A Proven 3-File System to Vibe Code Production Apps

Carson uses a Markdown prompt to have an AI ask clarifying questions, then produce a product requirements document. He demonstrates it with a proposed onboarding feature for Untangle, an app that helps people through divorce. The questions and PRD give the agent context before implementation begins.

A second prompt turns the PRD into a small set of parent tasks, which Carson reviews before asking for detailed subtasks. He adds creating a feature branch and asks for tests at each step. He also uses a more capable reasoning model, the Oracle, to check whether the plan is missing larger tasks.

The third prompt tells the agent to handle one subtask at a time, wait for approval to continue, mark completed work, and run tests before committing. Carson demonstrates the workflow with a relationship assessment feature, including a static UI mockup so he can review the design before investing further in implementation. He says this close supervision may be less necessary as models improve.

Carson describes Untangle as a Connecticut-focused divorce app, because the state has 14 forms with 277 fields and divorce rules vary by state. His sisters’ experiences led him to the problem, which he says could reduce the cost of a process that may otherwise cost $15,000 to $40,000. He contrasts this urgent need with Treehouse, which taught about a million people to code but addressed a less immediate need.

Carson says running a company with more than 100 employees took him away from hands-on building, while working solo on Untangle lets him ship and control his time. He argues that a small, cash-flowing business can support a good life without venture funding or a large team. He also suggests learners can use AI as a tutor while building projects they care about.