Why AI Gives You Random Results (and the 6-Part Fix)

Tomasz Dylik
Jul 19
4 min read

Updated: Jul 30

Most people don't get bad results from AI. They get random ones.

You write a prompt, the answer looks great, and you move on. Next day you write what feels like the same prompt, and the output is flat, off-topic, or just different enough that you can't use it. Nothing broke. You didn't get worse at writing prompts overnight. The problem is that most prompts leave too much to chance, and the model fills the gaps differently every time.

Over a hundred of you wrote back when I asked about your single biggest frustration with AI, and thank you for that. The answers clustered around three things, and all three come from the same root cause:

25% said “I never know how to phrase it.”
14% said “It gives me a different result every time.”
14% said “It just invents answers I can't trust.”

Phrasing, consistency, trust. Three complaints, one fix.

Why the same prompt gives different answers

A language model doesn't read your mind. It reads your words and fills in everything you left unsaid, using its best guess. Every unstated assumption is a place where the output can drift.

Ask it to “write a post about our new feature” and you've left almost everything open: who's reading it, what voice to use, how long it should be, what it's actually supposed to achieve, and what it should avoid. The model picks a lane for each of those, and picks a slightly different lane next time. That's where the randomness comes from. It isn't the model being unreliable. It's the prompt being underspecified.

The fix isn't a longer prompt or a magic phrase. It's a structure: a small set of inputs you fill in every time, so the model has nothing left to guess.

The 6 inputs behind a reliable prompt: CROFTC

CROFTC is six parts. Fill them in and you turn prompt roulette into something repeatable.

C · Context. What the model needs to know before it starts: the situation, the product, the audience, the background. Most weak prompts skip this entirely and the model improvises the setup.

R · Role. Who the model should act as. “A senior email copywriter” and “a helpful assistant” produce very different work. Naming the role narrows the whole response.

O · Objective. What this output is actually for. Not the topic, the goal. “Get readers to click through to the free toolkit” is an objective. “Write about prompts” is just a subject.

F · Format. The shape of the answer: a 5-bullet list, a 150-word intro, a table, a subject line plus body. When you don't specify the format, you're accepting whatever shape the model defaults to.

T · Tone. Professional, friendly, bold and punchy. Tone is what makes output sound like you instead of like a generic bot, and it's the first thing that drifts when you leave it out.

C · Conditions. What has to be true of the answer, both what it must include and what it must not. No jargon, no made-up statistics, stay under 200 words, don't mention competitors, and cite only sources you were given. This is the one most people forget, and it's the one that stops the model from inventing things you can't trust.

Notice how the six map straight back onto the three complaints. Context and Role fix “I don't know how to phrase it,” because the structure tells you exactly what to write. Format and Tone fix “a different result every time,” because the same inputs give the same kind of output. Conditions fix “it invents answers I can't trust,” because you tell the model what has to be true of the answer before you would accept it.

Want to try this without building prompts from scratch? Grab my free bonus pack: 500+ ready-to-use prompts plus the one-page CROFTC cheat-sheet. Get the free prompt pack.

What it looks like in practice

Here's a typical prompt:

Write an email about our summer sale.

And here's the same request run through CROFTC:

Context: We're a small skincare brand running a 20% summer sale for one week, aimed at existing customers who've bought before. Role: You are an email marketing specialist who writes for repeat buyers. Objective: Get readers to click through and buy again before the sale ends. Format: Subject line, plus a short body under 120 words, ending in one clear call to action. Tone: Warm and direct, not pushy. Conditions: No fake urgency, no invented discounts, don't mention any product we haven't sold before, and every claim about the sale must match the terms above.

The second one isn't longer because it's padded. It's longer because it's complete. Run it today, run it next week, hand it to a teammate, and you'll get the same quality every time. That's the whole point: the result stops depending on how you happened to feel when you typed it.

Start using it today

You don't need to memorize anything. Keep the six letters somewhere you can see them and treat them as a checklist. Before you hit enter, ask yourself whether you've given the model its Context, Role, Objective, Format, Tone, and Conditions. If one is missing, that's the gap where your next random result is going to come from.

CROFTC is the structure behind every prompt in my series, and it is the reason they stay consistent page after page. If you want the whole system rather than one piece of it, that is what I wrote AI Prompt Engineering Bible (7 Books in 1) for. Seven books in one volume, each of the six inputs taken apart and rebuilt, with ready-made prompts you can copy and adapt to your own work.

Get the paperback on Amazon →