Ukrainian AI Product Manager Vitaliy Kobzev created and launched MindLoop, a web platform with daily logic puzzles, in 30 days. During this time, the project went from the first commit to real payments, and a significant part of the work was performed by AI agents.
Ukrainian AI Product Manager Vitaliy Kobzev created and launched MindLoop, a web platform with daily logic puzzles, in 30 days. During this time, the project went from the first commit to real payments, and a significant part of the work was performed by AI agents.
Kobzev shared his experiment on LinkedIn. According to him, the goal was to first try to build your own product from idea to release using AI agents.
As a result, he made 186 commits in a month, agents used about 40 million tokens, and Claude’s costs were $20, not including design work.
MindLoop was conceived as a pet project, but gradually outgrew the initial idea. In addition to the code itself, Kobzev worked on product logic, subscriptions, SEO and GEO, analytics, advertising, and integration of several services.
MindLoop is a web hub of daily visual-logical puzzles that can be completed in a few minutes without registration. Currently, the platform has three games: Mini Sudoku for logic, Pattern Memory for memory, and Pentomino Fitting for spatial thinking.
After completing the game, the user sees their score, percentile among those who played that day, streak, XP, and level. The result can be shared.
The monetization model is freemium. Daily games are available for free with ads, and a subscription unlocks an archive of previous puzzles and premium games and removes ads.
Kobzev took the daily games model, used by LinkedIn and The New York Times, as a basis. In his opinion, this format has several mechanics at work: one attempt per day, a streak, comparing results with others, and the ability to share the result.
Kobzev built the MindLoop architecture on the principle of Core + separate game modules. Core is responsible for users, daily puzzle assignments, streaks, ratings, XP, and access. Each game has its own module with puzzle generation, solution checking, and score calculation functions.
He used Next.js and TypeScript for infrastructure, Vercel for deployment, Supabase for authorization and database, Paddle for payments, PostHog for product analytics, and Google AdSense to monetize the free version.
Kobzev paid special attention to tools for AI agents. He used Claude Code, MCP, and special plugins for working with services. Among the tools the author mentions are code-review and frontend-design from Anthropic, as well as humanizer and plugins for Vercel, Supabase, Paddle, PostHog, and Figma.
According to Kobzev, AI agents performed a significant part of the routine work, but product decisions remained with the human.
“Agents make up most of the puzzle, but it’s the human who closes the loop,” he notes.
At the same time, the experiment showed the author another side of AI development. According to him, along with productivity, cognitive load increased: there was less routine work, but more time was needed to set tasks, check results, and make decisions.
“Previously: thought — transferred thoughts — made a decision — routine work. Now: thought — transferred thoughts — made a decision. And so on in a circle,” describes Kobzev.
Another important element, he says, is documentation. In MindLoop, it’s a roadmap, a backlog, PRDs for functionality, instructions for agents, and dependencies between files. “If you’re really interested in the product you’re making and you have a vision for what it should be, you need to clearly document that,” he writes.
Kobzev now plans to focus on SEO/GEO and MindLoop marketing, and in the future, add new games and AI features.
He also notes that this experiment helped him as a product manager better understand the new "human + agent" work format, in which AI takes over part of the execution, and the human is responsible for the direction, decision, and outcome.
