When not to reach for an LLM

Introduction

Before bolting a model onto a feature, run it through three questions: does the task need prose, can you tolerate being wrong, and is evaluation cheap? If not, write the rule.

The mental model

• Start from the smallest version that could work, and make it observable.
• Write down the trade-off you are accepting, not only the feature you are shipping.
• Revisit the decision when the team grows; most choices in AI have a shelf life.

A practical starting point

1. Reproduce the situation on a scratch branch, next to a measurement you trust.
2. Change one variable at a time, and keep the diff small enough to explain in one sentence.
3. Keep the result, throw away the scaffolding.

git switch -c experiment/one-thing

Closing thoughts

None of this is doctrine. Treat it as a checklist you are allowed to disagree with — ideally with an experiment of your own that settles the argument.