Use the right environment
for the right decision.

Give each tool a job

AI-assisted development becomes harder to govern when every environment is treated as an interchangeable chat box. A cleaner workflow gives each system a responsibility.

• ChatGPT: clarify the product, reconcile context, define decisions and prepare implementation instructions.
• Codex or coding agent: inspect the actual codebase, make bounded changes and run technical checks.
• GitHub: preserve the durable history, branch state, pull request review and exact production authority.
• Preview environment: prove the rendered behavior before production.

Make the handoff explicit

Do not ask the coding agent to rediscover the product from scratch. Send an implementation packet containing current state, the decision, files or systems involved, preservation rules and acceptance criteria.

Use GitHub as memory, not just storage

A branch is a temporary work state. A pull request is an evidence boundary. The default production branch should represent the latest approved system, not a scrapbook of experiments.

For material changes, preserve the exact head commit, CI evidence, preview deployment and release decision. This makes recovery dramatically easier because “what shipped?” has a precise answer.

Verification happens outside the prompt

An AI can tell you that a change should work. That is not the same as observing that it does work. Verification should cross the actual boundaries of the feature: browser → request → server → data or external service → response → rendered result.

A compact build handoff template

Continue from the exact current repository state.

CURRENT STATE:
[branch / commit / deployed behavior]

DECISION:
[one approved change]

DEPENDENCIES + RISKS:
[what could break]

IMPLEMENTATION:
[specific bounded work]

PRESERVE:
[approved behavior and commercial truth]

VERIFY:
[tests + rendered evidence]

Do not restart discovery or redesign adjacent systems.