You typed something into ChatGPT. It came back bland. Generic. The kind of output you could have predicted before hitting enter. So you typed "be more creative" and hit send again. The next response was slightly different — maybe a little more energetic, maybe with a few unexpected word choices — but still fundamentally the same thing you got the first time, just wearing a different shirt.
This is one of the most common frustrations people have with AI, and it is entirely fixable. But the fix is not what most people think it is.
What "be more creative" actually tells the model
When you tell an AI to be more creative, you are giving it an instruction with no definition. Creative compared to what? Creative in what direction? More surprising? More poetic? More unconventional in structure? More emotionally resonant? "Creative" is not a style — it is an aspiration. And aspirations without specifications produce exactly the kind of output you already got, except with slightly more exclamation points.
Language models work by predicting what text should come next based on patterns learned from enormous amounts of human writing. When you ask for "creative" output without further definition, the model defaults to the statistical center of what "creative" writing looks like across all the text it has ever processed. That center is usually competent, readable, and completely forgettable — because the most common kind of creative writing is exactly that.
The model is not withholding creativity from you. It genuinely does not know what you mean unless you tell it.
Specificity is the actual mechanism
The writers, marketers, and developers who consistently get great AI output are not using magic prompts. They are being specific in ways that non-technical users are not. Here is what that actually looks like in practice.
Instead of "be more creative," they say things like:
- "Open with a counterintuitive claim that the reader will want to argue with."
- "Write in short, punchy sentences. Maximum 12 words per sentence."
- "Use an unexpected analogy from an unrelated field to explain this concept."
- "Avoid every adjective that could appear in a press release."
- "Write the way a smart friend would explain this over coffee — not the way a consultant would present it."
Every one of these is specific enough that two different writers could execute them and produce something recognizable as following the instruction. "Be more creative" is not specific enough for one writer to execute reliably, let alone a language model.
The three types of specificity that actually work
When people get reliably better AI output, they are usually being specific in one of three ways — and often all three at once.
Structural specificity tells the AI how the output should be organized. "Three short paragraphs," "open with a question, answer it, then complicate the answer," "end with one concrete action the reader can take today." These are not creative instructions — they are architectural instructions. But they produce more interesting output because they prevent the model from defaulting to its standard five-paragraph-essay structure.
Voice and tone specificity tells the AI how it should sound. Not "professional" or "casual" — those words mean almost nothing. Instead: "Write like you're slightly frustrated with how overcomplicated this has gotten." Or: "The tone should be warm but direct — the kind of thing your most competent friend would say." Or: "No hedging language. No 'it's worth noting that' or 'it's important to remember.' Just say the thing." These instructions give the model something it can actually execute.
Negative specificity tells the AI what to avoid. This is the most underused type and often produces the biggest improvement. "No buzzwords." "Don't use the word 'delve.'" "Avoid starting any sentence with 'I'." "No metaphors involving light, journeys, or bridges." Telling the model what not to do clears away the statistical defaults and forces it toward less common paths — which tend to be more interesting.
A quick test you can run right now
Open ChatGPT or Claude. Ask it to write a short paragraph describing why a particular habit is hard to break. Read the output. Note how many of the following appear: the word "important," a sentence starting with "It's," a closing sentence that begins with "By," the phrase "take the first step," or any form of the word "journey."
Now ask again with this added to your prompt: "No clichés. No motivational language. No sentences starting with 'It's.' Write like you're explaining this to someone who has heard every piece of advice about this topic and is skeptical of all of it."
The second output will be noticeably different — and almost certainly better. Not because you asked for creativity, but because you blocked the default paths and forced the model to find a different route.
The practical version of this
You do not need to understand language model architecture to use this. You just need one habit: before hitting send on a prompt that is asking for any kind of writing, add at least one specific instruction about what you do not want, and at least one specific instruction about structure or tone.
That combination — one negative constraint and one structural or tonal direction — consistently produces output that feels like it came from someone who actually thought about the task. Which, in a sense, it did. You just had to think about it first.
The AI is not the bottleneck. Your prompt is.