Stop Prompting AI.
Start Briefing It.

I keep seeing advice about the perfect AI prompt.

There are formulas. Acronyms. Cheat sheets. Vast collections of phrases that apparently unlock the machine if spoken in the correct order.

‘Act as a world-class expert.’

‘Think step by step.’

‘Your answer is very important to my career.’

Some of these instructions can be useful. Some have become the digital equivalent of rubbing a lamp and hoping the genie has worked in management consulting.

But I think we may be concentrating on the wrong thing.

Most disappointing AI results are not caused by a failure to discover the right incantation. They are caused by a poorly defined job, missing context, invisible expectations or nobody being quite sure what a good answer would look like.

In other words, the AI has not been prompted badly.

It has been briefed badly.

I have started to live and breathe AI to a degree that is probably not medically useful. I use it to research, write, code, test ideas, challenge arguments and help manage a growing collection of projects that seemed perfectly reasonable when I started them. I have used individual models, specialist tools and teams of agents with different roles.

The more capable the tools become, the more convinced I am that effective use is less about clever prompting and more about an old-fashioned skill: explaining the work clearly.

A prompt is not necessarily a brief

Consider this prompt:

Help me prepare for a difficult project meeting.

It is not terrible. The AI will produce something. It will probably suggest setting an agenda, listening actively and ensuring everyone has an opportunity to contribute.

You may even receive a table.

What it cannot know is why the meeting is difficult, what must be decided, who is resisting it, what authority you have, what has already been tried or whether ‘successful’ means agreement, escalation or simply surviving until lunch.

Now compare it with this:

I need to leave tomorrow's project meeting with agreement on who owns the delayed integration decision. The participants and their current positions are below. Identify the unresolved issue, the three most likely objections and the questions I should ask to reach a clear decision. Keep the tone calm and constructive. Do not write a speech or add generic meeting advice. Give me a one-page brief I can scan in two minutes.

There are no magic words in that prompt. It simply contains a job.

I find a useful AI brief answers five questions:

1. Intent: What are we trying to achieve, and for whom?
2. Context: What does the AI need to know for this particular job?
3. Constraints: What must it preserve, avoid or treat as non-negotiable?
4. Output: What should it actually produce?
5. Standard: What would make the result good enough?

You do not need to turn this into another acronym. The world has suffered enough.

The point is to move beyond telling the AI what activity to perform and explain the outcome the activity is meant to support.

‘Summarise this document’ describes an activity.

‘Summarise the decisions, owners and unresolved risks from this document for a project sponsor who has five minutes before the steering committee’ describes a purpose.

The difference is not decorative. It changes what information matters.

Do not think like AI. Think about what it can see

People sometimes say you need to think like an AI to use one well.

I am not entirely sure what that would involve, but it sounds tiring. I already have enough trouble thinking like myself before coffee.

You do not need to think like the model. You need to think about what the model can actually see.

It cannot see the meeting behind the minutes, the politics behind the request, the quality standard sitting silently in your head or the conversation in which everybody agreed that option three was dead. Unless that information is available in the current interaction, the AI is working without it.

This is why ‘make it better’ is such a hopeful instruction.

Better for whom? Better in what way? Shorter? Warmer? More rigorous? More persuasive? Less likely to make the legal team appear suddenly at your desk?

Before blaming the output, ask a mildly uncomfortable question:

If a capable colleague received only what I supplied here, would they know what I meant?

If the answer is no, the model may not be the main problem.

Context is a workbench, not a storage unit

Most of us interact with AI through a chat window. That makes the exchange feel like a conversation, and conversations encourage accumulation.

We add another question. Then a document. Then a correction. Then a different task. Three days later, the same chat contains a product strategy, a half-written email, some Python code and a discussion about whether the word ‘platform’ has finally lost all meaning.

All of this becomes context.

A context window is the amount of material a model can work with during an interaction. Depending on the system, that may include your instructions, conversation history, supplied documents, retrieved information, tool results and the model's own responses.

Modern context windows can be very large. This is useful. It does not mean filling one is automatically useful.

Capacity is not the same as comprehension.

