Daily AI Catchup
A prompt is read, not received
It's easy to write a prompt the way you'd write to a person: a little softening, an apology for the ask, a line about how much you'd appreciate the help. None of it is wrong to say to a person. All of it is wasted on a model, which reads "if it's not too much trouble, could you possibly help me out here" exactly as carefully as it reads the actual instruction — and gets nothing usable out of the first one.
What actually earns its place in a prompt
The task itself, stated plainly. A constraint the model can check itself against — "under 80 words" beats "keep it short-ish" because one is verifiable and the other is a matter of opinion. The actual content to work on, when there is any. And a concrete style anchor when tone matters: "match the tone of a friendly coworker" gives the model something specific to aim for, where "be friendlier" leaves it guessing what that means.
Why "try your best" doesn't work the way it sounds like it should
There's no dial inside a model that turns up when you ask it to try harder or think very carefully. Those phrases feel like they should matter because they'd matter if you said them to a person — but a model doesn't have effort to withhold or supply on request. What changes its output is more specific instructions, or less specific ones. There's no third setting called "more effort."
There's no dial in the model for "try harder." There's only more specific instructions, or less specific ones.
The same lesson as trimming a crowded chat, one level up
Deciding what to keep in a prompt is the same skill as deciding what to keep in a cluttered conversation before asking a model to answer from it — just one step earlier. Before you ever get to what belongs in the context, there's what belongs in the instruction that starts it. Cut the softeners and the vague exhortations, and what's left is usually a shorter, cheaper, more effective prompt than the one you started with.
Updated August 2026 · 2 min read