Kotoba Dojo — describe an image, AI redraws it, a sensei scores you
Kotoba (言葉) = words. A dojo (道場) = a training hall. This is a training hall for how you put things into words.
Round: see a reference image → describe it (type or speak) → AI redraws from your words only → a Shiba Inu sensei scores the match and shows which words changed the picture. Belts white → black. 22 languages.
Bet: precise description is the same skill as prompting well.
Build: I can't read or write code. No degree. After prison I drove trucks, then sold on Amazon and used AI for product copy before I knew the word "prompt." That became how I ship: I describe the product to Claude + an agentic IDE. 6th product since March. ~40 hours over two weeks. Solo, Tokyo, no funding.
Money: Stripe subs at $9.99 / $24.99 / $49.99 plus credit packs. No IAP.
Also live on Product Hunt today.
Want honest takes: does the scoring feel fair?
White belt is playable without paying — new accounts get 10 free rounds, no card required.
One thing I’d change while you’re validating the scoring: separate speed from the main score.
You mentioned the current breakdown is 65 for visual match, 25 for language analysis, and 10 for speed. That makes the total mix two different signals — description quality and response pace — which makes changes in the score harder to interpret.
I’d keep visual match + language analysis as the main score and show speed separately as a bonus or secondary stat. That keeps speed in the game without muddying what the main score represents.
I also offer a lightweight $99 Launch Check for AI-built MVPs, but even if you don’t need a paid Launch Check, this is the first thing I’d fix.
The fairness problem is separating language precision from generator variance. If the same description can produce materially different redraws, the learner is partly being scored on sampling luck.
I’d test each description across two or three redraws before turning a low score into feedback. I’d also split the result into semantic coverage, such as objects, attributes and relationships, and visual similarity. One combined score can hide whether the wording or the renderer caused the mismatch.
Great point — this is exactly the tension I wrestled with. Two things in the current design address it partially: the rubric already splits into visual match (65) + language analysis (25) + speed (10), and the language portion is scored on what you actually described, independent of the render. Sub-scores are clamped when an element wasn't described at all, so generator luck can't award points for things you never said.
But you're right that a single combined number can still hide renderer variance on the visual side. Multi-sample redraws before penalizing is a genuinely good idea — cost is the constraint (each redraw is a paid generation), but I could see it as a "second opinion" option. Adding it to the list. Thank you!