I Had No Ideas, So I Built an AI Board to Judge Ideas First
I had no ideas, so I built an idea-evaluation system first. How I set up a persona debate, and the record of my fumbling
I wanted to make money with a side project and had no ideas, so I used Claude Code to build the thing that judges ideas instead: six personas with different positions, beating up candidates against a scorecard out of 75. Once the setup was in place and I ran it for the first time, what came out wasn’t a debate.
What I set out to do
I wanted to make money with a side project, but I had no ideas. So instead of squeezing out ideas, I decided to use Claude Code to build a system that evaluates ideas first. My first prompt pinned down the nature of the project: a project for organizing service ideas that I, a solo developer, could actually commercialize. Then I set the evaluation method: several different personas would debate and score each idea by category.
Thinking alone leads to confirmation bias, so the idea was to bring in fictional judges with different positions and let them beat the ideas up.
How I asked
1) Assembling the panel. I got persona recommendations from Claude and settled on six: a realist developer, a startup PM, an investor, an AI native, an end user, and an idea bank. That last one does not explain itself from the name: its job was to summarize the evolution so far at each round and put forward the next proposal.
The scorecard was out of 75, made up of fifteen items across five areas, each scored 1 to 5. The areas were market, revenue, feasibility, AI leverage, and scalability.
2) I gave it my own info first. This mattered more than I expected. I entered my experience (9 years, backend), my main stack (Spring/Java, high-traffic systems), and my interests (stocks, games) in concrete terms. For the “realist developer” persona to make substantive objections, my own capabilities, the thing being evaluated, had to be in the prompt. And in the debates that followed, remarks like “can a solo developer handle this scope?” were actually made against my stack.
3) Starting the debate. I asked it to research application topics where solo developers had recently seen results, pick candidates, run the debate, and leave the results in a document: research, debate, and record-keeping all in one go.
Where it broke
The first debate output was disappointing. The six personas politely took turns presenting their impressions once each, and that was it. It wasn’t a debate; it was a recital. So I gave it the rules again.
“It looks like each persona just states a personal opinion and that’s the end. Run it again so they narrow or twist the topic and think about how to raise the score.”
The key is rounds. Not “everyone state your opinion,” but “over repeated rounds, narrow and twist each other’s proposals.” Only when you explicitly specify the interaction pattern itself does the LLM actually debate. After that, the debate started rolling: “general-purpose AI stock analysis is a red ocean” (rounds 1-2) → narrowing down to a disclosure-focused product, and so on.
One more problem: sometimes the debate drifted toward conclusions detached from my situation. The pattern I used for this was barging in with a role. Rather than correcting from outside, I entered the debate as a character (the CEO) and threw in constraints. When the price got set at 9,900 won a month, I barged in as CEO and pulled it down to the 1,000-2,000 won range, and I removed a feature wholesale with “bring-your-own-API-key (BYOK) is only for developers, and our target isn’t developers.” When the premise “a solo developer can’t afford a frontend” came up, I corrected it, saying that I had actually designed and operated a React admin from scratch, and that kept the web version alive. Injecting constraints as a character’s lines bends the direction without breaking the structure of the debate.
Result
Every time a handful of rounds got consolidated into a single proposal, I bumped the version and rescored it. That happened seven times across 34 rounds.
| Version | Rounds | Core pivot | Score |
|---|---|---|---|
| v1 | R1 | General-purpose AI stock analysis web app | 52 |
| v3 | R6~R10 | Revenue model, cost, regulation made concrete | 65 |
| v5 | R16~R22 | CEO intervention 1, shift to ultra-low pricing | 71 |
| v7 | R29~R34 | CEO intervention 3, shift to a web service | 72 |
What got picked was a service where AI summarizes Korean corporate disclosures. A disclosure is the business information a listed company is legally required to publish. There is a lot of it and the prose is stiff, so retail investors do not read it through. That was the opening for AI.
Honestly, I don’t know how much to trust that score. For now, I’m going with it.
Lessons
- A persona debate needs the interaction rules spelled out. “Everyone evaluate” becomes a recital; “narrow and twist in rounds” becomes a debate.
- Enter the info about the subject (me) first. Stack, years of experience, and available time are what make the “realist” objections realistic.
- Bend the direction by intervening as a role. A “CEO:” line kept the structure intact and worked better than instructions from outside the debate.