Ryan Carson's "Proven 3-File System to Vibe Code Production Apps" for AI-Assisted Development with AMP
🎙️ Introduction: Ryan Carson, a serial founder and former CEO of Treehouse, shares his innovative "3-File System" for leveraging AI in building production-ready applications. Working solo on his latest venture, Untangle, an AI-powered divorce app for Connecticut, Carson demonstrates how structured interaction with AI agents, specifically using AMP, can dramatically enhance productivity and output quality for solo AI founders. His system emphasizes guided development over reactive "vibe coding," likening it to providing a new engineer with clear instructions rather than vague requests.
📂 Ryan's 3-File System:
-
create PRD(Product Requirements Document) 📝:- Purpose: This initial file serves as a prompt to guide the AI assistant in generating a detailed PRD in Markdown format. It structures the initial feature idea, ensuring comprehensive planning.
- Process: The AI receives an initial prompt, then proactively asks clarifying questions about the problem, goals, target users, and desired output. This interaction prevents ambiguity, leading to a well-defined PRD that is then saved as a Markdown file. Carson stresses that this structured approach counters the "lazy prompting" often seen in unguided AI interactions.
-
generate tasks📋:- Purpose: Following PRD creation, this file prompts the AI to break down the feature into a detailed, step-by-step task list, guiding the subsequent implementation.
- Process: The AI analyzes the generated PRD to propose high-level "parent tasks" (typically 5-6). The user reviews and approves these parent tasks, providing an opportunity for refinement (e.g., adding a "create feature branch" step). Only after approval does the AI proceed to generate detailed, atomic "subtasks" (e.g., 1.1, 1.2) using dot notation for clarity. This two-tier task generation prevents the AI from running off with an unmanageable list of unapproved tasks.
-
process task list✅:- Purpose: This file governs the iterative execution of the task list, ensuring a controlled, test-driven, and approved workflow. It acts as a gatekeeper, controlling the AI's pace and validating its work.
- Process: The AI executes one subtask at a time, requiring explicit user approval ("y" or "yes") before proceeding to the next. Upon completion of all subtasks under a parent task, the AI is instructed to immediately run the entire test suite. If all tests pass, the changes are committed. This emphasizes a "code, test, then commit" cycle, mirroring best practices in traditional software engineering. User approval is also required for critical Git commands.
🧠 AI Development Philosophy: Carson sharply contrasts his structured approach with "vibe coding," which he describes as blindly giving AI a single prompt and hoping for a good outcome. He argues that this is akin to asking an unfamiliar engineer to "make me a super fun game" without providing context. The critical role of Test-Driven Development (TDD) is paramount for AI agents, as it provides the essential feedback loop ("the agent needs to actually know if it's doing things right") to validate their work and accelerate development, preventing endless "fix this, it's not working" cycles.
🚀 The Solo AI Founder & Untangle: Carson shares his personal journey with Untangle, an AI application designed to simplify the complex and expensive divorce process in Connecticut 💔. Inspired by his sisters' experiences, Untangle aims to mitigate lawyer fees by streamlining the completion of 14 forms with 277 unique fields specific to Connecticut law. He employs the "pain pill vs. vitamin" analogy: Untangle is a "pain pill," addressing an acute, high-impact problem users desperately want to solve, unlike his previous education platform, Treehouse, which was a "vitamin" for self-improvement. Carson champions the advantages of being a solo AI founder in this era, citing the control over his time, the joy of building, and the ability to focus on impactful niche problems without the pressures of scaling large teams or seeking venture capital.
💡 Learning to Code with AI: Carson views AI as a powerful coding tutor 🧑🏫, fundamentally changing how people learn. While AI can ship code without deep user understanding, he stresses the enduring value of comprehending the underlying mechanics (analogous to building a house with a robot—you need to understand the plumbing). His practical advice for beginners is to build what they're passionate about (e.g., a Warhammer tracker or a sewing pattern designer). By asking AI to build alongside them and "explain as it goes," learners can achieve profound understanding.
🛠️ Key Technologies & Tools: Carson's workflow relies on several key technologies: AMP (his chosen AI agent), Ghosty (a terminal for Mac), Neovim (text editor), and Jest (a testing framework for TypeScript/Next.js). For large language models, AMP primarily uses Sonnet 4 for the main agent, with 03 accessible via an "Oracle" tool call for deeper, more thoughtful reasoning (like consulting a senior engineer), and Gemini Flash for summarization tasks. He also mentions using Whisper Flow for voice prompting.
Final Takeaway: Ryan Carson's system demonstrates that the current AI landscape empowers individual founders to achieve significant development velocity by transforming chaotic "vibe coding" into a structured, iterative, and test-driven engineering process. This disciplined approach, coupled with AI's tutoring capabilities and its ability to tackle niche, high-pain-point problems, makes this an unparalleled era for solo AI founders seeking control over their work and lives, building substantial value without the traditional demands of large-scale startups.
