After two years of shipping AI features we have a small prompt library that shows up in some form in almost every project. Today two of them that saved us the most time.
The "strict extractor"
Task: from unstructured text (receipt, voice memo, email) build a strict JSON schema. The trap: models love to invent fields. Solution: schema in the system prompt, with the explicit instruction "only fields in the schema, everything else is dropped". Bonus: a confidence column per field where the model itself says how sure it is.
We use this in TaxCastle for receipts, in Wissensvorsprung for magazine indexing and in the realtor portal for voice memos. Three apps, one prompt, each with a different schema.
The "careful classifier"
Task: sort something into one of N categories, with an honest "don't know" option. The trap: models always want to give an answer. We force them to pick the unsure option when confidence is below a threshold.
Better ten times "unsure" than once confidently misclassified. False confidence costs us more trust than an open shrug.From our internal prompt guide
In RoadReview we use it to transcribe the instructors' voice notes. When the model isn't sure whether the words were "mirror check" or "shoulder check", it returns unsure and flags the spot instead of guessing. That's fine. Better than a wrong word landing in the student briefing. We also pass patterns like these on in our AI consulting — to teams that want to take their own prompts into production.
