The cost of building a first version has collapsed. An agent can scaffold an app in an afternoon. That makes the choice of idea the expensive decision — you can now waste a month on the wrong thing much faster than before.
This post is a three-round brainstorming method that uses AI as a sparring partner. It fits before step 1 of the AI agent workflow.
Round 1 — Diverge: get past the obvious
Start from something real: a problem you had this week, a workflow that annoyed you, a community you know well. Then ask the AI to widen it:
I keep losing track of which side project is stuck. List 25 product ideas around this problem,
one line each, for different kinds of people (developers, writers, students, small shops).
Include some odd ones.Add your own ideas to the list. Don’t judge yet. The first ten ideas are usually the obvious ones everybody has; the interesting ones come after.
Round 2 — Attack: find the reasons each idea fails
Pick the five ideas you find most interesting and ask the AI to argue against them:
For each idea, give the three strongest reasons it would fail as a paid product for a solo developer.
Be specific: existing tools, who would not pay, what is hard to build or to explain.This is where AI is most useful — it is tireless at finding objections and has no attachment to your ideas. Your job is to judge which objections are real. Some are fatal (“everyone already gets this free inside their editor”); some just tell you what the first version must prove.
Round 3 — Converge: one paragraph and a won’t-do list
Choose one idea and write the paragraph by hand:
For solo developers running several repos with AI agents, who lose track of status and sessions, v0.1 is done when they can see every repo’s current task on one screen and resume its session in one click.
If you can’t write that paragraph, go back to round 2. If you can, write the first won’t-do list immediately, while you still remember which features you were tempted by.
Kill ideas fast — and write down why
Most ideas should die in round 2. That is the method working. Keep an ideas.md with every killed idea and
its one-line reason:
- Habit tracker for developers — killed 2026-09: crowded, no reason to switch.
- Session search for Claude — parked: interesting, but part of a bigger problem (status across repos).“Parked” ideas are often the seeds of better ones: the reason for “not now” can stop being true.
What AI can’t do for you
- It can’t measure demand. It knows what has been written, not what people will pay for next month.
- It can’t know your advantage. Which idea you are unusually well placed to build is your call.
- It won’t tell you to stop. Ask it to find reasons an idea fails; don’t ask whether it is good.
After converging, test with people: a landing page, a post in a community you belong to, or a small launch.
Where ShotMatic fits
Brainstorming produces more projects than you can run, and that is fine — as long as each one has a phase and a next step you can see. Put ideas that survive round 3 on a board in the Backlog tab; they stay visible without competing with the work that is active.
