// writing /04 · prompt engineering · llms · ai literacy

It's not you, it's the prompt. Most people use ChatGPT like Google. That's the bug.

ChatGPT is not a search engine. It is a Waze for words. The address you type is the prompt. Vague address, weird route. Specific address, the answer you actually wanted.

By Michael Fleicher · PrincipalJul 10, 2025· 5 min readUpdated: June 2026

prompt engineeringLLMsAI literacyChatGPTAI workflowprompt engineering 2026

A ChatGPT-generated image of a golden retriever in an autumn park - the punchline of a post about specific vs vague prompts. — Same model, same request: "a dog in a park." One prompt got something generic. The other got this. The difference is the entire skill.

A short answer first. When ChatGPT writes nonsense, it is almost never the model. It is the prompt. People ask an LLM the way they ask Google - three keywords, half a thought - and then complain the answer is "generic." It is generic because you gave it a generic address. Prompt engineering is the skill of talking to AI in a way it can actually work with: a goal, a style, examples, constraints. It is not coding. It is not magic. It is the difference between "a dog in a park" and getting back a stock photo, vs. one specific sentence that gets you an image you actually wanted to send to someone.

If you have ever closed a ChatGPT tab thinking "this thing is overhyped," there is a decent chance you were not asking it the right way. In 2026, with models this capable, the model is not the bottleneck. You are.

AI is not Google. Stop prompting it like Google.

Google was trained on a simple deal: you type keywords, it returns links. The fewer the keywords, the broader the net. "Best pizza Tel Aviv" works because Google is matching documents.

An LLM is not matching documents. It is generating the next word, then the next, then the next, based on what you said. The instruction is the steering wheel. If the instruction is vague, the steering wheel is vague. The car still drives - it just drives somewhere generic.

That is why "write me something about marketing" gets you a Wikipedia-flavored paragraph, and "write me a 120-word LinkedIn post in a dry, builder tone, opening with a contrarian claim about B2B marketing, no emojis, no "in today's world"" gets you something you would actually publish.

Same model. Same minute. The prompt did all the work.

What is actually happening behind the scenes

Quick, intuitive version. No math, promise.

Think of the AI as a Waze for words. Your prompt is the address. The model is the navigator.

→Every word you type gets converted into a number. Actually, into a vector of thousands of numbers - a coordinate in a giant map of meaning.
→For every next word, the model computes "given everything so far, what is the most likely next word?" and picks from a probability distribution.
→A vague address ("a dog in a park") puts the navigator somewhere in a huge city of possible answers. It has to guess which neighborhood you actually meant.
→A specific address ("a golden retriever, autumn, blue bandana, golden-hour light") snaps the navigator to one block. The route gets short. The output gets sharp.

Every small improvement in how you phrase the request saves the model kilometers of trial and error, and saves you minutes of editing the output until it stops being embarrassing.

That is the entire mechanic. Once you internalize that the prompt is an address, you stop arguing with the model and start writing better addresses.

The one example that makes it click

Open an image model. Try this:

Prompt A: draw a dog running in a park

You will get something generic. Some kind of brown dog. Some kind of grass. A pose your eye slides off.

Now try this:

Prompt B: A golden retriever with a blue bandana running through an autumn park, leaves flying around, dynamic motion blur, golden hour lighting.

You will get a different image. Sharper. Alive. Like it was art-directed. Because it was - by you, in one sentence.

Both prompts asked for "a dog in a park." Only one of them told the model which dog, which park, which light, which mood. That is the entire skill compressed into one before-and-after.

You can run the same experiment in five minutes. Go do it. It is more convincing than any explainer.

