Most people who say AI is overrated are using it like a slightly smarter Google. Short query in, answer out, move on. It works — kind of — the same way typing a question into a search bar works. You get something back. But you leave most of the actual capability sitting on the table.
This isn’t a criticism. It’s the obvious default behavior, the one the interface invites. A blank text box is a blank text box, and years of muscle memory says: type a short thing, press enter, scan what comes back.
The problem is that AI models aren’t retrieval systems. They’re not fetching a page that already answers your question. They’re generating a response shaped entirely by what you gave them — which means a vague three-word input produces a vague three-paragraph output, and the model had no way to do anything else.
What Search Trained You to Do Wrong
Search engines reward brevity and keyword density. “best budgeting app 2024” works because the algorithm is pattern-matching your tokens against documents that contain those tokens. More context doesn’t help much. You’re querying an index.
AI chat works almost exactly backwards. More context produces dramatically better output, because the model is using everything you give it to narrow down which of many plausible responses actually fits your situation. “best budgeting app” is maximally underspecified — the model has to guess whether you’re a freelancer with lumpy income, someone juggling three joint accounts, or a person on a fixed salary who keeps blowing the food budget. It picks the most generic answer because that’s the safest guess given zero signal.
Tell it your situation — monthly income is irregular, you have a tendency to abandon apps after two weeks, you tried YNAB and bounced off the learning curve — and you get a completely different answer. Not because the model suddenly knows more, but because you gave it enough to stop guessing.
The Actual Shift to Make
Stop treating the first message as the complete request. Treat it as the opening of a working session.
That changes everything about how you write the first message. Instead of asking a question and waiting to be impressed, you’re setting up a problem. That means saying:
- What you’re actually trying to accomplish, not just the surface request
- What constraints matter — time, format, audience, length, tone
- What you’ve already tried or ruled out, so the model doesn’t spend three paragraphs on the obvious
- What a useful response looks like versus a useless one
That last one is the most underused. If you ask “help me write a follow-up email” but you actually need something two sentences long that sounds casual and doesn’t restate the whole previous thread — just say that. The model will write the two-sentence casual email. Without it, you get a five-paragraph professional-tone email you then have to cut down manually, which defeats the point.
A Concrete Example
Here’s the search-engine version of a prompt: “how to negotiate salary”
You’ll get a listicle. Research the market rate, know your worth, practice the conversation. You’ve read this a hundred times.
Here’s the same request reframed as an actual working session: “I have a verbal offer at £52k for a mid-level data analyst role in Edinburgh. Market rate for this level looks like £54–58k based on LinkedIn Salary and two job ads I’ve seen. My target is £56k. I’ve already expressed enthusiasm for the role. Help me draft a one-paragraph response that counters at £56k without being awkward about the fact that I already said I was excited.”
The second version gets you something you can actually use, or very close to it with one round of editing. The first gets you generic advice you already knew.
The difference in effort to write that second prompt? Maybe ninety seconds. The difference in output quality is not marginal.
Why People Don’t Do This
Two reasons, mostly.
First, writing more feels like more work, and the appeal of AI is supposed to be less work. Short-term thinking. Writing a detailed prompt once and getting a usable output is less total work than writing a short prompt, getting a mediocre output, editing it for fifteen minutes, and still not being happy with it.
Second — and this is the part worth being honest about — a lot of people don’t actually know what they want until they start asking. The blank text box is intimidating not because of the interface but because the question lurking behind it is “what exactly am I trying to do here,” and that question is harder than it looks. Search sidesteps it. You can be vague and still get something back. AI punishes vagueness more directly, which makes the vagueness visible in a way it wasn’t before.
That’s actually useful. If you can’t write a specific prompt, it often means you haven’t fully worked out what you need yet. The prompt-writing process forces a clarification that should have happened anyway.
One Habit That Helps Immediately
Before you send any AI prompt, read it back once and ask: if I got a generic, bland response to this, would I be surprised? If the answer is no — if the prompt is vague enough that a generic response is the obvious outcome — add one concrete constraint or one piece of context before sending.
Not five. One. “Make it two paragraphs.” “I’m explaining this to someone who doesn’t work in tech.” “Skip the part about why this matters and just give me the how.” A single specific constraint shifts the output more than you’d expect, because it rules out a large chunk of the plausible-but-useless response space the model would otherwise have to guess its way through.
The fastest way to find that one constraint: look at your prompt and ask what the most generic possible response to it would look like. Whatever that generic response assumes about you — that’s exactly what you need to correct.
