Plan G: AI Coding Workflow (Requirements to Delivery)

Claude / ChatGPT

Decide stack and architecture (language, framework, data structures, interfaces) first, with the model comparing options and trade-offs. Settle the plan before writing code.

💡Ask for the simplest runnable MVP first, then expand — steadier than designing everything up front.

Cursor / Claude Code

Initialize the project, directory structure, dependencies and base config until the code runs. Use templates or agent scaffolding to cut manual labor.

💡git init early and commit in small steps so anything can roll back.

Cursor / Claude Code / Copilot

Implement in small steps — finish one feature/module and verify it before moving on. Let coding agents (Claude Code / Cursor Composer) edit and run autonomously, so problems trace to a specific step.

💡Commit every verifiable unit; generating one giant feature at once is how big hidden bugs happen.

Test framework + AI

Write unit tests for critical logic, manually verify edge and error paths, and feed error logs to the model for diagnosis. Let AI write tests, run them and fix regressions.

💡Describe exactly how to reproduce before asking the model — hit rate jumps; test empty, huge and invalid inputs separately.

Code Review / review agent

Self-review the diff, run an AI security/correctness/performance pass, fix findings, and write a clear PR description (what changed, why, how tested).

💡Include test results and reproduction steps in the PR; reviews pass faster and rollbacks are easy.