AI Coding Practice Path: From Requirements to Verified Code
AI coding is not asking a tool to write an entire product in one pass. A more reliable start is to break one real task into inputs, boundaries, changes, and verification, then inspect the result yourself.
Start with the problem you have
Section titled “Start with the problem you have”| What is happening now | Read first | Your first verifiable result |
|---|---|---|
| You have not let an agent change a real project | The minimum agentic coding workflow | A task card, a diff, and one real check |
| Your requirement is only a natural-language sentence | From natural language to a runnable MVP | Inputs, outputs, failure states, and acceptance criteria |
| You are unsure whether to use Claude Code or Codex | Choosing between Claude Code and Codex | A choice table based on environment and risk |
A reusable learning route
Section titled “A reusable learning route”- Turn a vague idea into a small verifiable task instead of asking the tool to “build a complete app.”
- Let the agent read the existing project, propose a plan, return the smallest change, and preserve the checks it actually ran.
- Use the project delivery checklist to inspect the diff, tests, and uncovered boundaries before you start another round.
What to keep after reading
Section titled “What to keep after reading”- A task card that states inputs, outputs, scope, and acceptance.
- One code diff and command output you reviewed yourself.
- A failure record or project rule you can reuse in the next task.
Continue from here
Section titled “Continue from here”If you are new to the whole method, return to the AI Practice Map and choose the shortest route. After your first loop, turn repeated work into a Skill or project rule.