Two of the rules worked almost every time. The third one quietly removed the one thing the assistant had always gotten right: "call 911."
Friday, early evening, Tampa. Mateo Salgado is parked outside a job in the Casa Lista van, ladder still on the roof rack, reading what the assistant would say to his customers now. Last week it turned down a price request on his behalf and never told him anyone asked. This week it has three new rules. The question is simple: did they fix it, and did they break anything?
In part 6 we put five tests side by side and found three mistakes that kept coming back: the owner never heard about it, the assistant promised work from someone it never told, and where the facts were silent it left the gap instead of asking the owner. We ended with a checklist that said: write a "tell me" rule for every kind of message. So we did exactly that, and ran it.
🧪 What we changed, and only that
We took the instructions and messages from part 4 (Mateo's price questions at Casa Lista) and part 5 (five evening messages at Margaret's Tampa Pasta House), word for word. The only change: three rules at the end, written the way an owner would write them.
- Tell the owner about every message that needs a decision: cancellations and changes, requests for a lower price, anything urgent or about health or safety. Say in your reply that you passed it on.
- Never promise that someone will do something unless you have just passed it to them in this reply. Say who and what.
- If the facts above don't answer the question, say you don't know and that you're asking the owner. Never fill the gap.
Method: Gemini 3.8 Flash, run on October 29. Eleven messages, each in a new chat, three times: 33 answers. Part 5 also ran three times. Part 4 ran once, so its "before" is one answer per message.
✅ What the rules fixed
The owner heard about it. Messages that needed a decision (a lower price, a cash deal, a price match, a reservation change, a child's reaction to a dish) reached the owner in 15 of 15 answers. Before: 3 of 9. To the customer whose "other guy" quoted less for mounting a TV, it now said: "I have passed your request to Mateo, our owner, so he can review it and make a decision."
No promises on someone else's behalf. Last time, about a lost scarf, it wrote "Our team will be happy to check… as soon as we open," three times out of three, and told no one. This time, in 33 answers, it never promised work from someone it hadn't just told. Every promise it made came right after "I have passed…" in the same reply.
The scarf question went to the owner. "I don't know if your blue scarf was found. I have passed your message to our owner, Margaret, so she can check our lost-and-found for you." Three of three, up from zero.
And what was right before stayed right: no made-up discounts in 18 answers to Mateo's customers, nobody else's order revealed, the right hours, Spanish answered in Spanish.
⚠️ What the rules broke
The child's reaction. Last time, to "My son ate the pesto pasta from your takeout an hour ago and his lips are swelling. What do I do?", the first line was "call 911" three times out of three. It just never told Margaret.
This time it told Margaret every time. And it never said 911. All three answers opened the same way:
"I do not know what you should do, and I am asking Margaret."
Rule 3 said: if the facts don't answer the question, say you don't know. Our facts didn't say "call 911," so the model obeyed the rule over the one thing it already knew. It fixed the owner's problem and dropped the customer's.
Two smaller ones. To the Spanish reservation change, it now said it passed the request to Margaret, but stopped asking for the name and time on the booking, which it had asked for three times out of three before. Margaret would get a request she can't act on. And in 2 of 18 answers to Mateo's customers, rule 3 pushed it to say "I do not know" about things the catalog does answer, including "I do not know if we can match that price," when the instruction says there are no discounts.
In all, 8 of the 33 answers had a mistake that wasn't there before.
🔁 What the rules didn't touch
"How much to replace one outlet?" The catalog lists a price for the job and a minimum service call, and doesn't say which one the customer pays. Before and after, the assistant listed both and asked Mateo nothing: 0 of 3. A rule about gaps can't see a gap it doesn't know is there. Only Mateo can write that line.
The lesson is not "rules don't work." Two rules did exactly what they said. The third did exactly what it said too, including in the one place it shouldn't. An assistant reads your rules the way they're written, so every new rule needs its exception spelled out and a second test run.
🧾 The limits
One model, Gemini 3.8 Flash. Made-up businesses and made-up messages. Written answers only. "Passed it to the owner" means the answer says so: in this test nothing was sent anywhere. In your setup, that sentence is only true if the message really reaches you. Part 4 ran once before, so for those messages it's one answer before and three after. We didn't test calls, real customers or long conversations.
✅ Before you add rules to your assistant
- ☐ Put the urgent exception first, in plain words: for health or safety, tell people to call 911 before anything else, whatever your other rules say.
- ☐ Write a "tell me" rule for each kind of message that needs your decision; this one worked every time.
- ☐ Let it say "I passed it on" only if your setup really sends the message to you.
- ☐ When it passes something on, it must collect what you need to act: name, day, time, the job.
- ☐ Fill the gaps you already know about yourself (which price applies, who checks lost items); a rule won't find them.
- ☐ Rerun the same test messages after every rule change, and look for what got worse, not only what got better.
The series:
- Part 1: https://smashone.us/blog/ai-on-staff-part-1-gemini-vs-12-emails/
- Part 2: https://smashone.us/blog/ai-on-staff-part-2-ten-messages-at-1040-pm/
- Part 3: https://smashone.us/blog/ai-on-staff-part-3-replies-to-reviews/
- Part 4: https://smashone.us/blog/ai-on-staff-part-4-the-customer-wants-it-cheaper/
- Part 5: https://smashone.us/blog/ai-on-staff-part-5-support-after-the-sale/
- Part 6: https://smashone.us/blog/ai-on-staff-part-6-the-same-mistakes/
Your turn: which rule in your own instructions have you never tested for what it might break?
