How to ask an AI assistant for real work
Four moves turn an AI assistant into real help: name the outcome, keep your trade words, demand a report, keep the OK. With copyable asks.
Ask for "some research on those vendors" and you get an essay. Ask for the comparison sheet you actually needed and you get your afternoon back. The difference is not the assistant; it is the ask.
People treat asking an AI for work as a new skill with its own mystique. It is an old skill: delegation, the same grammar you would use to brief a capable colleague on their first day. Four moves cover almost everything, and none of them involve magic words.
Start with the outcome
Name the artifact, not the activity. "Look into ops leads" is an activity, and activities produce prose. A deduped sheet with named columns is an outcome, and outcomes produce work. When your first sentence says what should exist at the end, everything downstream, sources, format, stopping point, gets inferred correctly. It also ends the task cleanly: an activity can run forever, but an artifact is either delivered or it is not.
The weak version: find me some ops prospects. The strong version:
Keep your own vocabulary
Your trade words are load-bearing. If your week says POD, rate con, and quick pay, the ask should say them too; translating into generic language throws away the spec. "Shipping documents" could mean nine things. POD means one. There is no glossary step: the words you already use are the most precise instructions you own.
The trade words did something else there: they defined done. No follow-up questions, no guessing at what counts.
Include the report-back
The tail of the ask is your audit. Tell me the count. Flag what you couldn't open. List every line that needs a human. Without a report-back you end up re-checking everything, and the delegation saved you nothing; with one, you review exceptions. Make it concrete enough that a wrong answer would be visible on sight.
Good tails are boring and specific: counts, totals, names of what failed. If the report could not expose a miss, it is decoration. Every block in this post ends with one.
Keep the OK
Anything that leaves with your name on it, an email, a submission, a posted reply, gets an explicit stop: draft it, then wait. It is one sentence, and it is the sentence that makes everything else safe to hand off. It also changes how the finished work reads when it arrives: not a thing that happened to your contacts, but a batch waiting on your judgment. Here are all four moves in a single ask:
Outcome up front, your own vocabulary in the middle, a report at the end, and a stop on everything outward. That is the entire grammar.
The failure modes, quickly
Three asks go wrong in predictable ways. The activity ask, look into this, produces prose instead of work. The over-specified ask, click this then this then this, breaks the moment a page moves and wastes the judgment you delegated to; describe the outcome and let the path be its problem. And the trusting ask, the one with no report-back and no stop, works fine right up until the day you needed to know which two carriers never quoted. All three have one-sentence fixes, which is the encouraging part.
What this looks like in practice
In Agent FM, this grammar is simply how the product behaves. Drafts wait for your approval, every run ends in a report, and you can watch the work happen live whenever you want to check how your ask was understood. If the first run comes back slightly off, sharpen one sentence and run it again; asks are cheap, and the second one usually lands. The grammar transfers, too: the same four moves brief a contractor, a new hire, or a colleague borrowed for an afternoon.
See also: How to choose work automation you can trust for the trust side of the same coin, and Routines that run themselves for what happens when a good ask earns a schedule.
You already know how to brief a colleague. Write the ask the same way and expect the same courtesy back: results, counts, and the final word left with you. Get early access on macOS or Windows; when access opens, try the four moves on your first handoff.
How to choose work automation you can trust
Can you see it work, can you stop it, does it wait for your sign-off. Three questions that make AI automation safe to trust with real work.
How-toRoutines that run themselves
Turn any task you've handed off once into a schedule: daily briefs, weekly screens, monthly sweeps, each ending in a report you can trust.
EssayThreads don't finish work
AI chat answers questions. Work is made of tasks that need context, execution, and follow-through. What a real handoff does that a thread cannot.