When an AI answer comes back wrong, the cause is usually missing information, and better wording rarely fixes that. Prompting matters for setting the task and the format. Beyond that, quality depends on what the model knows about your situation: your facts, your rules, an example of good work and the decisions you've already made.

Most people do the opposite. They add "be concise", then "think step by step", then a paragraph of role instructions, get something slightly better on the fourth try and decide prompting is a skill they need to learn. We think they're working on the smallest lever in the room.

A supplier email, four rewrites later

Say you run operations for a mid-sized logistics firm in Manchester. A packaging supplier writes asking to raise prices by 8 per cent from next month. You ask an assistant to draft a reply that's "firm but polite".

The draft is fluent and gently accepts a smaller rise. So you try again: "firmer, and don't concede anything". Then "push back on the timing". Then you add "you are an experienced procurement manager". Each version reads differently, and none is right, because the assistant doesn't know three things: your contract fixes prices until April, you turned down a similar request from this supplier in March, and you've got a second supplier quoting for the same work.

No rewording supplies those facts. Two sentences of context would have produced a usable draft first time.

The fourth rewrite is the tell. If the answer is still wrong in the same way after several attempts, the gap is upstream of the wording, and you're trying to phrase your way around something the model was never told.

What can a prompt fix, and what can't it?

A prompt sets the task and the shape of the answer. Ask for a table and you get a table. Ask for three options under a hundred words each and you get roughly that. Ask it to take a position instead of summarising and it usually will.

Prompting is a real skill, and dismissing it would be its own kind of laziness. Clear framing, explicit constraints, output format and worked examples all change results, and a precise request to a weaker model often beats a vague one to a stronger model. Vendors' documentation has said so for years.

What a prompt can't do is supply facts. If the model doesn't know your customer's history, your margin, your house style or why the last draft was rejected, you'll get a fluent answer to a slightly different question. That's the most dangerous kind, because it looks like progress.

The way we'd order it: model capability sets the ceiling, context decides how close to it you get, and wording moves you around a little within that range. Once the task is clearly stated, further rewording mostly shuffles the answer around.

Why doesn't pasting everything in work?

The obvious next move is to paste in the lot. The research says that backfires.

In 2023, Liu and colleagues found that language models often do best when the relevant information sits at the start or end of a long input, and measurably worse when it's buried in the middle (Lost in the Middle). Where a fact sits matters as well as whether it's there.

Chroma's July 2025 report tested 18 models, including the leading frontier systems of the time, on deliberately simple tasks and found performance growing less reliable as inputs got longer, a pattern they call context rot (Chroma).

Anthropic's engineering team made the same case from the builder's side in September 2025, calling context "a finite resource with diminishing marginal returns" and describing the job as context engineering: deciding what goes in, in what order, and what stays out.

So the aim is the right material in the right order. That's a writing job, and most organisations haven't done it.

What's worth writing down once?

Most of this is reusable. You write it once and it serves every later task. It usually falls into four kinds of material:

  • Facts. Who you are, what you sell, who you work with, how the business makes money, what matters this quarter. Short, plain sentences.
  • Rules. Things that are always true about your work. Never promise a delivery date. Quote in the client's currency. We don't say "solution". Rules are cheap to write and among the easiest things for a model to follow.
  • Examples. One piece you'd happily send and one you wouldn't, each with a sentence on why.
  • Decisions. What you settled last time, and the reason. Without this, the model will cheerfully suggest the thing you rejected in March, and you'll spend the conversation arguing it again.

Of the four, examples do the most and get skipped most often. Ask for "professional but warm" and you get a guess at what those words mean. Show two paragraphs and say "like this one, not that one" and the model can see sentence length, how directly the point lands, whether contractions appear and how much hedging is normal. The sentence explaining the difference is what turns imitation into a rule you can refer back to. It also makes correction quicker: you can say which example a draft drifted towards, which is the habit described in correcting your agent.

Together this comes to a page or two, and it's the same page every time. If your tool has memory, this is what belongs in it; how AI agent memory works explains how that's stored. If it doesn't, keep the page somewhere you can paste from.

The same caution applies here as with pasting: context means the shortest well-ordered set of material that would let a capable new colleague do the task. If a new starter couldn't do it from what you've written, the model probably can't either.

Next time an answer comes back wrong

  • Ask "what would it need to know to get this right?" before you reach for different wording, then supply that.
  • Write the facts-and-rules page: a couple of hundred words on the business, then ten rules. Half an hour pays back on every task after it.
  • Put background first, then the task, then the format you want. It costs nothing and follows what the position research found.
  • Halve whatever you pasted last time and compare. The shorter version often wins.

If you're building this up for the first time, getting the best from an agent covers the routine, and once your pages settle down they can become the reusable recipes described in agent skills.

Common questions

What's the difference between a prompt and context?

A prompt is the instruction for this task. Context is everything the model knows that makes the instruction meaningful: facts, rules, examples and earlier decisions. You write a prompt every time, while context is written once and reused, which is why it pays back more.

How much context is too much?

Too much is when the material stops being read carefully. Reliability tends to fall as inputs grow, and details in the middle fare worst. If you can't point to the few sentences that decide the answer, the input is probably too long.

How do I know my context page is good enough?

Give it and the task to a new colleague and see whether they could do the job from it. Whatever a competent person would have to ask about, the model is missing too. It's a free test and it catches most gaps.

Is this worth doing for one-off questions?

Usually not. For a single question, one clear sentence with the key facts is enough. Context pays off on repeated work, such as the monthly report, the weekly replies or anything that follows a house style.