The First Question Isn't How to Prompt It — It's Whether This Should Go to an Agent at All

Work shows up, you open the chat. That's the first move, almost thoughtless — like reaching for your phone when your hands are free. Then the rest of your mind pours into one question: how do I phrase this so the agent gets it? You tune the wording, add context, hand it an example.

But you skipped a fork before that one — a far more important fork: does this belong to an agent? This whole garden, every piece, assumes you're already standing at "agent, chosen." This one steps back behind that assumption, to where you actually should stand first.

01Two axes decide it, not one

The two pieces beside this one each light up a single axis. One asks: does this job have one determinate right answer? — if so, it belongs to a command line, not a brain. The other asks: does the saving beat the overhead of opening a chat each time? Both are right, but each is only one slice.

Put together they get much sharper. Set the two axes at right angles — one axis is the nature of the work (one right answer, or needs judgment), the other is frequency (done once, or repeated) — and you get a plane of four cells. And here's the surprise: the agent doesn't cover all four. It clearly wins in exactly one. The other three have a cheaper tool, a surer tool, or just your own hands being faster.

Notice: only one of the four cells carries a ✓. The left column — deterministic work — is never the agent's, once or repeated; a machine does it exactly and for free. The right column is the land of judgment, and even there, "once" still leans toward doing it yourself. The agent is most valuable in the bottom-right corner: judgment × repeated.

02Why each cell has its own tool

This frame isn't a tidy taxonomy — it mirrors where each tool is genuinely strong. Deterministic work has one right answer; anything with a right answer, a machine produces surely and reproducibly, while an agent merely guesses at what was computable — paying tokens for a less trustworthy reply. That's why the whole left column leaves the agent.

The frequency axis decides whether you should invest in setup. A deterministic job done once: type the command, done. The same job done ten times: wrap it in a script or a button so you never re-think it — setup paid once, returns forever. Same on the judgment side: a single run lets the open-the-chat-then-verify overhead swallow the gain, but repeat it enough and that overhead amortizes, and the part the agent carries — judgment that won't reduce to a rule — turns out to be the thing no other tool can stand in for. Four cells, four different economic reasons, not four labels.

03Three screening questions to know which cell you're in

You don't have to draw the matrix each time. Three questions, in order, land you in the right cell within seconds — and they're worth asking before your hand touches the chat box:

These three ask about the nature of the work, not the overhead — unlike the "how long to do it yourself / does the agent have an edge / can I verify it" trio next door. That set weighs the profit-and-loss of a single run; this one locates which cell of the map you're standing in.

There's a quiet payoff to screening from the start: most of the waste in working with agents isn't a weak prompt — it's opening the chat for something that never needed a chat. You pay a brain's overhead for a one-line job, or grind by hand through a repetitive job a script could carry. Get the tool choice wrong and every cleverness afterward is just decoration on a losing pick.

So next time, before you reach for the prompt, pause a beat and ask which cell you're in. The agent is a wonderful tool for exactly one corner of the map. Your job is to know when you're standing there — and when to close the tab, open the terminal, or simply do it yourself.