Pick.
How to choose what to build with AI when there are too many ideas on the table.
You don’t have a focus problem.
You’ve been picking projects that aren’t yours.
I lead multiple programs developing AI models at Google. Twenty years in tech. The same trap has caught me, every cycle.
Below are the three exercises I actually run when there are too many AI ideas on the table and nothing on it is finished.
The trap.
You watched a tutorial yesterday. Stripe dashboard held to the camera. Thirty thousand a month. AI agent built in two weeks.
You took notes. You bookmarked the tutorial. You opened a doc.
You started something.
You stopped.
It’s not the tutorials. The tutorials are fine.
The problem is copying them.
The project in that tutorial was shaped by someone else’s drive, someone else’s customer, and someone else’s timing. When you copy it, you’re trying to graft someone else’s life onto yours.
When the work gets hard, the energy runs out. Because it wasn’t yours to begin with.
That’s why your folder fills up with half-started projects and finishes nothing. Not because you lack discipline. Because the decision to start happened before you ever did the work of deciding what was actually yours to build.
The three exercises below are what fix that.
They come from people who have spent decades thinking about how humans choose what to do with their lives. They aren’t a sequence you run end-to-end. They’re three different lenses on the same question.
Use one. Use all three. Whichever lands first.
Exercise 1: Yearning Octopus.
From a Tim Urban article called How to Pick a Career That Actually Fits You.
Urban’s idea: everyone has a yearning octopus inside their head. Five tentacles. Each wants something different. They don’t get along.
For someone trying to build with AI, the five tentacles look like five drives.
Personal. Making something that means something to you.
Social. Respect from peers.
Lifestyle. Freedom from a day job.
Moral. Helping people, having an impact.
Practical. Making money, paying the bills.
Different drives want different builds.
The respect drive picks an app, because peers see apps and apps are status. The lifestyle drive picks something that runs without you, because the goal is freedom, not more hours. The moral drive picks a problem you can’t ignore. The practical drive picks a service, because services pay this month.
When the drive and the project don’t match, the work stalls around week three. The project isn’t bad. The drive behind it just isn’t yours.
This is why “just pick an idea and start” advice fails. The drive doing the picking has to be the drive that finishes. If you let the wrong one pick, you start, then quit.
Run it.
Step 1. Pick an AI project you’ve thought about. Maybe one you almost started.
Step 2. Ask which of the five drives is actually pulling you toward it. Not the one that sounds best out loud. The one that did the actual pulling.
Step 3. If nothing comes to mind, ask why you haven’t started. “I’d look stupid” is social. “I’d lose money” is practical. “It wouldn’t matter to anyone” is moral. The fear points at the drive too.
That’s your top drive. The honest one. The one you wouldn’t say to a stranger at a dinner party.
Then ask Urban’s question: is this actually mine, or did society hand it to me?
If the answer is “society handed it to me,” you’ve found the cleanest reason most projects stall.
Exercise 2: 5/25 Rule.
Attributed to Warren Buffett. Doesn’t matter if he said it.
It works.
You’ve considered more AI projects than you remember today. Twenty-five, easy. Maybe more. They’re in sticky notes, in voice memos, in dead Notion docs, in half-renamed folders.
The 5/25 Rule cuts the noise.
Run it.
Step 1. Write down 25 things you’ve considered building with AI. Force the number. Pull from every dead doc, every voice memo, every napkin.
Step 2. Pick the 5 most important to you. Circle them.
Step 3. Cross out the other 20.
The part most people miss.
Crossed out doesn’t mean backlog. It means off the table. Buffett’s framing: they get no attention until the 5 are done.
This is the move that actually changes something in your head.
Your brain has been cycling 20 silent ideas in the background. Every time you sit down to work on one, the others pull at your attention. That’s the tax you’ve been paying.
Killing the 20 stops the cycling. The 5 start getting the focus the whole list was eating.
It feels uncomfortable. That’s the point. If it feels easy, you didn’t bury them.
Your top 5 only matters when the other 20 are off the table.
Exercise 3: Napkin Test.
Customer development thinking, distilled.
If you can fit your project on a napkin and you know one specific person to give it to, that’s the project.
You now have 5 candidates from Exercise 2. They all feed the drive from Exercise 1. They all came from your own list.
But only one of them is the project you actually start.
To find it, you run each through two filters.
Filter 1: The napkin.
Can you describe what this project does in one sentence?
Not a paragraph. Not three sentences. One. Short enough to fit on a napkin.
If you can’t, you don’t understand the project well enough yet. That’s not a flaw in the project. It’s a signal that the project is still vague in your head, and vague projects don’t finish.
Filter 2: The name.
Can you name one specific person who would actually use this?
Not “small business owners.” Not “indie hackers.” A specific person. By name. Someone you could text right now and they’d recognize the problem.
If you can’t name the person, you don’t have a customer. You have an audience in your imagination. Audiences don’t pay. People do.
Scoring.
Pass both filters: that’s the project you start. Fail either: back to the dead list. Two candidates pass both: you have a luxury problem, pick by gut.
The reason this filter is so cruel is also why it works. Vagueness hides on a list of 25. It hides on a list of 5. It does not survive one sentence and one name.
One sentence. One name. That’s the project.
A worked example.
Say you have 25 ideas that range from “AI tutor for my niece” to “an automation that summarizes legal contracts.”
Exercise 1. You run the yearning check. Your top drive is lifestyle. You don’t want another job, you want freedom from the one you have.
That kills half the list. The agency-style services are out. The “10k MRR SaaS” is out. They’re more work, not less.
Exercise 2. From the 25, you circle 5 that fit lifestyle. Cross out the other 20.
You now have 5 candidates. All five fit your drive.
Exercise 3. You napkin-test all 5.
The legal contract summarizer fails Filter 2. You don’t know any lawyers.
The niece tutor passes both, but it doesn’t feed the lifestyle drive on its own.
The “morning briefing for solo founders” passes both. You can describe it in one sentence. You can name three specific solo founders who’d use it.
That’s the project.
Notice what just happened. You went from 25 vague ideas to one specific project with a real first customer, in three exercises. No tutorial required.
How you know.
Run each candidate through the three exercises before you start anything.
A drive that’s actually yours. Five candidates pulled from your own list. One project you can fit on a napkin and hand to someone real.
That’s how you know what to build with AI.
Not someone else’s thumbnail. Not someone else’s vision. Not the project that pays the loudest creator on YouTube.
The project that survives all three exercises is the one that finishes.
What’s next.
The full walkthrough is in this video, with examples and the exact prompts I use:
Next week: how to build the smallest possible version of the survivor project in a weekend. No coding. No buying another course.
If this helped, send it to one person who keeps starting AI projects they don’t finish. That’s how Growsembly grows and how these stay free.
— Guney
Growsembly is a weekly newsletter for non-technical creators and professionals who want to actually finish what they start with AI. Subscribe here.