One person, one month, a full platform: five rules for building with AI agents
AI coding agents write most of the code. The rules are what turn it into production software instead of a demo.
By Hassan Ed-dyb · · 1 min read
Short answer: AI coding agents like Claude Code, Codex or Cursor can write most of the code, but the person running them owns the rules. Five rules turn their output into production software: one task at a time, a strict definition of done, specs instead of long prompts, a lessons file, and an agent that is told to push back.
1. One task at a time
Everything lives in one task list, split into phases, done in order. Never "build the dashboard". A small task fits in the agent's attention; a vague one gets a confident, wrong answer.
2. "Done" has a definition
Type check, lint, tests and build pass. End-to-end tests pass at phone and laptop width. Every new page passes an accessibility check. Screenshots are compared to a written design bar. Then the live site is tested. "I wrote it" is not done. "It passes" is.
3. Specs, not long prompts
Write one spec per area: product, design, tech. Each task points to the spec it needs, so the prompt itself is one line. The knowledge stays in files, not in a chat that disappears.
4. Every mistake becomes a rule
Keep a lessons file the agent reads before every phase. When something goes wrong, the fix is not only in the code but in the rules. The same mistake never happens twice.
5. Make it push back
An AI agent will build a bad idea perfectly. So the instructions say: warn me when something doesn't make sense, and give me options, not paragraphs.
AI didn't make building easy. It made the boring parts cheap enough that I stopped skipping them.