Most prompts are written once, tested on three examples that happen to work, and then defended in meetings on the strength of a feeling. When the assistant starts producing the wrong thing, nobody can say when it began, which edit caused it, or whether the fix helped. This course treats a prompt as something you build, version and measure, not something you phrase well.
The first day is prompt architecture. You take an overloaded prompt apart into a chain of small steps that can be tested separately, then separate the instruction from the context and the data so each can change without rewriting the others. You choose few-shot examples that correct a specific failure rather than padding the window, and you learn when step-by-step reasoning earns its cost and when it just adds tokens and noise. The afternoon moves to the parts that touch other systems: forcing output against a JSON schema and validating it before anything downstream sees it, repairing malformed responses instead of retrying blindly, writing a tool description the model calls with the right arguments, and grounding an answer on your own documents so it cites its source and refuses honestly when the answer is not there.
The second day is evaluation, which is the part most teams skip. You build a golden set out of real failures from your own task, write a rubric that a second person would score the same way, and separate the checks a machine can do reliably from the judgements that need a model or a human. You run a model-graded evaluation and then examine where the judge is unreliable, including position bias, a preference for longer answers and a tendency to favour output from its own family. You measure how far the judge agrees with a human rater, because a judge nobody has checked is just another unverified opinion.
The last module is what makes the work survive contact with production. You compare two prompt versions with a pairwise test, run a regression suite before a change ships, track cost, latency and token use next to quality so a quality gain is not paid for twice, and harden the prompt against injection from untrusted content such as a web page, a pasted email or a user-uploaded file. You leave with a versioning and logging scheme that lets you trace any bad answer back to the prompt that produced it.
Every exercise can be completed in a spreadsheet and an assistant's ordinary interface, so no programming is required. A Python notebook is provided for anyone who would rather script the harness. Bring a real task from your own work with a few good outputs and a few bad ones, and you will leave with an evaluation set for it. The class runs live online or in the classroom in Colombo, priced in LKR for Sri Lanka and in USD internationally.
