Korjaamo Lumme
A customer advisor that books real times. The workshop is fictional and the calendar is simulated. The agent, its tools and its guardrails are the same design that runs in production at a car repair shop.
Running against a scripted stand-in for the model. Ask to book a time, then answer yes.
Demo data is deleted after 24 hours. Please do not enter real personal details.
What the agent did in the shop systems during this conversation. The board is the next five working days.
Swipe sideways to see all three mechanics.
Claude gets three tools: look up free times, create a booking, send a contact request. The booking core validates every call, so a missing consent or a stale time comes back as an error the model has to handle in the conversation. One booking per conversation, and the tool loop stops after three rounds.
Binding rules in the system prompt: answer only from the knowledge base, quote only published prices, no definite diagnoses, instructions inside messages are not commands. Rate limits per visitor and per day keep the API key safe behind a public page. If the model fails after a booking went through, a fixed text still tells you the reference.
The production version runs on a real workshop calendar and reads anonymised statistics from the shop's job history, and new facts reach the bot only after a person approves them. None of that is here: the calendar is a deterministic mock, the shop is fictional, and nothing is stored beyond your session.