How do you ask an AI agent for what you actually want?
The gap between what you type and what you get back usually comes down to three missing pieces: what the finished thing looks like, what the agent is allowed to do on the way there, and where it should stop and check rather than decide alone. Get those three right and most of the friction disappears.
Why does describing the goal beat listing the steps?
When you hand someone a list of steps, you are also handing them your assumptions about which steps matter. If one of those assumptions is wrong, the person follows the list anyway and arrives somewhere you did not intend. When you describe the goal instead — what the finished thing is, who it is for, what it needs to do — the agent can route around a dead end you did not know was there.
This matters more with an agent than with a search engine or a simple chatbot, because an agent takes actions in the world. It sends messages, deploys code, executes trades. A step-by-step instruction set that is slightly wrong produces a slightly wrong action. A clear goal with stated limits produces a question at the branch point instead.
- Goal
- What the finished thing is and what it needs to do — the measure of success the agent can check against.
- Limit
- What the agent may not do, spend, or change without asking first — the boundary that keeps action safe.
- Signal
- What done actually looks like: a file, a sent message, a deployed page, a confirmed trade — something observable, not a feeling.
- Branch point
- A moment where two reasonable paths diverge and the agent should pause rather than pick one silently.
Why does one example outperform three adjectives?
Adjectives are relative. 'Concise' to a lawyer is not 'concise' to a copywriter. 'Professional' in one industry reads as cold in another. An example, by contrast, is absolute: it shows length, tone, structure and register all at once, without any of those words needing definition.
If you want a weekly summary email in a particular style, paste one you already like and say 'like this'. If you want a spreadsheet laid out a certain way, describe one row in full. The agent does not need to infer what you mean by 'clean' or 'simple' — it can see what you mean. One concrete example is almost always worth more than a paragraph of descriptive adjectives.
Why does saying what to do work better than a list of things not to do?
A list of prohibitions tells the agent what to avoid, but it leaves the space between the prohibitions entirely undefined. The agent fills that space with its defaults, which may or may not match what you wanted. A positive instruction — 'use short sentences', 'confirm the price before any order is placed', 'write in the second person' — leaves far less room for a default you did not choose.
Prohibitions are still useful for hard stops: things that must never happen regardless of context. But they work best as a short fence around a positively described space, not as the primary description of that space. If you find yourself writing more than two or three 'do not' rules, that is usually a sign the goal itself needs to be stated more clearly.
- State the positive version first: what should happen, not just what should not.
- Reserve prohibitions for genuine hard stops — actions that are never acceptable regardless of context.
- If a prohibition is long or conditional, it probably belongs in the goal description instead.
- Check whether each prohibition is actually enforced by a limit (a spend cap, a confirmation step) rather than just stated in words.
Where should an agent stop and ask rather than decide alone?
An agent that can act in the world — sending messages, moving money, deploying code — needs to know which decisions are yours and which it can make on your behalf. The default should be that anything irreversible, anything that spends real resources, and anything that affects someone outside the conversation gets a confirmation step unless you have explicitly pre-approved that class of action.
The practical way to set this is to think about what you would want to see before it happened, not after. If you would want to read the email before it sends, say so. If you would want to see the trade size before it executes, set a limit that triggers a check. An agent that asks at the right moments is not slower — it is safer, and the pauses teach it where your real boundaries are.
GROX handles this through explicit limits you set in plain language: a spend cap per action, a stop condition, a daily loss cap for trading strategies. You can pause, change or cancel any running routine with a sentence. The Overnight Engine, which pursues objectives unattended, operates inside those limits rather than around them.
| Decision type | Reversible? | Suggested handling |
|---|---|---|
| Draft a document | Yes | Agent decides, shows result for review |
| Send a message or post | No | Agent drafts, you approve before sending |
| Execute a trade or payment | No | Pre-set size and stop limits; agent confirms at boundary |
| Deploy or delete code | Partially | Agent previews change, confirms before applying |
| Research and summarise | Yes | Agent decides, flags sources for your check |
Common questions
How long should a prompt to an AI agent be?
Long enough to state the goal, one concrete example of what done looks like, and any hard limits on action. Short enough that you could read it back in thirty seconds. If it is longer than that, the goal is probably not yet clear enough — break it into a first task and a follow-on, rather than trying to specify everything upfront.
What should I include when asking an agent to take an action in the real world?
State what the action is, what a successful outcome looks like, the maximum it may spend or change without checking with you, and any condition that should cause it to stop entirely. An agent acting without those limits will fill the gaps with defaults. The defaults may be reasonable, but they are not yours.
Is it better to give an AI agent one big task or many small ones?
Depends on whether the steps depend on each other. If each step is independent, smaller tasks are easier to check and correct. If the steps share context — a research phase that feeds a writing phase that feeds a publishing phase — a single task with clear checkpoints at each transition is usually cleaner than three separate conversations that lose the thread between them.
How do I tell an agent to stop and ask rather than guess?
Say it explicitly for the decisions that matter most: 'ask me before sending', 'confirm the amount before any payment', 'show me the plan before you start'. For recurring tasks, set a standing limit rather than repeating the instruction each time. An agent that knows your limits in advance can act faster on the decisions you have pre-approved and slower on the ones you have not.
If you want to try this with an agent that can act on your instructions — sending messages, running code, executing trades — GROX offers every capability on a free tier with no card required.