Most bad AI output isn’t a model problem. It’s a prompt problem.
I was talking with a fellow coder friend of mine named Matthew Breau about prompting. He and I have both been using Claude Code and AI tools in general for about ten months now, roughly as long as most coders I know.
Matthew pointed something out to me recently. Most people don’t know how to prompt. “Make me look pretty,” someone types into ChatGPT. The AI returns a picture of them wearing red lipstick.
The primary problem most people have with prompting is understanding context. As part of AI prompt eval work I’ve been learning, I’ve spent time reviewing broken prompts and figuring out why they break. The failures fall into a deterministic set of categories.
Name these problems, and you can fix them. Most people using AI today don’t understand the fundamentals of how it works. So over the next five posts, I’m covering the basics: how to prompt, context windows, and the rest of the foundation everyone skips.
Over the next month and a half, I’ll cover the foundational pieces of understanding LLM prompting. This is worth your time whether you’re writing the code, doing the business analysis, or working in strategy.
What’s Missing from Most People’s Prompts
No role or context
A prompt that jumps straight into a task usually skips the role and the context. It never says who the AI should act as or what the output is for. The model fills that gap with a guess.
I saw this play out with a founder testing an AI tool for customer emails. She typed “respond to this complaint.” The output read like a form letter. She added one line: “You’re a support rep for a boutique skincare brand, warm but professional, and this customer has been loyal for two years.” Same complaint, same model. The second response actually sounded like it came from a person who cared.
Vague instructions
“Make it better” doesn’t mean anything to a model. Better how? Shorter? More formal? More persuasive?
I’ve watched people burn ten minutes rerunning the same vague request and getting frustrated with the results. The model isn’t being difficult. It genuinely can’t infer taste it hasn’t been told. Swap “make it punchier” for “cut it to three sentences and open with the strongest claim,” and the model has something to actually execute on.
No output format
If you don’t say how you want the answer structured, you’ll get whatever the model feels like giving you.
A colleague asked an AI tool to “list the risks in this contract.” It came back as three dense paragraphs. She wanted five bullet points she could forward to her team. The fix took four words: “Give me five bullet points.” Same question, same contract, completely different usefulness.
Missing examples
If you want a specific tone or structure, show one.
I ran into this building an email-drafting workflow. Telling the system to “sound conversational but professional” produced generic output every time. Feeding it three real emails I’d actually sent, and telling it to match that voice, got me something usable on the first try. A model can describe a style in the abstract. It’s far more reliable when it has something concrete to match against.
Conflicting instructions
“Be concise” and “cover everything in detail” cannot both be true.
I see this constantly in prompts people paste together from two different goals without noticing they fight each other. Read your prompt back before you send it. If two lines pull in opposite directions, pick one. The model will try to satisfy both, and the output suffers for it.
No success criteria
If you don’t know what a good answer looks like, the model doesn’t either.
This one’s easy to miss because it feels implicit. It isn’t. Say what you’re optimizing for. Accuracy. Brevity. Getting someone to click a link. A prompt with a stated goal produces a sharper answer than one that just describes a task.
Doing too much at once
One prompt tries to summarize, analyze, and recommend, all in a single pass. Each of those is a different task with a different shape of good output. Stacking them means none of them get done well.
I learned this the hard way building a research pipeline. One giant prompt asking for a summary, a risk assessment, and next steps produced mediocre versions of all three. Splitting it into three separate steps, each one feeding the next, produced three genuinely good outputs.
Why this matters more now
Most people skip all of this, then wonder why the output feels generic. The model isn’t bad. You never told it what you wanted.
That’s the whole lipstick problem. Matthew’s colleague wanted to look better. She typed “make me look pretty,” and the AI gave her the closest literal read of that: lipstick. Not because the AI is dumb. Because it had nothing else to go on.
A year ago, most people’s AI use was casual. A quick question, a draft, a summary. The stakes of a bad prompt were low, because the tasks were low.
Now, people are building workflows, automating real work, and making decisions off AI output. A vague prompt used to cost you a mediocre paragraph. Now it can cost you a wrong analysis, a broken workflow step, a decision made on bad information.
Prompting well isn’t a nice-to-have anymore. It’s as basic as writing a clear email. Skip the context, skip the specifics, and you’ll keep getting lipstick on a pig instead of the thing you actually needed.