David K Piano says tell agents what you want, not what to do

David K Piano posted a single line about how he prompts agents. Telling your agents what you want beats telling your agents what to do. He hedged it himself up front, calling it either a hot take or painfully obvious.

David K 🎹

@DavidKPiano

X

Telling your agents what you want >>>>>> telling your agents what to do

Kent C. Dodds replied with his own framing of the same idea, that you should be talking with your agent rather than at it.

Kent C. Dodds 🐨

@kentcdodds

X

You should be talking with your agent, not at your agent.

What is actually being argued

This is the old specification versus instructions split, moved into agent prompting. The what-to-do style is a list of steps, edit this file, add this function, run this command, and the agent executes them in order. The what-you-want style states the outcome and leaves the route to the model, which is the whole premise of handing work to something that can plan.

Neither post comes with a codebase, a benchmark or a before-and-after. It is two experienced developers stating a preference, not a result, and the style guide for believing things applies here. One line on X is one line on X.

The caveat that neither post mentions is that outcome prompting only pays if you can state the outcome precisely enough that you or the agent can check it afterwards. A vague want is just a vague instruction with fewer words, and you find out at review time. That is the same ground Matt Pocock was working when he made his prompt of the day how do we make this change fully verifiable, which is the question that turns a want into something a model can actually aim at.

For working engineers, the practical read is that the two styles fail differently. Step-by-step prompts fail by doing exactly the wrong thing you asked for. Outcome prompts fail by doing something plausible that you never specified. Which failure you prefer depends on how well you can describe the finished state before the work starts.