You know that golden retriever at the park who sprints off the moment you throw a stick, tail helicoptering with pure joy, only to return thirty seconds later with a shoe? Then a pine cone. Then someone's half-eaten sandwich. That dog has enthusiasm in spades but zero quality control. It's just so happy to bring you things, any things, all the things.
That's ChatGPT when you ask it to brainstorm.
The AI will generate ideas with the same indiscriminate excitement as a retriever who thinks every object in a five-acre radius might be exactly what you want. And here's the thing: that enthusiasm is actually useful. You just need to teach it which sticks are worth carrying home.
Why Your First Fetch Always Brings Back Garbage
When you prompt an AI to generate ideas, it has no built-in sense of what makes one idea better than another. The model is designed for volume. Ask for blog post topics and you'll get fifty of them, ranging from genuinely clever to "did you just suggest I write about my houseplant's LinkedIn profile?"
The model predicts what words might plausibly come next based on patterns it learned from millions of documents, but it's not evaluating whether those words form a good idea or a terrible one. That's a completely different task.
Ask ChatGPT for remote work blog topics. You'll get genuinely useful suggestions like "How to set boundaries with always-on Slack" right alongside absolute duds like "10 ways your cat is your coworker now." Both are plausible blog topics that exist somewhere on the internet. The retriever fetched them with equal confidence.
This isn't a flaw. Brainstorming should cast a wide net. The problem is when you stop there, staring at your screen like someone holding a damp sandwich a dog just delivered, wondering what you're supposed to do with this.
Teaching the Retriever to Sniff Before Delivering
Here's the trick: you need two prompts, not one.
First, you generate. Ask for way more ideas than you could possibly use. Then you ask the AI to evaluate its own output.
The model can absolutely apply judgment when you explicitly request it. You just have to give it criteria. This is your evaluation rubric: the specific standards you want applied to separate the useful from the useless.
Try this. After ChatGPT dumps fifty remote work blog topics in your lap, send a follow-up: "Rank these by how actionable they'd be for someone working from home for the first time. Eliminate anything that's pure clickbait or requires expertise they wouldn't have."
Watch what happens. The AI will ruthlessly cut the fluff. It'll surface the practical gold and push the "your cat is your coworker" nonsense to the bottom of the list, or drop it entirely. The exact same model that generated the mess will now curate it, because you've switched it from fetch mode to quality control mode.
Your follow-up prompt structure matters. Be specific about what "good" means: feasibility, originality, audience fit, tone, whatever matters for your context. The clearer your criteria, the better the filtering.
Training for Specific Sticks (Not Just Any Sticks)
You can make the initial fetch cleaner by defining constraints upfront.
Instead of "give me ideas," try "give me ideas that fit these criteria." Specify your tone, format, audience, and constraints before the AI starts running. This narrows the search space. The dog now understands it's looking for sticks specifically, not every object in the park.
Want email subject lines? Don't just ask for subject lines. Ask for "subject lines under six words, no question marks, must mention a specific pain point the reader experiences daily." You'll get fewer generic "Are you struggling with X?" variants and more targeted options.
But here's the thing: you still need that second filtering step. Even with constraints, the AI will produce uneven results. Some will nail your criteria. Others will interpret them creatively (read: incorrectly). Constrained fetching is better than unconstrained fetching, but it's not a replacement for quality control.
Once you've filtered and found the best results, you can iterate. Use the winners to prompt for more in that direction. "Give me ten more like number three, but make them slightly more casual." You're teaching the retriever what good looks like by showing it examples.
When to Let the Retriever Keep the Pine Cone
Sometimes a terrible idea is actually useful, just not in the way you expected.
Let's say you're brainstorming team-building activities and the AI suggests "host a silent meeting where everyone communicates via interpretive dance." Your first instinct is to delete this immediately and question your life choices. But hold on.
Ask the AI why it suggested that. "What problem were you trying to solve with the interpretive dance idea?" The model might explain it was addressing over-reliance on verbal communication or the dominance of loud voices in meetings. That's actually a useful insight, even if the proposed solution belongs in a comedy sketch.
Terrible ideas can be springboards. They point toward a problem or angle you hadn't considered. The trick is distinguishing between bad ideas and unexpected ones. Bad ideas solve the wrong problem or create more problems than they solve. Unexpected ideas solve the right problem in a weird way.
Before you toss something, ask the AI to explain its reasoning. You might find there's a pine cone worth keeping after all, even if you're not going to throw it.
So What Can YOU Do With This?
The two-step method works for almost any brainstorming task. Generate more than you need, then filter ruthlessly. Here's how to apply it:
For email subject lines: "Give me 20 subject lines for a sales email about project management software." Then: "Rank these by likelihood to get opened by a busy operations manager. Eliminate anything that sounds spammy or vague."
For meeting agendas: "Generate 15 potential agenda items for a product team sync." Then: "Filter these down to what's achievable in 30 minutes and remove anything that should be an async update instead."
For social media posts: "Give me 25 different angles for a LinkedIn post about our new feature." Then: "Cut anything that sounds too salesy or uses jargon a non-technical person wouldn't understand."
For project names: "Generate 40 potential names for a customer feedback tool." Then: "Rank these by memorability and clarity. Remove anything too cute or too generic."
For article outlines: "Produce 10 different ways to structure an article about remote team communication." Then: "Critique each structure for logical flow and eliminate any that front-load too much theory before giving practical advice."
Copy these templates. Adapt them. The universal pattern is always the same: generate volume, then apply specific criteria to filter it down.
TL;DR
- Ask the AI to generate way more ideas than you need because quantity beats quality in round one.
- Use a second prompt to make the model evaluate and rank its own suggestions based on criteria you provide.
- Define constraints upfront to get cleaner initial results, but still filter afterward since even constrained output produces uneven quality.
- Sometimes weird suggestions point toward useful insights, so ask the AI to explain its reasoning before you dismiss them entirely.
Just remember: eventually you need to stop throwing sticks and actually play fetch with the good ones, or you'll spend all afternoon holding a shoe and a sandwich while the dog wags expectantly at your feet.
Top comments (0)