Common Prompting Mistakes and How to Fix Them
Most bad AI answers trace back to a handful of avoidable prompt mistakes. Here they are, each with a quick, concrete fix.
Most bad AI output isn't the model being dumb. It's a prompt that quietly assumed the model could read your mind, guess your format, and infer the context sitting in your head. The good news is that prompting mistakes are boringly repetitive. Once you can name them, you spot them in your own drafts before you hit send. Here are the ten that trip people up most, each with a fix you can apply in the next five seconds and, where it helps, a weak-to-fixed rewrite.
The model only knows what you put on the page
Three of the most common mistakes share a single root: you're carrying information the model never received.
The first is writing a request so open-ended it could be answered a hundred ways. "Summarize this" — how long, for whom, keeping which details? You already know the answer, so the gap feels invisible.
The second is skipping the role or context entirely. A model with no framing defaults to a bland, average-of-the-internet voice. Give it a job and an audience and the register snaps into focus.
The third is the sneaky one: assuming the model knows your private context. Your product name, your customer, last week's decision, the tone your boss hates — none of that exists to the model unless you type it. It will happily invent a plausible substitute instead.
The fixed version isn't longer for the sake of it. Every added clause closes a gap the model would otherwise fill by guessing. When answers come back generic, the prompt is usually generic too, and the fix is to hand over the specifics you were keeping to yourself. If you want a second pair of eyes on how much context you're actually supplying, run a draft through the prompt analyzer before you blame the model.
- Name the audience and the goal in the same sentence.
- Paste the facts the model can't know: names, numbers, prior decisions.
- Give it a role only when the role changes the answer (a lawyer, a teacher, a skeptic).
Tell it what "done" looks like
You can describe a task perfectly and still get something unusable because you never said what shape the answer should take. Do you want a table, three bullets, a JSON object, a single paragraph? The model picks for you, and it rarely picks what you had in mind.
Closely related is leaning on vague quality words. "Make it professional." "Write something good." "Keep it engaging." These feel like instructions but carry almost no signal, because "professional" to you might mean terse and to the model might mean stuffed with corporate filler. Replace the adjective with an observable trait.
And when the format you want is unusual, describe it and then show one example. Rules alone leave too much room; a single sample collapses the ambiguity instantly. This is the everyday case for few-shot versus zero-shot prompting — you don't need ten examples, one well-chosen one usually does the work.
One prompt, one job
There's a strong temptation to write one mega-prompt: research the topic, outline it, write it, fact-check it, and format it as a newsletter. The model tries to do all five at once and does each of them at maybe seventy percent, then drops one entirely because it ran out of attention. Reasoning quality drops as you pile jobs onto a single request.
Split the work. Get the outline, look at it, then ask for the draft from that outline. You keep a checkpoint between each stage where you can correct course, and each step gets the model's full focus. The extra round-trip costs you a few seconds and saves you rewriting a tangled block of output that missed half your intent.
- Multi-part deliverable? One stage per message.
- Keep the output of each step as the input to the next.
- If a prompt has more than one "and then," it's probably two prompts.
Give it room to sound human
The opposite failure is over-correcting. You've been burned by vague prompts, so now you write a wall of constraints: exactly 100 words, exactly four sentences, must include these six phrases, no commas, formal but friendly, and so on. The model obeys, and the result reads like it was assembled by someone filling out a tax form. Every constraint you add is a degree of freedom you take away, and past a point the writing goes stiff and lifeless.
Constrain the things that genuinely matter — length, format, a fact that must appear — and leave the rest open so the model can actually write. If you're not sure which of your rules are load-bearing, an prompt optimizer can help you strip the prompt back to the constraints that change the output and drop the ones just adding noise.
Plan for the moment it doesn't know
Models hate saying "I don't know." Left to their own defaults, they'll produce a confident, well-formatted, completely fabricated answer rather than admit a gap. The mistake is never telling it what to do when it's unsure, so hallucination becomes the path of least resistance.
Give it an escape hatch and permission to use it.
Naming the fallback — "say you're unsure," "flag it," "list what you'd need to answer" — turns an invented answer into an honest one. It's the single cheapest guard against confident nonsense, and it costs one sentence.
Fix the prompt, don't burn it down
When output disappoints, the instinct is to trash everything and rewrite from scratch. Usually that's wasteful. A near-miss tells you exactly which one thing to adjust. If the tone is off, correct the tone and keep the rest. If it ignored a constraint, move that constraint to the top or make it louder. Treat prompting as a dialogue with small edits, not a slot machine you keep re-pulling.
Iterating beats restarting because each version teaches you something about how the model read your words. Save the prompts that work, note what you changed, and your next first draft starts better. If you want the full method behind this loop, how to write AI prompts and prompt engineering fundamentals walk through building prompts that need fewer rounds to begin with.
None of these fixes are clever. They're just the difference between telling the model what you want and hoping it guesses. Do that consistently and the "the AI is bad at this" complaints mostly disappear.
