From Learner to Business Owner: an Alumnus on Interviewing the AI First, Criticising Plans, Free First Projects, and Working from Telegram

Vignesh's central technique. Instead of 'build me a poker app', he gives the agent a role-set and an obligation: research the best existing products in the space, identify what is missing, then interview the user in depth as a team of senior designer, developer, engineer and product specialist, producing a plan and a to-do list for approval before any building. He says the interview alone can take 7-8 hours and is what keeps his output ahead of others building the same thing. The prompt also states constraints up front: non-technical builder, a $20 plan, use sub-agents and delegation, plan with the top model and execute with cheaper ones, phase-by-phase tested builds, and scale to 1,000-2,000 users from day one rather than a throwaway MVP. When you do not know what to build, interview real people (a business owner) and feed that interview into the model as a second interview.

Second prompt in his chain: have the agent assemble a team of the most critical thinkers to assess the plan across dimensions, list what is missing and show a before/after of the improved plan. He repeats this three or four times and alternates which model criticises which - a plan made with one model is critiqued by another, then critiqued again - and says the plan ends up crisp. He notes Outskill teaches the same habit of not trusting the first output.

Many learners have no project, so the course feels abstract. His fix: start building something during the course and use office hours to ask experts about it; find the project by asking friends, family and local shop owners what problems they have, and keep talking about AI in social settings. Do the first projects small, even free, because their value is confidence, a showcase and paid access to a top-tier AI subscription (he suggests aiming for a $100 plan funded by clients) - not profit. Lowest-hanging fruit: invoice generators and inventory management for small shops, sold with a small fee and a short lesson in how to use them. Do not chase six-figure clients at the start. For research on a new venture, he again points to the interview, not desk research alone.

Timeline as told: a 13-year career in another field ended; a free 2-day mastermind, a friend-gifted 14-day course and later Catalyst; a poker-club operations app first built on Replit with a browser agent helping him connect a database (a hit-and-miss first version that spent a friend's money); then a friend at a gaming company facing closure joked he had 'five agents', which led a marketing team to fund a 1-lakh rupee experiment. He interviewed five senior agency staff (creative director, designer, content writer, video director), put the recordings into a ChatGPT/Codex project, and worked out how to digitise each role's principles into many specialised versions. He used Catalyst office hours for RAG, vectorising and Supabase questions. He states the company was funded in January (about 3 crores rupees, roughly $300k as he estimates) and built its first version on Emergent for 4-5 months before abandoning it for an agentic approach.

He shows his agent (a customised Hermes layer with a persona, engines, skills, memory and security set up behind it) taking a live request in Telegram: 'build a burger-place website' using sub-agents, deep research and Codex CLI for images, in an isolated new project folder, returning an HTML file. He stresses there is a large system behind the simple chat. He shows a game-trailer-style video made from storyboards, character sheets and iterations (failures included), uses Minimax and another video model, and a 10-minute local video he says cost $0 on a home GPU (a 24 GB used RTX 3090) with Minimax locally. He runs Ollama with GLM 5.3 Flash as his main model, says his total AI spend is around $120 a month, uses a free voice-dictation tool, and recommends a cheaper video vendor with a discount code he states he works with. He believes the future of apps is away from the computer.

Advice pulled together: stay on one tool long enough to learn its limits instead of moving each time a lab releases a model; take from every teacher the way they think - n8n taught him memory, agent and decision structure even though he never used it - and 'understand the why'; do not give up when apps break, persistence and rebuilding matter most; look at the model's thinking to learn how databases and front/back ends connect; a functional app takes at least three to four months and should not be built alone; with your own project you become blind to bugs ('brain rot'), so use 4-5 testers or early customers; ask the interview to decide which features to keep. He currently favours Codex on cost for beginners and says he has not tried Claude's cheaper option.