Featured termA fresh pick each visit
Prompt-and-pray
- Vibe Work
- Practitioner slang
Prompt-and-pray is relying on an AI prompt to produce usable results without adequate checks or a plan for consequential failures.
A prompt that works on a handful of tries looks finished, though those tries say little about the next hundred. A model can answer the same input differently from one run to the next, and later inputs bring cases the tries never covered.
Where you’ll meet it
You set up an AI to read your household bills from email and add each amount to the family budget spreadsheet. It handled the three bills you tried, so you let it run every month. In April it reads a water bill's account number as the amount due and enters $40,712.
What to do
Correct the entry against the bill. Add a rule outside the model that holds any amount over twice that bill's usual total for review, and check that the extracted value comes from the amount-due field. Record each bill, the extracted amount, and the destination cell so an odd value can be traced back.
Easy to mix up
Vibe coding leaves the code behind an app's visible behavior unread. Prompt-and-pray leaves an AI workflow's outputs unchecked, with no plan for the failures that matter.
Check yourself
For months, a prompt has drafted the weekly update your team emails to clients, and you've approved each draft before it went out. Your manager now wants the updates sent automatically. How much does that change the risk?
Go deeperEvaluation and validation address different kinds of failure
The phrase is informal slang among people who build AI features. Evaluation-driven work uses saved inputs with trusted answers or explicit criteria to assess changes to the prompt or model. A change that fixes one case and breaks three older ones can then be caught before anyone relies on it.
Structured output can constrain a reply's data format using a schema: rules for its fields and value types. JSON is one common format. Code can reject a reply that violates the schema, but correct structure does not establish true content. When available, a fixed, dated model version reduces unplanned model changes; checks on output are still needed.
Sources
- Introducing Structured Outputs in the API, OpenAI
- GPT-4.1 Model, OpenAI documentation