How I Brainstorm With AI Without Getting Generic Ideas

The first time I used AI to brainstorm I came away convinced the tool was useless for creative work. I asked for "ten ideas for a launch tweet for Super Prompts." What I got back was the same ten tweets every SaaS founder has seen — "Tired of losing your prompts? 🚀 Introducing…" followed by nine variations with different emojis. Technically ten ideas. Functionally one idea, repeated.

I almost stopped using AI for ideation entirely. The problem wasn't the model. The problem was that "give me ideas for X" is the worst brainstorm prompt you can write, and it took me months to figure out why.

The short version: when you ask for ideas with no constraints, the model gives you the *mean of the internet*. It returns the average of every example it has seen for that task. The average is generic by definition. To get ideas worth using, you have to make the model travel away from the mean — and that's a job constraints do, not enthusiasm.

Below are the four constraints I now add to every brainstorm prompt, and three real sequences I run when I need actual ideas.

Why "give me ideas for X" returns sludge

A model trained on the internet has seen ten thousand "launch tweets for a SaaS." Most of them are mediocre, because most launch tweets are mediocre. When you give it no further direction, it samples from the dense middle of that distribution — the tweets that look the most like all the other tweets.

This is not a flaw. It's exactly what you asked for. You asked for the most likely thing.

The trick is that you almost never *want* the most likely thing. You want the unusual angle, the framing nobody else is using, the idea that makes someone stop scrolling. That lives in the tails of the distribution, not the middle. Getting the model into the tails takes intentional pressure — and the pressure shape that works is constraint, not encouragement.

"Be more creative" does almost nothing. "Be creative" is in the middle of the distribution too. But "no abstract nouns, no exclamation marks, must reference a physical object" forces the model into a corner of the space where the average tweet doesn't live. That's where the good ideas are.

The four constraints I add to every brainstorm prompt

Every brainstorm prompt I write now includes at least three of these. Most include all four.

1. A banlist. The phrases, framings, and clichés to actively avoid. "No 'transform your workflow,' no 'unlock your potential,' no 'in today's fast-paced world,' no emojis, no rhetorical questions." This is the single biggest unlock. You're not telling the model what to do — you're telling it what *not* to do, which forces it out of the dense middle.

2. A reference anchor. A person, brand, or piece of writing whose style I want to borrow. Not "be witty" — "write the way [specific person] writes on Twitter." The model has read them too. Anchor it to a specific style and the output stops being generic.

3. A structural constraint. A non-negotiable shape: word count, sentence count, must-contain element, must-not-contain element. "Exactly 12 words. No verbs in the first three." Structure is a corner of the space. Corners produce interesting output.

4. A "what it sacrifices" question. I ask the model to tell me what each idea gives up, not just what it does. This forces it to differentiate. If five ideas would all sacrifice the same thing, they're variations of the same idea — and I want it to throw four of them out before I see them.

You can stack these. You should stack these. A prompt with all four reliably produces ideas I'd actually use; a prompt with none reliably produces sludge.

Sequence 1: Launch tweets

This is the prompt I now run instead of "ten ideas for a launch tweet."

I'm writing the launch tweet for Super Prompts, a tool that lets creators save and rediscover AI prompts. Give me five tweets that fundamentally differ from each other — not five rewrites of the same angle. Constraints: under 25 words each, no emojis, no rhetorical questions, no "introducing" or "tired of," no abstract claims like "transform" or "unlock." Each must reference one concrete thing (a moment, a number, an object, a feeling). For each, tell me what kind of reader it sacrifices and what kind it attracts.

The output looks nothing like the sludge from before. The model returns one tweet built around a specific moment ("the prompt I wrote at 2am that I'd kill to find again"), one built around a number ("I have 184 saved prompts. I use 6"), one built around a feeling ("the worst kind of déjà vu is rewriting a prompt you already nailed"), and so on. Each one has a real point of view. Each one excludes a different audience.

The "what it sacrifices" line is what makes me pick. It's not "which tweet is best?" — it's "which sacrifice am I willing to make?"

Sequence 2: Product feature ideas

When I'm thinking about what to build next, "give me ten feature ideas for Super Prompts" returns the same ten features every productivity SaaS has on its roadmap. Folders, tags, sharing, AI suggestions, dark mode. The list is correct and useless.

The prompt I run instead:

I want feature ideas for Super Prompts that would only make sense if you'd lost a great prompt and felt the specific pain of trying to remember it three weeks later. Forget generic productivity features — folders, tags, dark mode, AI suggestions. I'm specifically looking for ideas that fall into one of these categories: (1) features that fight memory failure, (2) features that turn rediscovery into a habit, (3) features that would feel weird at first and obvious after a month. Five ideas total. For each, describe the specific user moment it solves, not the feature spec.

This prompt does three jobs at once. The banlist ("forget folders, tags, dark mode…") strips the mean. The category constraint forces the model into specific corners of the space. The "describe the user moment, not the feature spec" framing makes the model think about *why* a feature would matter, not just what it is.

What comes back is something like "prompt resurrection notifications — three weeks after you save a prompt, the app shows you the prompt and asks if you'd still use it; if yes, it floats it back to your home screen; if no, it asks why." That's a feature I'd actually consider building. "Folders" is not.

Sequence 3: Personal decision ideas

This one is for when I'm stuck on a decision and want to break it open before committing. Generic "give me options" prompts return the obvious three options I already considered. That's not useful — I need angles I haven't seen.

I'm trying to decide [DECISION]. Don't give me the obvious options I've already thought through. Give me three angles I almost certainly haven't considered. Constraints: (1) at least one option should sound risky or socially awkward, (2) at least one option should be radically smaller in scope than what I'm currently considering, (3) at least one option should reframe the problem so the original question is the wrong one. For each angle, tell me the specific situation in which it would be the right call.

The "at least one X" constraints are the key here. Without them the model returns three safe options that all live in the same neighborhood as my original framing. With them, it's forced to occupy three different neighborhoods — risky, small, reframed. At least one of the three is usually the angle that breaks the decision open.

How to know your brainstorm prompt is working

The signal I look for is *internal disagreement*. If a brainstorm returns five ideas and I can rank them — best to worst — within ten seconds, the prompt was probably too narrow. The ideas are all variations of the same thing.

If a brainstorm returns five ideas and I have to actually think about which one to pick, because they sacrifice different things, attract different readers, or solve different problems — the prompt did its job. The model traveled away from the mean and gave me a real tradeoff to make.

Ranking is easy. Tradeoffs are valuable. Brainstorm prompts that produce real tradeoffs are the only ones worth keeping in your library.

What I'd do if I were starting today

Pick a brainstorm task you've been disappointed by — launch copy, feature ideas, content angles, anything. Find the prompt you used and add three things: a banlist of five clichés you keep seeing in the output, one structural constraint, and a "what does each option sacrifice" line.

Run the new prompt once. Compare the output side by side.

You'll never write "give me ten ideas for X" again.