Priya read her colleague’s ChatGPT answer over his shoulder on Tuesday morning. Same tool, same three-line question about restructuring a client’s timeline. His answer came back as three ranked options with a trade-off named for each. Hers, sent an hour earlier, was a paragraph of hedging. It named nothing. A weaker ChatGPT answer is, in most cases, the output of an under-specified prompt: one missing a stated role, the situation’s context or a named output format, not proof that the person asking is worse at this than the colleague beside them. That gap is documented independently across several technology outlets, not one study, and it is checkable in your message history in the time it takes to reread three prompts.
What exactly is missing from a weak prompt?
A weak ChatGPT prompt is short on three things: who the model should be, what the situation actually is, and what shape the answer should take. Ask ChatGPT to “help me write a difficult email to a client” and it has no way to know whether you’re an account manager softening bad news or a lawyer setting a hard deadline, so it defaults to a generic, safe tone that fits nobody’s situation well. Add “act as a senior account manager, the client is three weeks late on payment, write four sentences that state the new deadline without threatening the relationship” and the model now has a role, a fact set and a length. Agni, an AI-literacy app in development for iPhone, tracks this exact gap because it turns out to be one of the most common and most fixable differences between people getting strong and weak results from the same model. The mechanism is not exotic: a missing instruction, not a missing skill.
Is this actually documented, or is it a theory?
The link between prompt structure and answer quality doesn’t rest on one study. It rests on a pattern independent technology outlets keep landing on separately. PCWorld’s piece “Bad ChatGPT answer? Maybe you’re asking the wrong question” makes the case that with ChatGPT, Claude and Gemini alike, a prompt that is too vague produces a wide, unfocused answer, and one that is too narrow produces an answer too blinkered to be useful, so the practical fix is often to have the model check the question itself before it answers. How-To Geek’s “6 Ways to Get Better Responses From AI Chatbots” names the other half of the same fix: give ChatGPT a role or persona, such as instructing it to act as a maths tutor or a specific kind of expert, before stating the task, because the role narrows which of the model’s plausible answers it reaches for. Two separate outlets, one mechanism: give the model a role and the right scope, and the hedging stops.
What does a fixed version of the same prompt look like?
A fixed prompt keeps the underlying question but adds three things a weak one leaves out. Priya’s original read: “Help me restructure this project timeline.” It named no role, no fact about the client relationship, and no format, so ChatGPT had to guess at all three and picked the safest, vaguest version of each. Her colleague’s version read: “Act as a senior project manager. The client is upset about a two-week slip and expects an update today. Give me three options as a numbered list, one sentence of trade-off each.” Same task underneath, three added pieces of information, a different answer entirely.
| Missing element | Priya’s prompt | Colleague’s prompt |
|---|---|---|
| Role | none stated | ”act as a senior project manager” |
| Situation | none stated | ”client upset about a two-week slip, needs an update today” |
| Output format | none stated | ”three options, numbered list, one sentence of trade-off each” |
| Result | a paragraph of hedged generalities | three ranked, usable options |
The model didn’t get smarter between the two prompts. It got told more.
How do you check whether it’s your prompting?
Whether a run of weak ChatGPT answers comes down to a prompting habit, or something else entirely, is a specific question with a checkable answer. Start with your own message history: pull up the last five prompts where an answer disappointed you and check each one for a stated role, a stated situation and a named output format. Missing two or more of the three in most of them matches the pattern the outlets above describe, and it’s fixable within an afternoon of rewriting habits, not months of practice. For a structured version of the same check, Agni’s AI Literacy Score is a fourteen-question quiz, about three minutes long, covering four dimensions of how someone uses AI tools, that returns a score out of 100 plus a percentile against people in your job and country: couragehorizon.com/agni.
Common questions
Does telling ChatGPT to “think step by step” fix a vague prompt? It helps with reasoning-heavy tasks but doesn’t replace a role or a format. That instruction changes how the model works through a problem, not what it assumes about who’s asking or what the output should look like, so a step-by-step instruction stacked on an unspecified role and format still tends to land somewhere generic.
Is a longer prompt always a better prompt? No. PCWorld’s reporting on the same issue points out that an over-specified, overly narrow prompt produces an answer that misses the bigger picture just as easily as a vague one does, so the fix is coverage of the same three gaps, role, situation, format, rather than sheer length.
Is there a way to measure this instead of guessing? Yes. Agni’s AI Literacy Score checks exactly this, scoring how someone instructs AI tools as one of its four dimensions and returning a percentile against people doing similar work, so instead of guessing whether a run of weak answers is about prompting, there’s a number attached to it.
Sources
- PCWorld, “Bad ChatGPT answer? Maybe you’re asking the wrong question”, 2026.
- How-To Geek, “6 Ways to Get Better Responses From AI Chatbots”, 2023.