Here’s a moment that has happened to everyone who’s tried this stuff.
You ask for something reasonable — “write a summary of this project for the leadership team” — and back comes 400 words that are grammatically perfect, structurally sensible, and completely useless. Bland where it should be pointed. Generic where the whole value was the specifics. You read it, feel a small deflation, and quietly conclude the tool is overhyped.
Almost every time, the model did exactly what was asked. The problem is that the request contained roughly four per cent of what you actually knew about the job.
The reframe: you’re briefing, not searching
Most people’s instincts here were trained by two decades of search engines. In a search box, shorter is better — you type the fewest keywords that will find the page, and the machine’s job is to guess your intent from a fragment.
That instinct is exactly backwards now, and it’s the single biggest thing to unlearn.
You’re not querying an index. You’re briefing a contractor — one who is fast, widely-read, capable, willing, and who has never met your company, doesn’t know your audience, has no idea what your last deck looked like, and will not ask you a single clarifying question unless you explicitly invite one. They will simply take your four-word brief and make reasonable-sounding assumptions about all of it.
If you handed a talented freelancer a note saying “write a summary of this project for leadership” and nothing else, you’d expect exactly what you got. Not because they’re bad — because you didn’t brief them.
The four things a good ask contains
You don’t need a template. You need four pieces of information, and once you’ve internalized them you’ll supply them in one natural paragraph.
1. Context — who and what for. Who’s reading this, what do they already know, what happens after they read it? “This goes to our CFO, who hasn’t been in any of the working sessions and cares mainly about whether we’re going to overrun the budget.” This single sentence changes the output more than everything else combined, because it determines what gets included and what gets cut.
2. The goal — what it’s for, not just what it is. “Summarize this” describes an operation. “Summarize this so she can decide whether to approve phase two” describes a purpose — and purpose is what lets the model prioritize. Anything you can phrase as so that… is worth phrasing that way.
3. Constraints and shape. Length, format, tone, structure, what to leave out. “Under 300 words, three bullets then a recommendation, no marketing language, don’t restate the project background — she wrote the brief.” Constraints feel like they limit quality; in practice they’re most of it, because a shape you specified is a shape you’ll actually use.
4. What good looks like. The most-skipped and highest-value one — see below.
Show, don’t adjective
Watch what happens with words like professional, engaging, clean, punchy, modern. They feel like specification. They aren’t. Every one of them is an average taken across everything the model has ever seen described that way — which is precisely how you get output that reads like it came from nowhere in particular.
“Professional” gets you the mean of all professional writing on the internet. You almost certainly didn’t want the mean of anything.
The fix is mechanical: replace the adjective with an artefact. Paste two summaries you thought were good and say “match the register of these.” Attach last quarter’s version and say “same structure, new content.” Give one example of the thing done right and one of it done wrong, and say which is which.
This works so well because it converts an abstraction into evidence. It also does something an adjective can never do — it carries all the things about your house style that you’ve never articulated and probably couldn’t. You don’t have to be able to describe your voice. You just have to be able to point at it.
Give it the material, not a description of the material
Related, and it’s the difference between a tool that’s mildly useful and one that changes your week.
Compare:
- “What are the main risks in our Q3 project?” — the model has no Q3 project. It will compose plausible risks for a plausible project. Fluent, confident, invented. (Why that happens, and how to spot it, is why AI makes things up.)
- “Here’s the project brief, the last three status updates, and the risk log — what’s the biggest risk that isn’t already being tracked?” — now it’s reading comprehension over your actual situation, which is something these systems are genuinely excellent at.
Same question. Entirely different reliability, because the second one is grounded in real material.
This is why tools that work on your files rather than on what you paste are such a step up for ordinary work. Something like Claude Code Desktop opens a folder and can read what’s actually in it, so “the material” stops being a copy-paste chore and becomes the default. The best prompt in the world can’t compete with the mediocre prompt that has the real documents attached.
Ask for the plan before the work
One habit, disproportionate returns, and it costs about fifteen seconds:
“Before you write it, tell me your outline and what you’re assuming.”
You get back a five-line plan. You read it, and either it’s right — say “go” — or, far more often, you spot in three seconds that it’s aimed at the wrong audience, or missing the thing that actually matters, or about to spend 300 words on background you didn’t want.
Correcting a five-line plan is trivial. Correcting a finished 800-word document is a rewrite, and worse, it’s a rewrite you’re psychologically reluctant to demand because something already exists. Most of the frustration people report with these tools is the cost of discovering a misunderstanding after the work was done rather than before.
The variant for anything ambiguous: “What do you need to know from me to do this well?” The model will ask three questions, and typically at least one of them will be something you knew and hadn’t thought to mention.
Steer — don’t restart
Here’s where the gap between casual and effective users is widest. When the first answer isn’t right, most people delete it and rewrite the prompt from scratch.
Don’t. The second message is where the quality lives, and it’s cheaper than the first:
- “Too long. Half the length, and cut the background — she wrote the brief.”
- “The second bullet is the actual point. Lead with it and expand it.”
- “Right content, wrong register — this reads like a press release. Plainer.”
You’re doing exactly what you’d do with a colleague’s draft, and you’re doing it in the medium where it’s cheapest to do. Three rounds of ten-second steering will beat any single perfectly-crafted opening prompt, essentially always.
Two things make steering work better: be specific about what was wrong (not “I don’t like it” — “the tone is too formal”), and keep the good parts explicitly (“keep the structure, redo the language”) so you don’t lose ground on each pass.
A before and after
Before:
Write a summary of the website project for leadership.
After:
Here are the project brief and the last four status updates. Write a summary for our CFO — she hasn’t been in the working sessions and mainly cares about whether we’ll overrun. Under 300 words: three bullets on where we are, then one clear recommendation on phase two. Plain language, no marketing tone, don’t restate the background. Here’s last quarter’s version for the register. Before you write it, give me your outline.
Not clever. Not engineered. Just the four things — context, goal, shape, standard — plus the actual material and a checkpoint. That’s the whole skill, and you now have it.
When prompting stops being the lever
One honest caveat to end on, because it’s where the field has actually moved.
Everything above is about a single request. The moment your work becomes repeated — the same weekly report, the same review, the same document type across a team — wording individual prompts stops being where the quality comes from. What matters then is what the model can see: the right files, the right background, your standing conventions written down once, and the absence of irrelevant clutter.
That’s a different discipline, and it’s the one that scales. We cover it in context engineering, and the distinction between the two in prompt engineering vs context engineering.
But you don’t need any of that yet. Brief it like a contractor, show don’t adjective, hand over the real material, ask for the plan, and steer instead of restarting. Do those five things this week and the tool you thought was overhyped will look like a different product.