Most bad prompts aren't missing words. They're missing decisions.

"Classify this support ticket" sounds complete until the model has to guess your categories, your output format, and what to do with a ticket that fits nowhere.

Every guess is a place the output can drift. Adding more paragraphs does not help if the decisions are still missing.

Make the decisions explicit

• The task: what exactly is being classified, and what is out of scope.
• The output contract: which labels, in what format, and nothing else.
• The fallback: what to return when the ticket is unclear. If you cannot write the fallback rule, the model cannot follow it either.
• Examples: only where the boundary between labels is genuinely fuzzy.

That last one is when few-shot earns its tokens. If zero-shot already separates billing from technical, examples add length, not clarity.

Reasoning models change less than you think

Asking for a long step-by-step chain of thought is not a free upgrade. A clear goal and a strict output format usually do more, and they cost less per call.

Then test it like code

Freeze a set of real tickets, including the awkward ones, and rerun them every time the prompt changes. Otherwise "better" just means the last answer felt better. This is the same discipline we apply to the pipeline that reviews pitch decks: every prompt change runs against a fixed set before it ships.