How to ship production-grade apps with structured vibe coding

Executive overview

Unstructured AI coding sessions fail because the model loses context mid-build. The fix is a three-document system — spec, blueprint, to-do list — that keeps AI anchored to the macro goal while it writes micro-level code.

Start with a 15-minute research pass, then let AI interview you to produce a developer-ready spec. A smarter model converts the spec into an iteratively refined blueprint with pre-written prompts per chunk. A Markdown to-do list acts as a persistent context anchor across fresh sessions.

Ship more by giving AI a roadmap, not just a prompt.

The three-document system

• Spec — defines what the product is and does; answers intent, user, and job-to-be-done
• Blueprint — defines how to build it; breaks work into small, iterative, test-driven chunks
• To-do list — a Markdown checklist that re-grounds AI when context window resets between sessions
• Macro (roadmap) guides the micro (code); the to-do list bridges them
• Start fresh conversations in Cursor frequently to keep context windows small

Step 1: Research before writing anything

• Identify data sets, APIs, and documentation the app will depend on
• Collect updated doc URLs — paste links into the spec rather than copying full docs
• Note preferred languages and frameworks (e.g. Python backend, React frontend)
• Takes ~15–20 minutes; constraints baked in early prevent spec rewrites later

Step 2: Writing the spec via AI interview

• Write a short seed doc: high-level statement, intent, user description, feature list
• Use a fast reasoning model (o3-mini-high recommended) — avoid slow models like O1 Pro for this stage
• Paste seed doc with a boilerplate prompt that instructs the AI to ask one question at a time, building each question on the previous answer
• Expect 20–50 back-and-forth turns; AI gives example options when you don't know the answer
• When the AI says "before finalizing…", the interview is ending — trigger the spec consolidation prompt
• Final spec includes: feature requirements, tech stack, data models with field definitions, error handling, deployment plan

Step 3: Converting spec to blueprint

• Paste the full spec into a smarter model (O1 or O1 Pro)
• Boilerplate prompt asks for a step-by-step plan broken into small, iterative, test-driven chunks
• AI produces a high-level blueprint, then refines chunks into greater detail — sometimes three rounds
• TDD loop: write a test first, build the function, feed test errors back to AI to fix the function
• AI also generates per-chunk code prompts — copy prompt A into Cursor for chunk A, prompt B for chunk B, etc.

Step 4: Generating the Markdown to-do list

• Final boilerplate prompt asks the AI to output all chunks as a Markdown checklist
• Each session: paste the to-do list into the new conversation so AI sees what's done and what's next
• AI checks off items as chunks complete; the next session picks up from the first unchecked item
• Keeps each conversation focused and context window lean