The most useful thing an AI did for me this month wasn’t writing code. It was telling me my idea was generic.
Most people think “brainstorming with AI” means asking it for twenty ideas, or saying “build me X” and watching it produce something. That’s the shallow use. The real value isn’t generation — it’s interrogation. Being asked the right questions, in the right order, before you commit to anything.
Here’s what that looked like when I added a tool to this site.
The wrong way to do it
My first instinct was the obvious one: a project scope estimator. Punch in some details, get a timeline. I told Claude exactly that and expected it to start building.
It didn’t. It ran a structured brainstorming process instead — one question at a time — and the very first thing it asked was who the tool was actually for. Not “what should it do.” Who reads the result, and what do they need to feel when they see it. That single question reframed everything that followed.
What it actually did
Then came the pushback that mattered.
When I said “scope estimator,” it challenged me: estimators are everywhere, and any single number a tool produces is contestable. Someone can disagree with the math, and the moment they do, the credibility is gone. So we threw out the idea of a magic number. The product became an honest range — wide when a project is still vague, tight when it’s well defined — with the reasons for the spread shown openly. The cone of uncertainty stopped being a concept and became the whole point.
I wanted it to handle projects at different phases, and I assumed that meant a live AI chatbot doing the thinking. Claude flagged the trap before I walked into it: that needs an API key and a backend, which breaks the one rule this entire site runs on — everything works without me, for free, indefinitely. So we made it deterministic and client-side. It runs entirely in your browser. No server, no cost, nothing to maintain.
And when the first version’s ranges came out too wide to be useful, the process didn’t defend them. It grounded the model in concrete inputs — features by complexity, integrations, team size — so the spread tightened and every figure became traceable back to something real.
None of that came from me asking for ideas. It came from being made to defend the idea until what survived was something I could stand behind in a room.
Why this is the PM job in disguise
That process — questions before activity — is almost exactly what good project management is.
The value of a PM was never in producing motion. It’s in asking the uncomfortable questions early. Who is this really for? What happens at handoff? What breaks the day no one is watching it? Answer those before you build, and the thing you ship survives contact with reality. Skip them, and you spend the next six months patching a system that was wrong from the first sprint.
Brainstorming with an AI that pushes back is just a fast, cheap way to run that interrogation on yourself. The tool I ended up with took an afternoon to build. The afternoon was mostly questions.
It’s live, if you want to try it: the project estimate tool. Pick a phase, add your scope, and it gives you a defensible range with the full working shown — exactly the kind of estimate I’d hand a client on a call.
Try this yourself
The brainstorming process I used isn’t built into Claude by default. It comes from an open-source skills framework called Superpowers, by Jesse Vincent. You’ll need two things: a paid Claude plan (Pro or Max), and Claude Code — Anthropic’s command-line tool.
First, install Claude Code in your terminal and sign in. On most setups that’s:
npm install -g @anthropic-ai/claude-code
Then run claude to start it. Once you’re in, here’s the path I took:
- Type
/plugin— it opens a small plugin browser. - Search for “superpowers” and install it.
- Restart, and you’re set.
The part that tripped me up: nothing shows up until you install it this way and restart. And it lives in the Code environment specifically — you can run it in the terminal or inside the Claude app’s Code mode, but a normal chat window won’t see it.
To use it, start your message with /brainstorming and add what you want to build — for example, /brainstorming a tool that estimates project timelines. From there it takes over: one question at a time, before a single line of code gets written. That’s the whole trick. It makes you think before it makes anything.
The lesson I keep relearning: the quality of what you build is set long before you build it, by the quality of the questions you were willing to sit with.
I build systems that stand on their own.