How to Talk to AI Agents: Closing the Communication Gap, a Beginner’s Q&A
If you have ever asked an AI agent to build something, got back a perfectly reasonable result, and still thought “that is not what I meant”, this article is for you. It is not a list of clever prompt tricks. It is about how to talk to AI agents when the real problem is the gap between what is in your head and what you actually wrote down, and why that gap is a human problem that stays with you no matter how good the models get. Everything here is written in plain English for people just starting out: designers, writers, founders, students, anyone who has to get real work out of an AI and does not come from engineering.
Learning how to talk to AI is mostly learning how to notice what you have left out. The agent cannot read your mind. It does not know your goal, your taste, what “done” looks like to you, or what you would never accept. Before every request, ask one question: what does the agent not know about me and about what I want? Then write that down. Most of the rest follows.
Why “I want this” means less than you think
In everyday life, “I want this” works because the person listening already shares most of your context. Your colleague knows the client, the deadline, the thing that went wrong last time, and roughly what you consider good. They fill in the blanks without noticing they are doing it.
An AI agent has none of that. It has your words, whatever files or context it can see, and a very large amount of general knowledge about how other people usually do things. So when you say “build me a booking form”, it builds the most ordinary booking form in the world, because that is the most likely reading of the words you gave it. It is not being lazy. It is being statistically sensible with far too little information.
Ask a stranger to “make me a tea” and you will get a tea. Milk or not, how strong, which mug, sugar: the stranger guesses, and the guess is based on what most people want, not what you want. You would not call the stranger bad at making tea. You would say you did not tell them enough. Talking to an AI agent is making tea for someone who has never met you, every single time.
Engineers learned this lesson the hard way, long before AI. A one-line feature request from a client becomes a week of back-and-forth, because “add a login” hides fifty decisions: email or phone, password rules, what happens when someone forgets, what happens on the third wrong attempt. Our guide to briefing a development agency exists for exactly this reason. The agent is the same, only faster and more literal, so the gaps show up in minutes instead of weeks.
The communication gap, in Matt Pocock’s words
The clearest short explanation of this we have found comes from Matt Pocock, the Oxford-based developer educator behind Total TypeScript and AI Hero, who now teaches AI-assisted engineering full time. In a 60-second short he makes the point that people massively underestimate the communication gap between themselves and the agent. His framing: working with an agent is like working with a colleague you have just hired, and like any colleague there is a barrier between what you mean and what they hear. The part most people get wrong is assuming the barrier will vanish as the models improve. It will not, because, as he puts it, “it is not a model problem. It is a human problem.”
His reasoning is worth sitting with. We often do not know exactly what we want. We carry complicated, half-conscious priorities about the output we are after, and we care about many things in a particular order. The agent knows none of that, and we rarely know it well enough ourselves to say it out loud. His practical advice is a single habit: keep the gap in mind, and keep asking what the agent does not know about you and about what you want. The rest of this article is that habit, broken into the questions beginners actually ask.
How to talk to AI agents: the beginner’s Q&A
Do I need “prompt engineering” to learn how to talk to AI properly?
No. Prompt engineering is a real craft, and the official guides are worth reading later, but the beginner’s version is just good briefing. If you can explain a task to a new colleague on their first day, you already know how to talk to AI agents. The skills are the same: say what you are trying to achieve, say what good looks like, say what must not happen, and say what already exists. Anthropic’s own prompt engineering overview starts with “be clear and direct” for a reason. The clever techniques come after that, and most of the time you will not need them.
It built something different from what was in my head. Whose fault is that?
Almost always yours, and that is good news, because learning how to talk to AI is something you can get better at tonight. When the result is wrong, read your own request again as if you were a stranger. Nine times out of ten you will find a word that could mean two things, a priority you never stated, or a whole requirement that lived only in your imagination. The agent picked the most common interpretation. Your job next time is to make the uncommon one explicit.
What does the agent not know about me?
Here is the list that catches most of the gaps. Run through it before any task that matters.
How much should I write? I do not want to write an essay for every request.
Write in proportion to the cost of being wrong. A quick question gets a quick prompt. A feature your customers will use for a year deserves ten minutes of writing, which is still nothing compared with the hours you lose when it comes back wrong. A useful test: if you would send this request to a freelancer and expect a sensible result without a follow-up call, it is detailed enough. If you would expect the freelancer to ring you with questions, write those answers down now.
