A prompt is a contract between your business and a model. Like any contract, it should be specific, testable and boring to read.
Separate the what from the how
We split every agent instruction into four parts: the job, the inputs, the rules and the output format. Each part is versioned and tested on its own.
Job: one sentence describing the outcome
Inputs: exactly what the agent will see
Rules: numbered, each one testable
Output: a schema, never free text
Every rule gets a test
If a rule says “never offer a discount above 10%”, there’s a test conversation that tries to get a 15% discount. Rules without tests are wishes.
Treat prompts like code
Prompts live in version control, change through review, and ship only when the evaluation suite passes. It sounds heavy. It’s what lets us change things quickly without fear.

Written by
Priya Raman
Leads the team that designs, evaluates and ships our agents. Former research engineer in applied NLP.



