Building AI apps with three documents: spec, blueprint, and to-dos

Executive overview

Every AI-assisted app build starts with three documents — not code — that give the AI precise context at every stage. The spec defines what to build, the blueprint defines how to build it with copy-paste prompts, and the to-do list compensates for the AI's fading memory across conversations. Non-technical builders can create all three using AI itself, making the system accessible without coding knowledge. Choosing the right model at each stage and persisting through errors — then embedding lessons into a rules file — is what separates working apps from abandoned prototypes.

The specification: interview the AI to define what you're building

• Use a structured interview prompt that forces the AI to ask one question at a time, back and forth 15–20 times.
• Speak your answers using dictation to give as much context as possible.
• If the AI asks something you can't answer, tell it to answer its own question given the constraints you've provided.
• Resist the AI's suggestions to add features — keep the scope to the simplest version that solves your problem.
• Use ChatGPT in auto mode for speed during the interview; switch to thinking mode at the very end to produce a thorough spec.
• The finished spec is a step-by-step document you could hand to a developer.

The blueprint: a phased build plan with embedded copy-paste prompts

• Paste a detailed prompt plus your spec into a high-end model (Claude Opus 4.5 recommended for longest output window).
• The blueprint breaks the build into phases, each phase into small steps, each step into a prompt you copy-paste into your coding AI.
• Each step should be small enough for an AI to implement and safely test in isolation.
• Specify real API calls and real data in tests — mock data causes tests to pass while the app still fails.
• The coding AI should receive all three documents (spec, blueprint, to-dos) every time it starts a new task.

The to-do list: a persistent roadmap that extends AI memory

• AI context fades over long sessions and across new conversations — the to-do list counteracts this.
• Generate the to-do list by appending a single one-liner to the bottom of the blueprint prompt.
• The output is a markdown checkbox list organised by phase and step.
• Start a new conversation after each completed step to refresh the AI's memory; it reads the to-do list to see what's done and what's next.
• Checked boxes let the AI orient itself without needing to remember the full build history.

Choosing the right tool and model for code generation

• Cursor is the recommended coding environment — better UI than alternatives and production-ready output.
• Replit Agent and Google AI Studio are useful for prototypes but not production apps; AI Studio is limited to Gemini only.
• Current daily driver: Codex (GPT-5.2) via the Cursor plugin — frequently one-shots entire phases with no errors.
• Claude Opus 4.5 excels at front-end work and visual taste.
• Gemini 3 Pro is best for hard bugs and UI testing — it can click through the app, collect errors, and self-correct.
• Model recommendations change monthly; A/B test regularly as new releases drop.

Handling errors and embedding lessons

• Errors are inevitable — persistence is the skill, not avoiding errors entirely.
• Most errors stem from model knowledge cutoffs: the AI doesn't know about APIs or models released after its training date.
• Fix: feed the AI official documentation for any new API or model version you need it to use.
• Once an error is resolved, ask the AI to add a brief, information-dense lesson to your rules file so future sessions never repeat it.
• Rules file location depends on tool: agents.md for Codex in Cursor, .cursor/rules for native Cursor, CLAUDE.md for Claude Code.
• Keep the rules file concise — it loads into context every session, so bloat wastes memory.

Scaling up: bigger chunks as models improve

• Four months ago you had to feed the blueprint one prompt at a time; current models can handle an entire phase in a single prompt.
• A full phase can take 25–45 minutes to complete, but Codex typically gets it right first time, making the wait worthwhile.
• As models improve, the right chunk size will keep growing — test regularly to find the new ceiling.