Vibe coding is a discipline, not a shortcut.

1. planning

Before touching any tool, I plan the flow end to end, but only at the core level. I define what the system should do, what the user sees, what goes in, what comes out, and what absolutely must not fail. The point is not to over-design it. The point is to make the path clear enough that the AI can build against something real.

2. prepare the context package

The quality of the output is determined before the first message is sent. I gather API docs, MCP configs, sub-agent responsibilities, data shapes, constraints, and the exact expected output format. It is not a PRD. It is a context package the AI can actually use.

3. let the ai run the loop

I let the AI write the code, execute the tests, inspect the mistakes, and keep fixing until the result is correct. Once that stage is stable, I move to the next phase. Every mistake becomes a lesson entry so the process gets more self-healing over time instead of repeating the same errors.

4. don't rebuild the wheel

I reuse patterns that already work instead of reinventing them for every build. If a prompt structure, deployment flow, component pattern, or debugging checklist has already earned trust, I start there. New work should compound on previous wins, not pretend every project begins from zero.

5. notes are better than memory

Clean, current documentation matters more than trying to remember everything later. Good notes reduce future development time, lower re-explanation cost, and make it much easier to dive back into a project after weeks or months away. Notes are how fast builders stay fast.