When should you ask an AI to write code from scratch versus only review or refactor existing code?

Question

Accepted Answer

The rule of thumb: let the AI **write** where mistakes are cheap and obvious, and use it to **review/refactor** where mistakes are expensive or subtle. The deciding factor is how much it costs you to verify the output.

## When to let it write from scratch

- **Greenfield and boilerplate** — a new util, a config file, a CRUD endpoint, a regex. There's no hidden context to break.
- **Unfamiliar APIs/libraries** — the AI knows the surface of a library you don't, and you verify against the docs.
- **Throwaway/prototype code** — speed matters more than perfection.

In these cases the cost of a bug is low and you can read the result top-to-bottom.

## When to use it only to review/refactor

- **Business-critical logic** — billing, auth, money math. A plausible-looking wrong answer is dangerous.
- **Subtle code** — concurrency, security, edge-case-heavy algorithms where "looks right" isn't "is right".
- **Large existing systems** — the AI can't see all the invariants your codebase relies on, so generating into it risks silent breakage.

Here you keep the AI as a **second pair of eyes**: it suggests, you decide.

```text
                 Cost of a bug LOW          Cost of a bug HIGH
New code          ✅ let AI write           ⚠️ AI drafts, you verify hard
Existing code     ✅ AI refactors           ✅ AI reviews, you write the change
```

## Tradeoffs and risks

Writing from scratch is fast but the AI may invent APIs (**hallucination**) or miss context it can't see. Reviewing is safer but slower and only as good as the context you give it. Either way you are the **accountable author** — the AI is a tool, not the signer of the commit.

## Why it matters

Misjudging this is where AI-assisted dev goes wrong: people let the AI generate critical logic they then can't fully verify, or they hand-write trivial boilerplate the AI could have produced in seconds. Matching the mode (**write** vs **review/refactor**) to the **cost of being wrong** keeps you fast on the cheap stuff and careful on the dangerous stuff. The skill being tested is judgment about where to spend your verification effort — not whether you can prompt an AI.