A four-part frame for any prompt

If you do nothing else, structure your prompts around four things. In any order, in any language:

01Goal. What is the output for? A LinkedIn post? An email to a CFO? A bullet list for your own notes? A model that does not know the goal will pick the most boring possible version of it.
02Style and voice. Formal, dry, founder-blunt, customer-service polite, journalist, teacher. One adjective costs you nothing and changes everything.
03Examples. One or two short examples of "the kind of thing I want" beats a paragraph of description. Paste a sentence in your voice. The model will copy the cadence.
04Constraints. Length, format, things to avoid, audience. "Under 200 words, no emojis, no 'in today's fast-paced world', third person." Constraints are not limits - they are the rails.

That is the whole framework. Goal, style, examples, constraints. People sell six-week courses around this. The courses are useful. The framework fits on a sticky note.

Why this is the highest-leverage skill of the next five years

Imagine a superpowered intern who can do almost any cognitive task you give them - draft, summarize, analyze, brainstorm, code, design - but only if you brief them properly. That is the deal with current LLMs. The intern is sitting in the room. Most people are mumbling at them.

People who can brief well get 10x the output in 10x less time. Not a marketing number - an actual one, if you compare a sloppy prompter to a deliberate one on the same task. This is already showing up in meetings, hiring, sales, content, support, product, and engineering. The gap between "uses AI" and "uses AI well" is now wider than the gap between "uses AI" and "does not."

And no, you do not need to be technical. You do not need to understand transformers. You need to be specific. That is it.

The next time ChatGPT writes garbage

Stop before you write "this thing is broken." Re-read your own prompt the way the model read it. Not "what did I mean?" but "what did I actually say?"

Nine times out of ten you will find the bug in your own sentence. Add the goal. Add the style. Add an example. Add a constraint. Hit send again. The output will be different. Usually a lot.

It's not you, but it kind of is. It's the prompt.

FAQ

Q

Is "prompt engineering" actually engineering?

No, and yes. There is no code, no compiler, no algorithm you have to write. In that sense the name oversells it. But it is engineering in the older sense: deliberately shaping an input to get a reliable output from a complex system you do not fully control. Same instinct as designing a good search query, a good brief for a designer, or a good ticket for a developer. The "engineering" word is just there to remind you it is a craft, not a vibe.

Q

Do I need to learn special syntax or commands?

No. LLMs respond to plain language. You do not need "magic words," role-play tricks, or three-paragraph "you are a world-class expert" preambles to do basic work. Be clear about the goal, the style, the audience, and the constraints. That covers 90% of everyday prompting. The advanced techniques exist, but most people are leaving 80% on the table before they ever need them.

Q

What is the single biggest mistake people make?

Treating the prompt like a Google search. Three keywords, hit enter, judge the model on the result. LLMs are not retrieving documents - they are generating an answer based entirely on what you wrote. If you give them one line, you get the most average possible interpretation of that line. Give them five lines of context and the output jumps a category.

Q

How do I get better at this quickly?

Two habits. First, when an answer is bad, do not rewrite it yourself - rewrite your prompt and send it again. You will learn faster from the diff than from editing the output. Second, save your good prompts. The ones that worked for emails, summaries, posts, code reviews. Reuse them. Over a month you will build a personal library that is more valuable than any course.

Q

Does this apply to image and video models the same way?

Yes, even more so. Text models can fill in gaps with general knowledge. Image and video models are extremely literal: if you do not specify the light, the angle, the style, the medium, the mood - you will get the model's default, which is usually some glossy, generic stock-photo flavor. The dog-in-a-park example in this post is the canonical demonstration. Specific words equal specific pixels.

Q

Is prompt engineering going to be obsolete when models get smarter?

Partly. Models will get better at guessing your intent from less. They are already doing that compared to a couple of years ago. But "telling a capable system what you actually want, clearly" is not going away. It is the same skill that makes you good at briefing a designer, scoping a project, or writing a ticket a junior engineer can ship. The interface will get more forgiving. The skill underneath it will not.

Written by Michael Fleicher, Principal at Bina Labs. Two-time CTO. We build AI systems that actually run in production, and we train the humans around them to brief, evaluate, and trust them. If you want your team to stop fighting their LLM and start using it like a coworker, start here.