Most people clean up before they talk to an AI. They organize their thoughts, structure the request, smooth out the contradictions. It feels like good practice — hand the model something polished and you’ll get something polished back.
This instinct is mostly wrong, and it costs you the most useful thing the model could actually do for you.
What cleaning up actually removes
When you tidy a brief before prompting, you’re not just removing noise — you’re making decisions. Which details matter. Which tensions are worth surfacing. Which directions are worth exploring. Those decisions are exactly what you might want help with. By the time you’ve written a clean, organized prompt, you’ve already done the cognitive work. The model is now executing your plan, which is fine, but it’s a much more limited use of the tool.
Messy input forces the model to do something more interesting: find structure you haven’t imposed yet. It will identify which parts of your notes are actually in tension, which threads look promising, what’s missing. A clean prompt forecloses that, because clean signals that the structure is already settled.
This shows up directly in how the topic generation step in our blog pipeline works. It doesn’t start with a polished brief — it starts with past posts, rough directional notes, and a fairly loose instruction. A tightly specified brief just replicates what we’d already decided. The looseness is what gives the model room to surface something we hadn’t considered.
What “raw” actually means in practice
Raw doesn’t mean unformatted stream of consciousness dumped into a text box. It means not pre-solving the thing before you hand it over.
A useful raw input might look like: everything you know about a problem, including the parts that contradict each other. Things you’re unsure about, in no particular order. Notes from a meeting or reading session you haven’t processed yet. A half-formed argument where you can see both sides and haven’t picked one.
What you’re really doing is handing the model a thinking task rather than a production task. “Here’s what I have — help me figure out what it means” is a different job than “here’s what I’ve concluded — write it up for me.”
Both are valid. But most people default to the second without realizing the first is available.
The specific case where this matters most: decisions
The highest-value version of this is decisions with genuine trade-offs. Not “write me a pros and cons list for X” — that’s just asking the model to format something you could format yourself. The more useful move is to dump the actual conflicting signals: what your gut says, what the numbers suggest, the constraint you’re trying not to name, the option you’re leaning toward but feel slightly bad about.
When I’ve done this with real operational calls — scope decisions, prioritization questions — the model consistently does something a clean brief wouldn’t allow: it names the actual tension rather than listing factors. “It sounds like the real constraint you’re working around is X, not the one you framed it as” is genuinely useful. You can’t get that if you’ve already resolved the framing in your prompt.
One honest caveat: the model can misread the structure of messy notes, especially when a lot of threads are competing for attention. It will sometimes amplify the wrong signal. So raw input isn’t a substitute for reading what comes back carefully — it just changes what you’re reading for. You’re not checking whether the output matches what you said; you’re checking whether the model found something real in the mess.
When clean prompts are still right
There’s a class of tasks where cleaned-up input is clearly better: anything where the model is executing a precise specification. Format this JSON. Convert these units. Rewrite this paragraph in a voice you’ve already defined. For those, vagueness introduces errors. The model needs constraints to execute against, not latitude to interpret.
The rough dividing line: if you already know exactly what the output should look like, clean up the prompt. If you’re trying to figure out what the output should look like, don’t.
Most people apply the clean-input discipline uniformly, including to the second category. That’s where the opportunity gets missed.
A concrete way to try this
Pick something you’ve been avoiding thinking through — a plan that feels half-formed, a project with unclear scope, a decision you keep deferring. Open a blank prompt. Write everything you actually think about it: the contradictions, the uncertainties, the option you’re not sure you should even mention. Don’t organize it. Don’t resolve anything before you type it.
Then add one line at the end: “What’s the real question here?”
What comes back is often more useful than anything a structured brief would have produced — because the model is doing work you hadn’t already done for it.
The version of this worth sitting with: the tasks where you feel most prepared to write a clean prompt are probably the ones where you’ve already done too much of the thinking yourself.
