You open an AI chat, type "give me business ideas," and get back a list you could have written yourself: a dropshipping store, a dog-walking app, an AI newsletter. You close the tab slightly more discouraged than when you opened it.
Here is the takeaway up front, before anything else: AI is not bad at ideas — it is very good at giving you the average answer to an average question. Its output is a mirror of your input. Vague prompt, vague list. The whole skill is learning to hand it constraints sharp enough that the average answer stops being available. That skill is the same one good human brainstorming runs on, and you can learn it in one sitting.
This guide walks through a five-step method you can run with any AI tool you already have open, using a real worked example. Nothing here requires a subscription or a clever "secret prompt."
Why AI ideas come out generic
Language models produce what is most likely to follow your words. Ask a broad question and the most likely continuation is the most common answer — the one that appears everywhere, which is exactly the one that feels stale.
Three things cause the blandness, and all three are fixable from your side of the keyboard:
- No constraints. "Business ideas" has no edges. "Business ideas for a physiotherapist in a town of 20,000 who can only work Saturdays" has edges everywhere, and edges are what force something specific.
- Too few ideas requested. Ask for five and you get the five most obvious. The interesting material tends to sit at positions 15 through 40, after the clichés are used up.
- You judged too early. If you react to idea number two, you never get to number thirty. Generating and judging are different jobs and they interrupt each other — the same reason classic brainstorming techniques insist on separating them.
That last one is worth sitting with. Most people describe AI as disappointing after reading about a quarter of what it produced.
The five-step method
Step 1 — Write your constraint block first
Before you open anything, write four lines on paper. This is the whole trick, and it takes about three minutes.
- Who it is for, as narrowly as you can stand.
- What is true about you — your skills, hours, money, location, what you refuse to do.
- What already exists that you find boring, so it can be excluded.
- What "good" would look like for you specifically.
Notice that none of this is about the ideas. It is about the walls of the room the ideas have to fit inside.
Step 2 — Ask for a lot, in one go
Request 30 to 40 ideas rather than five. You are not going to read them all with equal attention, and you shouldn't. You are trying to get past the obvious ones fast, and a long list does that cheaply. Tell the tool explicitly that repetition and mediocrity are acceptable — it will otherwise pad each idea with sales language that makes skimming slow.
A useful instruction: one line per idea, no explanation, no enthusiasm.
Step 3 — Point at what is nearly right
This is the step almost everyone skips, and it is where AI earns its place. Pick three items from the list — not the best three, the three that are nearly right — and say what is nearly right about each. Then ask for twenty more in that direction.
You are using the model the way you would use a sketch artist: not "draw me something good," but "more like this one, less nose."
Step 4 — Flip a constraint
When the direction stops producing anything new, break one wall on purpose. If everything so far is online, force offline. If everything is a service, force a product. If everything assumes customers pay monthly, force one-time payment.
Constraint flipping is a standard ideation move and it works particularly well with AI, because the model has no ego about the previous direction and will not argue with you.
Step 5 — Judge on paper, away from the screen
Copy your shortlist into a notebook or a plain document, then close the AI tab. Judgement made inside the chat window tends to be judgement made in the tool's voice, and it is unusually easy to be talked into liking something. Our separate guide on telling a good idea from an idea you just like covers the questions worth asking at this stage.
A worked example: content topics for a bookkeeping business
Say you run bookkeeping for small trades businesses and you have run dry on things to write about.
The lazy prompt: "Give me blog post ideas for a bookkeeping business." You will get "5 Tax Tips for Small Business," "Why Bookkeeping Matters," and other things nobody has ever searched for on purpose.
The constraint block instead:
- Who: electricians and plumbers running one-to-three-person businesses, who do their invoicing from a phone in a van.
- Me: twelve years of bookkeeping, strong on cash-flow, unwilling to give tax advice that varies by jurisdiction.
- Boring already: generic tax-deadline reminders, "top 10 accounting software" lists.
- Good looks like: something a plumber would actually read at 7pm and forward to a mate.
The prompt that follows from it: "Give me 35 article topics for electricians and plumbers running one-to-three-person businesses who invoice from their phone. One line each, no explanation. They must be about cash-flow timing, quoting, and getting paid — not tax advice. Avoid generic 'top 10 software' listicles."
Now you get things with texture: what to do when a builder pays 60 days late and you have a van payment due; how to price a callout so the third one of the day isn't losing money; whether to invoice on completion or in stages on a two-week job.
Then step 3: "The late-payment one and the callout-pricing one are closest — they're both about the gap between doing the work and holding the money. Twenty more in that gap."
That is a content plan for a quarter, generated in about fifteen minutes, and it sounds like a person who knows the trade — because the knowledge came from your constraint block, not from the model.
Where the AI part actually helps
Be clear-eyed about the division of labour. The model contributes volume and recombination — it will cheerfully produce the fortieth variation when you have run out at seven, and it will connect two of your constraints in a way you hadn't. You contribute the constraints and the taste. Both parts matter, and only one of them can be outsourced.
This is also why a dedicated generator can be easier than a blank chat: the fields do the constraint-collecting for you. Mellow Ideas is building exactly that — our generators for business names, taglines, content topics, and side-hustle ideas ask for the vibe, audience, and limits up front, then hand back a batch of starting points to react to. They are ours, so treat this as us describing our own tool rather than a neutral review. And they are starting points, not verdicts: any business name still needs availability and trademark checks with the relevant registry, and ideally a trademark professional, before you build anything on it.
Three habits that keep the quality up
- Keep a "nearly right" file. Ideas you rejected but couldn't quite delete are the best raw material for the next session.
- Brainstorm one job at a time. Names, taglines, and topics all want different constraints; mixing them produces mush.
- Expect most of it to be duds. That is not the tool failing. Brainstorming is a volume business, by hand or otherwise — if you are getting three usable ideas from forty, that is a completely normal session. If you are getting none, look at the constraint block first, not the tool.
If the blank page is the actual problem rather than the tool, our piece on why you can't think of ideas works through the usual causes before you go anywhere near a prompt.
FAQ
Does using AI to brainstorm make the ideas "not mine"?
The ideas are starting points; what you keep, cut, and combine is judgement, and that stays yours. A useful test: could you explain to someone why you picked this one over the other thirty-nine? If yes, you did the work that matters.
How many ideas should I ask for at once?
Thirty to forty is a good default for a first pass. Fewer and you only see the obvious answers; many more and skimming becomes a chore. Ask for one line each so the list stays scannable.
Why do my AI ideas all sound the same as each other?
Usually because the prompt gave one constraint instead of several, so the model has only one direction to walk in. Add a second and third constraint — a specific audience, a specific limitation — and the list spreads out.
Should I tell the AI about my budget and skills?
Yes, and be blunt. "I have six hours a week and no money for ads" removes a large family of unusable suggestions immediately, which is the fastest quality improvement available.
Is a generator better than just using a chat window?
It is mostly a matter of effort. A chat window is flexible but you have to supply the structure; a purpose-built generator collects the constraints for you and returns a batch in one step. Both need you to do the choosing.
Stuck on the constraint block, or just want a batch of starting points to react to? Generate business name, tagline, content and side-hustle ideas at Mellow Ideas — calm, honest starting points, and the judgement stays with you.
