.

The trap is subtle. AI-generated context looks high quality. It is clean, structured, and comprehensive. But most of it is repackaged information the model already knows. It doesn't add signal. It adds noise, delivered with confidence.

High-performing teams operate differently. They don't ask, "What else should we add?" They ask, "What can we remove?"

Most prompts don't fail because they lack context. They fail because they have too much of the wrong kind.

Longer prompts don't necessarily lead to better outcomes.

Tighter context wins.


First published on LinkedIn, 18 April 2026. Lightly revised. Original post.