Put the rule on the server, not in the prompt

For the Kasmaji Business Meetup I built an AI chat with one unusual rule: it refuses to do the work until the participant answers its key question.

The point was to make people think before they delegate. An AI that writes the business plan for you teaches you nothing; an AI that first asks “who exactly is your customer?” and waits for a real answer does.

Why not just write it in the prompt?

You can tell a model “don’t answer until the user explains X”. It will mostly comply. But “mostly” isn’t good enough when a room of participants is actively trying to get around it, and a clever rephrase is all it takes.

So the Socratic gate is enforced on the server:

• The server tracks whether the key question for the current step has been answered.
• Until it has, the request that would produce the finished work is never sent to the model.
• The prompt only shapes how the AI asks; the code decides whether it may proceed.

The general lesson

Anything that must always hold, such as an approval step, a budget limit, or a message that needs an admin’s sign-off, belongs in code. Prompts are for tone and judgement. The same idea runs through other Yogency systems: in Bimasakti, WhatsApp messages to parents only go out after admin approval, and that check is a rule in the system, not a request to an AI.

The chat was used by meetup participants on 20 September 2026.