Research into long-context models has found that relevant information can be used unevenly depending on where it appears, a problem commonly described as being ‘lost in the middle’. Models and techniques continue to improve, but the practical lesson remains sound: making information available is not the same as making it salient.

Think of the context window as a workbench, not a storage unit.

Put on the bench what the current job requires. Keep the rest nearby, organised and available if needed. If the model must search through six obsolete drafts to discover which paragraph is authoritative, you have not provided rich context. You have assigned archaeology.

The aim is not the smallest possible context. Starving the model of necessary information is not a sophistication.

The aim is the smallest sufficient context.

That usually means:

• giving it the sources that govern the task;
• removing duplicates and superseded versions;
• explaining which material is authoritative;
• separating background from instructions;
• stating the current decision or question near the point of work; and
• keeping enough room for the answer, critique and revision that follow.

More context can help. Better context helps more reliably.

End the meeting when the meeting ends

Chat interfaces encourage us to keep everything in one conversation because continuity feels useful. Sometimes it is. Sometimes it is the digital equivalent of forgetting to stop the recording after a Teams meeting.

The project team finishes its discussion and leaves the room. Another group arrives to debate the contents of the canteen vending machine. They are followed by the marketing team, who spend an hour planning next year's campaign. All the while, the original meeting is still being recorded.

A week later, somebody asks for a summary of the project meeting.

The transcript contains the project decisions, a surprisingly heated disagreement about salt-and-vinegar crisps, and next year's brand positioning. The information you need is in there. Unfortunately, so is everything else.

Long-running AI conversations can develop the same problem. Old instructions, abandoned ideas, unrelated tasks and superseded decisions remain in the context. The model must work out which parts still matter, which have expired, and whether the vending machine is somehow a project dependency.

When the job changes substantially, start a new conversation.

Carry across a short briefing note:

• what has been decided;
• what remains open;
• which source material governs;
• what constraints still apply; and
• what you want to do next.

A fresh chat is not throwing away useful context. It is curating it.

It is also ending the meeting before somebody starts discussing the crisps.

Show the standard

Instructions tell the AI what you want. Examples show it.

If you want the AI to write in your voice, provide one or two representative pieces. If you want a particular report format, supply a good example. If you need it to classify support tickets according to a policy, give it the policy and a few correctly classified cases.

This is often more effective than adding another paragraph of adjectives.

‘Make it concise, warm, professional, engaging, authoritative but approachable’ leaves considerable room for interpretation. It also sounds like a brand workshop that has gone 20 minutes over time.

‘Match the tone and approximate length of this example, but do not reuse its argument or phrasing’ gives the model something it can compare against.

There is an important limit here. One example may contain accidental features you do not want copied. Several examples can reveal the pattern more clearly. Tell the model what to imitate and what is merely incidental.

Good context is not just information about the subject. It is evidence of the standard.

Give it stages, not a heroic assignment

Another common approach is to ask the AI to research a subject, decide the argument, draft the article, check the facts, improve the style and produce the final version in one prompt.

It may comply. AI is very accommodating in this way.

The result can look remarkably complete while hiding a weak assumption made near the beginning. Every later step then builds neatly on top of it, like a beautifully furnished house on the wrong block of land.

For substantial work, separate the stages:

1. Frame the problem.
2. Identify missing information.
3. Generate or compare approaches.
4. Choose the direction.
5. Produce the draft or analysis.
6. Critique it against explicit criteria.
7. Revise it.
8. Verify consequential claims.

You do not need to use every stage for every task. Nobody needs an eight-step workflow to improve a birthday invitation.

The principle is to pause at the points where an early error would become expensive. Check the framing before producing 40 pages. Choose the argument before polishing the conclusion. Verify the evidence before sending the recommendation to somebody whose title contains the word ‘Risk’.

Yes, it can be a team of assistants

People often say we should treat AI like an assistant. There is more truth in that phrase than its slightly breathless use on social media suggests.

A good assistant does not merely wait for commands. They help organise information, prepare options, notice gaps, draft material and reduce the friction between an idea and a completed piece of work. AI can do many of these things quickly and, in some cases, extremely well.

It can also behave like a team of assistants.

