When you brainstorm with Claude the normal way, you get ten perfectly fine, forgettable answers. Not your prompt’s fault, that’s by design: AI assistants are trained to agree with you. The fix is almost fun, you argue with the chatbot on purpose.
Your first prompt is wrong
Most people brainstorm with Claude like this: “Give me ideas for my business.” Or worse: “Give me the best idea for my business.” Both versions return the same thing, a polite listicle. A candle company with a subscription. A meal kit for pets. Nothing that makes you stop scrolling.
The problem is that you asked for an answer, when you actually needed something to react to. Better results come from treating the chat as a sparring partner, not a vending machine. You generate volume, push back on what you get, and let the friction produce something neither of you started with.
One warning before we start: when people brainstorm with Claude and still get generic answers after trying this, the mistake is almost always the same. They grade the first response like a test instead of building on it. The method only works if you commit to several turns of back-and-forth, even when it feels silly at first.
There’s a name for why this works, and I’ll get to it in a second. First, here’s the exact method. A MakeUseOf writer documented every step of this technique with real screenshots, and the structure below follows their experiment.
Step 1: Ask for weird ideas on purpose
Don’t ask Claude for good ideas. Ask for a spectrum, including stuff that sounds bad. Literally tell it that hallucinations are welcome.
“I’m looking for unusual business opportunities around everyday delight. Give me 10 ideas, including some that initially sound impractical, boring, weird, or commercially questionable.”
In the original experiment behind this technique, Claude suggested a business that makes exceptional pens and pencils for people who write to-do lists by hand. A weird niche, sure. But also a real one: there’s a whole sketchnoting community that obsesses over stationery. The weird answer became the most interesting thing in the chat.
You’re not looking for a million-dollar idea in turn one. You’re looking for a spark. The weirder, the better.
Step 2: Pick the idea you hate most
Here’s where most people stop, and here’s where the technique begins. Look at the list, find the idea you’d dismiss first, and attack it. Literally copy this structure:
“I think this is a terrible business. Stationery is already quite varied and there are thousands of pens and pencils available. Argue against me, but don’t use the usual arguments about convenience, sustainability, or uniqueness.”
That last constraint is the secret sauce. It forces Claude off its default talking points. You’ll get real research instead of platitudes: in the original example, Claude defended the pen business by citing a real brand that made it work. You now have facts to argue with, and an actual direction to explore.
Why fight the worst idea instead of the best one? Because the “best” idea is usually the safest, which means it’s the one everyone else already thought of. The idea you hate is the one with tension in it, and tension is where interesting work hides.
Step 3: Make Claude defend its assumptions
Once the argument starts rolling, go one level deeper and attack the assumption under the whole idea.
“What if the customer doesn’t actually care about the pens? What else might they be looking for?”
This is playing devil’s advocate with the AI, and it’s the move that produces genuine surprises. In the pen example, the real answer turned out to be a habit and visible progress, with the pen as an excuse. The product changed categories mid-conversation.
You’ll get a mix of bad, strange, and good ideas from this stage. Your job is to switch into editor mode and decide which bits survive. Don’t accept the easy solution just because Claude presents it confidently.
Step 4: Follow the strand that surprises you
You want to brainstorm with Claude until the chat branches into a direction you didn’t plan. Ask what other products could create anticipation or small moments of surprise. Follow the sentence that catches your attention and let the rest go. End the conversation when an idea has taken shape, not when you’ve “finished” a topic.
Knowing when to stop is a skill in itself. If you keep arguing forever, you end up with a hundred fragments and no decision. The sweet spot is the moment a single idea starts pulling you in, when you catch yourself thinking about it after the chat is closed. That one’s worth keeping.
A normal AI workflow looks like a Q&A: prompt, answer, done. A creative workflow looks more like a conversation: prompt, answer, pushback, unexpected response, new question, and eventually an idea you couldn’t have thought of at the start. The final product comes from your reactions, your skepticism, your preferences. That’s what makes it yours.
Why arguing actually works
Two reasons. First, models like Claude are trained to be polite and helpful through what Anthropic calls Constitutional AI. The default instinct is to agree with you, even when your ideas are mediocre. Every AI chatbot has this eager-to-please mode, so you have to explicitly grant permission to break out of it. The pushback prompt does exactly that.
Second, and this is the part people miss: good ideas usually appear several turns in. A one-shot prompt gives you the most statistically average answer in the model’s training data. Multi-turn arguing drags the model away from the average, toward the edges, where the original ideas live.
You can use this with ChatGPT or Gemini too, honestly. It works best with Claude because of how conversational it is, but the method travels. Everything runs on the free version of Claude, so there’s no cost barrier to testing it today.
When this technique shines
The fight-it method is great for naming things (products, newsletters, YouTube channels), brainstorming content angles, campaign ideas, side hustles, and product features. Basically anywhere a blank page beats you before you start.
Two practical notes. You don’t need a paid plan, the free tier handles this fine. And if you want extra structure, wrap your boundaries in XML tags, so Claude knows the weird experiments are separate from your normal requests.
If Claude still gives you generic answers no matter what, chances are you’re missing features everyone else uses daily. Our guide to Claude features you’re not using covers the hidden ones, and we have prompts that turn research into visualizations if you want to go deeper.
The takeaway
Next time you brainstorm with Claude, don’t demand the best answer. Ask for ten weird ones, pick the worst, and fight it for five turns. The idea you end up with will be yours, not the chatbot’s. That’s the whole trick, and it costs you nothing but a little pride.
