3-File AI Coding System — Ryan Carson

Original source: Full Tutorial: A Proven 3-File System to Vibe Code Production Apps | Ryan Carson

TL;DR

  • A lightweight, production-oriented workflow built around three files: a concise PRD, a list of atomic tasks, and TDD-style tests.
  • Use AI templates to generate and refine the spec, decompose work into bite-sized tasks, and drive implementation through tests and small commits.
  • Practicing TDD with AI reduces bugs and rework; use the model as a tutor to fill gaps and explain trade-offs.
  • Great for solo builders: ship painkillers (urgent value) before vitamins (nice-to-haves); start tiny, iterate, and automate checks.

Core points

  • The 3 files (and why)

    • PRD: clarifies scope, constraints, success criteria, and non-goals in one page.
    • Atomic tasks: granular, independently verifiable steps that keep momentum and aid parallelization (human/AI).
    • Tests: executable spec; stabilize behavior, enable refactors, and define “done”.

  • Workflow loop

    • Draft PRD with AI template → ask questions until constraints are crisp.
    • Decompose into atomic tasks (each ~5–30 minutes); include acceptance hints.
    • Write/auto-generate tests first; implement the minimal code to pass; refactor.
    • Rinse and repeat per task; keep commits small and reversible.

  • TDD with AI

    • Let AI propose edge cases, fixtures, and failure modes; then you prune and enforce.
    • Prefer code-based checks over fuzzy prompts; make judges binary when using LLM-as-judge.
    • Treat test failures as guidance, not blockers; iterate spec/tests together.

  • Using AI as tutor

    • Ask for alternative designs, complexity trade-offs, and gotchas before coding.
    • Have it explain failing tests, suggest smaller steps, and surface missing constraints.

  • Business/solo builder notes

    • Prioritize painkillers over vitamins; solve an urgent, repeated problem.
    • Start as a one-person company with strict scope, clear docs, and automated tests.

Simple checklist

  1. Create a 1-page PRD with goals, constraints, and non-goals.
  2. Break into atomic tasks with clear acceptance criteria.
  3. Write tests first (unit/integration); run them continuously.
  4. Implement smallest change to green; commit; refactor.
  5. Use AI to draft, critique, and reduce task size; keep a lessons log.