One can research the topic. Another can test the logic. Another can edit for clarity. Another can look specifically for unsupported claims, security risks or the objection everybody else has politely avoided.

These may be separate AI agents, separate conversations or simply separate passes with clearly defined roles. The underlying benefit is the same: creation and criticism are different kinds of work, and separating them reduces the temptation for the drafter to approve its own homework.

There is, however, a detail missing from the ‘team of assistants’ metaphor.

You are now the manager.

You must define the work, decide who receives what context, resolve disagreements and judge when the result is ready. If four assistants receive the same vague brief, you may get four polished versions of the same mistake. This is not collective intelligence. It is synchronised confidence.

More agents do not automatically produce a better answer. They can also produce more cost, more duplication and a small committee that needs to be summarised by another agent.

Use multiple roles when the work genuinely benefits from different perspectives. Do not assemble an AI department to rename a spreadsheet tab.

Ask it to challenge you

AI systems are designed to be helpful, and helpfulness can sometimes look suspiciously like agreement.

If your framing is weak, a fluent answer may strengthen it without questioning it. That is pleasant. It is not always useful.

Build challenge into the brief:

• What assumption am I making that may not hold?
• What is the strongest reasonable objection?
• What evidence would change this conclusion?
• Separate known facts, evidence-based inferences, uncertainty and guesses.
• Where is this most likely to fail in practice?
• What important question have I not asked?

Do not ask for relentless opposition. We have all attended that meeting too. Ask for specific challenge at the point where it can improve the decision.

AI is particularly useful as something to push against. It can generate a plausible opposing case without becoming offended, forming a subcommittee or replying to your email at 11.47 pm with ‘a few initial thoughts’.

Tell it when to stop

AI makes expansion almost frictionless.

There is always another option, another framework, another table, another implementation roadmap and, if you show a moment's weakness, a proposed governance model.

Define the stopping point:

• one page;
• three viable options;
• enough detail to make the decision;
• no implementation plan yet;
• flag missing information instead of filling the gaps;
• stop and ask if a missing fact would materially change the answer.

This is not merely a way to receive shorter responses. It tells the AI what stage of the work you are in.

A brainstorming session should not quietly become a commitment. A first draft should not impersonate an approved policy. A plausible answer should not acquire authority simply because it has headings.

Good briefing defines the boundary of the work.

Keep the judgement

The final habit is the least exciting and the most important.

Check the work.

AI can help frame, explore, compare, draft, critique and revise. It can find inconsistencies you missed and produce options you would not have considered. It can also misunderstand the task, rely on the wrong source, invent a connection or state an uncertain conclusion with the serene confidence of somebody who will not be attending the consequences.

The more polished the answer looks, the easier it is to forget this.

For low-consequence work, checking may take seconds. For important work, ask for sources, trace claims back to them, distinguish fact from inference and have a qualified human make the decision. The level of verification should rise with the cost of being wrong.

‘The AI said’ is not accountability. It is a description of where the text came from.

A briefing template

If you want a practical starting point, use this:

Intent: I need to [achieve an outcome] for [audience or decision].

Context: The relevant background is [brief background]. Use [named sources or attached material] as authoritative.

Constraints: Preserve [important elements]. Do not [specific exclusions]. If essential information is missing, flag it rather than guessing.

Output: Produce [format, length and structure].

Standard: A good result will [criteria by which you will judge it].

Before beginning, identify any ambiguity that would materially change the result.

You will not need every field every time. A quick task deserves a quick request. But when the work matters, a useful brief is usually cheaper than three rounds of disappointed clarification.

The real skill is not persuading the AI to perform.

It is making the work legible.

Treat AI like an assistant. Give it a clear job, the relevant material, an example where useful and a standard it can work towards. Treat it like a team of assistants when the work benefits from separate roles.

Just remember that you are still the one whose name appears at the bottom.

Management. It gets you in the end.

Source Notes

• Liu et al., Lost in the Middle: How Language Models Use Long Contexts, Transactions of the Association for Computational Linguistics, 2024. Useful evidence that the ability to accept long context does not guarantee that relevant information will be used equally well throughout it.