Can I just ask the agent to ask me questions?
Yes, and for anything non-trivial you should. End your request with: “Before you start, ask me the five questions whose answers would most change what you build.” This is the single cheapest way to find the gaps you cannot see, because the agent is often better at spotting ambiguity in your words than you are. Answer the questions, and only then say “go”. Many coding tools have a planning mode built for exactly this; Anthropic’s Claude Code best practices recommend asking the agent to make a plan and checking it before any code is written.
The agent did exactly what I said, and it was still wrong. Now what?
This is the communication gap showing itself, and it is the moment to learn rather than to get annoyed. Ask the agent to list the assumptions it made. You will usually find two or three that you would have overruled had you known about them. Add those to your request, and keep a note of them for next time. Over a few weeks you will build a personal checklist of the things you always forget to say, and that checklist is worth more than any prompt template on the internet, because it is about you.
I am not technical. How do I check the work?
You check it the way a customer would, and you ask the agent to translate the rest. Three habits: first, ask for a plain-English summary of what was built and what was deliberately left out. Second, try the thing yourself, on your own phone, as a real user, and write down every surprise. Third, ask the agent “what would a senior engineer worry about in what you just built?” and read the answer slowly. If you are maintaining something an AI built for you, our guide to Claude Code for non-developers covers the safe ways to keep going without breaking it.
My English is not perfect. Can I still learn how to talk to AI?
Yes, and this matters, because a lot of people hold back for this reason. Agents read imperfect grammar, mixed languages and plain lists perfectly well. Structure beats polish every time: a messy list of “must”, “must not” and “done when” will get you a better result than an elegant paragraph that forgets the constraints. If thinking in your first language is easier, write in it. The agent will cope. The communication gap is about missing information, not missing vocabulary, and learning how to talk to AI has nothing to do with accent or grammar.
Does it matter if I am polite?
Not much, in the sense that “please” does not change the result. It matters a lot in a different sense: if you talk to the agent the way you would talk to a capable colleague, you naturally give it context, reasons and the chance to push back, and all three improve the output. People who bark one-line commands tend to get one-line thinking in return. Treat it as a conversation. If the first result is nearly right, say what is wrong with it and ask for another pass instead of starting over.
Will knowing how to talk to AI matter less as models get better?
Different, not easier. Knowing how to talk to AI will matter more, not less. Better models make fewer silly mistakes and understand vaguer instructions, which is a real gain. But a model that understands you better still cannot know the priorities you have not stated or the preference you have not discovered yet. Simon Willison, the British developer behind Datasette and one of the most careful writers on this subject, describes his own production work with LLMs as treating the model like a diligent intern who needs detailed instructions, in his guide on how he uses LLMs to write code. The instructions are the work. That does not go away, and it is the part that makes you valuable.
What is a good first project for learning how to talk to AI?
Something small, real and yours. A page for a thing you actually sell, a tool that fixes a chore you actually have, a prototype of an idea you have been carrying around. Avoid tutorials that have nothing to do with your life, because the whole point is to practise noticing what you want, and you only know what you want when the result affects you. If the project goes well and starts to matter, our prototype-to-production audit is the next thing to read.
A brief you can copy for any AI task
This is the shape we use ourselves, the most practical answer we have to how to talk to AI about real work, and it maps directly onto the “what does the agent not know” list above. Fill in every line, even if the answer is “none”. The empty lines are where the gaps hide.
Done looks like: how I will check it, step by step
Must: the non-negotiables
Must not: the things I will reject on sight
Priority order: what wins when two of the above collide
What already exists: files, links, examples, naming
Taste: two things I like, one thing I hate, and why
Before you start: ask me the questions that would change your plan
For bigger pieces of work, the same thinking expands into our full app development brief template, and if you are not sure whether to keep prompting or bring in a person, vibe coding vs hiring a developer sets out the honest trade-offs.
Where to learn more about how to talk to AI
And when you hit a word you do not recognise along the way, our glossary for founders explains the 35 terms that come up most when talking to engineers, human or otherwise.
FAQ: how to talk to AI, the quick answers
Getting a team started with AI agents?
We build web and mobile products with AI agents every day, and we have made every mistake in this article at least once. Tell us what you are trying to make and where the gap keeps showing up.
