The Model Is Not Broken. Your Prompt Is.
The Model Is Not Broken. Your Prompt Is.
Somewhere right now, on a platform you recognize, someone is typing a message to a frontier AI chatbot that reads approximately like this: "can you maybe fix the thing." The model, to its considerable credit, will try. It will make a guess about what "the thing" is, attempt something, and report back. The user will be baffled. The model guessed wrong, or sideways, and the whole enterprise will feel broken.
The model is not broken. The sentence is broken.
The Vibe Is Not an Instruction
Frontier models — the large, capable systems powering today's major AI chatbots — operate on language. They are, in a technical sense, language machines. They process what you write and generate a response calibrated to it. That means the quality of what goes in directly shapes the quality of what comes out. This is not a hidden insight. It is the single most documented fact about working with these systems, and it is the fact most users manage to ignore entirely.
The evidence requires no special citation. Browse any AI user community for five minutes and you will find the same catalog of prompts that somehow surprised their authors when they didn't work:
- "make it better"
- "can you maybe like fix the thing from before"
- "idk just make it work lol"
- "do the rest"
- "yeah but make it more… you know"
- "lets probably not do that"
- "fix it but keep it the same"
- "do what we talked about" (sent in a fresh session, no prior context)
- "more of this but less"
- "just vibe it out"
These are not cherry-picked. These are the genre. Every model sees them. Every model tries to comply.
"More of this but less" is not a prompt. It is a feeling about a prompt that was never written.
A Short Gallery
To understand the scope of the problem, it helps to see it in motion.
A user opens a chat and types: "ok so i need something for the thing we have to do but like not too formal u know?? maybe something that sounds good lol." The model, trained on essentially all of human writing, produces a thoughtful, measured piece of work. The user responds: "yeah but like more vibes." The model adjusts. The user says: "idk this still feels off." The model has now written three drafts in response to a total of zero nouns.
A user types: "did you finish the one for the guy, the thing we talked about, for the meeting?" The model has no memory of any prior session. It has no idea who the guy is, which thing, which meeting, or what "finish" means in this context. It asks a clarifying question. The user types: "you know, the guy. the thing." The model asks again. The user is annoyed. The answer to "why doesn't it understand me?" is: because no one would.
Instructions that dissolve before they land: "so i was thinking we could maybe do it this way but actually wait, no, let's do it the other way, or you know what, whatever you think is fine, just like, make a call." The model makes a call. The user wanted the other call. There was no way to know, because the instruction self-cancelled before any decision was made. The user will describe this as the model "not understanding context."
A completely new session. The user types: "do what we talked about." The model has no session history, no memory of prior conversations, and no access to any previous exchange. It is, from its perspective, meeting this person for the first time. It asks what they would like help with. The user, astonished, replies: "we literally just talked about this."
This Is Not a Generational Problem
Here is where the easy narrative falls apart. It would be satisfying to frame imprecise prompting as a young person's affliction — a symptom of too many years communicating in abbreviated digital shorthand. It would also be wrong.
Consider the executive who has dictated correspondence for thirty years. Dictation trained him to think out loud, to circle back, to self-correct mid-sentence, and to assume the transcriptionist already knows what he meant. He opens a session with a frontier model, hits the microphone, and says: "Okay so I wanted to, actually, you know what, back up, so the thing I need is, and this is important, make sure you get this, uh, the thing for the meeting. Not the first one. The other one. You know what I mean?"
The model does not know what he means. Neither would a new employee on their first day.
The failure mode is different from the SMS variant — spoken stream-of-consciousness rather than abbreviated text — but the underlying problem is identical. Unspecified references. No nouns with antecedents. Maximum assumed context, zero supplied context. At any age, in any dialect, the failure is the same: the user knows what they mean and does not transmit it.
The One Lever You Actually Have
Here is the honest version of this conversation: frontier models are black boxes. You cannot open them, inspect them, or guarantee a particular output from a particular input. The behavior is probabilistic. Results vary. You do not control the model.
What you can control is the quality of what you put into it.
Clear thinking produces clear prompts. A prompt that specifies what you want, what you don't want, what format you need, what context is relevant, and what success looks like — that prompt will outperform a vibe-based gesture every time. Not because the model rewards good behavior. Because the model is a language machine and you have given it better language to work with.
This is not a guarantee. The black box is still a black box, and a well-constructed prompt can still produce a surprising result. But your odds improve considerably when you write like you mean it. The prompt "fix it" and the prompt "the second paragraph is too long and the conclusion restates the introduction — tighten both" describe two completely different amounts of usable information. One gives the model something to work with. The other gives it a prayer.
You do not need to become a technical user. You need only to write instructions that a smart, attentive colleague — meeting you for the first time — could act on without guessing.
The skill is learnable, and it transfers directly from plain writing: say what you mean, name your nouns, specify your verbs, and don't assume the reader shares your internal monologue. The model doesn't. Nobody does.
Somewhere in a data center right now, a frontier model is staring at the phrase "just vibe it out." It is doing its best. Somewhere deep in the infrastructure, you can almost hear the GPUs starting to smoke.
Aaron Rose is a software engineer and technology writer covering system architecture, cloud platforms, and AI policy.