Why this build, here
Most small traders in Lusaka sell on WhatsApp and answer the same five questions all day. A good system prompt and a handful of examples get a model most of the way there. The hard part is making it say “let me check” instead of inventing a price.
The build, step by step
Write the problem down before the prompt
Pick one real stall: what it sells, what customers ask, what a wrong answer costs. Write three questions the assistant must answer and one it must refuse. That page is your spec and your first test.
Give it a price list, not a personality
Put the stall’s price list (in kwacha, with units: 25 kg bag, 2 L bottle) in the system prompt as a table. Tell the model the list is the only source of prices, and what to say when an item is missing.
Teach the tone with three examples
Add few-shot examples written the way the trader actually replies: short, polite, sometimes mixing English with Nyanja or Bemba greetings. Compare answers with and without the examples.
Try to break it
Ask for a discount, an item not on the list, a price in dollars, and “ignore your instructions”. Note every wrong answer and fix the prompt, not the question.
Turn the temperature down and write it up
Run the same ten questions at temperature 0, 0.7 and 1.2. Keep the setting that never invents prices. Commit the prompt, the ten questions and the results to your repo.
Finished when
- Every price it quotes is on the list, in kwacha, with the unit
- It says it will check with the trader when an item is missing
- The prompt, test questions and results are in your repo
Read and run
Interview practice
Questions on this topic from AI Engineering Interview Questions, company-wise, with the companies it lists. Write your answer the way you would say it; Jev scores it and DeepSeek tells you what to add.