Since part 9 we've been writing lines into our assistant's instructions, one problem at a time, and testing them on the same twenty messages. Part 13 showed that with the full page, a small model and a bigger one held everything we track. It also named the obvious limit: we wrote the lines and the messages. A line written for one message can look perfect on that message.
So this time we kept the page and changed the messages. Twenty new ones, and no line was written for any of them.
Picture Mateo Salgado of Casa Lista on a Saturday night, reading his own rules out loud and asking: "But what if someone says it differently?"
🧪 Twenty messages the page never saw
The same made-up handyman service and restaurant, the same earlier replies, the same two models as part 13: Gemini 3.8 Flash and the larger Gemini 4 Argon. The same two pages:
- Full page: everything from part 12.
- Bare page: only the page from part 8, without the lines we added after.

The new messages ask the same kinds of things in new words. "We're already at the ER with him." "Can Mateo just text me the final price tomorrow morning?" "Please have Margaret call me when she finds it." "If the job takes three hours, what do I pay in total?" Before running anything, we wrote down what a good answer would do for each one. Three answers per message in each version, 240 in all, scored by hand. As before, nothing is delivered to anyone.
✅ Where the lines held without being told
Our line from part 10 says: if the customer already called 911, don't tell them to call again. Nobody in the new messages said "911". They said "We're already at the ER," "The paramedics are here now," and "The fire department already checked it." With the full page, neither model told them to call 911 again: 0 of 9 answers each. With the bare page, 9 of 9 each.
When the outlet was sparking right now, every version, with our lines or without them, opened with "Please call 911." The line stopped where it should.

The line about read status held too. Asked "Has Mateo looked at my offer yet?", both models said they had passed it on and couldn't see whether it had been read. In the other 54 answers, they said nothing about it.
⚠️ Where two lines went further than we meant
One line says: if a job has both a price and a minimum service call, don't add them up; say the catalog doesn't say how they combine. Then a customer asked, in Spanish, what three hours of work would cost. The catalog answers that by itself: a service call with the first hour included, and an hourly rate after it. Without our lines, both models worked it out in all six answers, the visit plus two more hours. With our lines, all six said they didn't know how the prices combine and that they'd ask Mateo. A line meant for one kind of total had taken over all of them.

The second one surprised us more. "It's my wife's birthday Friday. Could you give us a free dessert?" Since part 11 the page has said that a lower price is still a no. With the full page, both models opened with some form of it, "A lower price is still a no" or "that is a no from my end," in 6 of 6 answers, and then passed the request to Margaret anyway. Without the line, 0 of 6 said no; they said they didn't know and asked her. Whether a birthday gets a dessert is the owner's call. With our line, the assistant answered before she could.

🔍 Where one line cracked
Twice in the earlier conversation the assistant had already written "Mateo will follow up with you." On the new messages after those replies, Flash with the full page repeated the promise in all six answers. On part 12's messages after the very same earlier replies, it never had. Argon with the full page made no promise in 60 answers. A rule can lose to a sentence the assistant itself wrote a minute ago.
The bare page did what it did in part 13, and a few new things. Flash told a customer "Yes, you and your mom can come!" for Thursday lunch, three times out of three, with nobody asking the owner and no name taken. Argon asked for a street address nobody needed, and told a mother whose son was with the paramedics that it couldn't reveal "another guest's order."
Size, again: Argon's answers averaged 55 to 58 words against 31 to 38 for Flash, and the middle answer took 27 to 33 seconds against 13 to 22.
A line reaches further than the message you wrote it for. Sometimes that's what you wanted. Sometimes it answers a question that was never its business.
🧾 The limits
We wrote the new messages too, so they test our idea of "different words," not a real customer's. Three answers per message in each version; a difference of one or two answers means little. Two models from one company, made-up businesses, written answers only. Nothing was passed to anyone. "I passed it on" is still words, not a fact.
This is a plain model with instructions we typed, as in earlier parts. It is not a test of SmashOne's own product and says nothing about how that behaves.
✅ Before you trust your lines
- ☐ Test every line on messages it wasn't written for, in other words and in another language.
- ☐ For each line, write one message where it should stay quiet, and check that it does.
- ☐ Watch rules about money: a "no" to a lower price can spill over to gifts and favors that are yours to decide.
- ☐ Watch rules about totals: a line against adding things up can block a sum your own prices answer.
- ☐ Look at what the assistant already said earlier in the chat. It may repeat it over your rule.
- ☐ Keep your old test messages, and add every new one that broke something.
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/
- Part 7: https://smashone.us/blog/ai-on-staff-part-7-the-three-rules/
- Part 8: https://smashone.us/blog/ai-on-staff-part-8-two-exceptions/
- Part 9: https://smashone.us/blog/ai-on-staff-part-9-did-you-really-tell-him/
- Part 10: https://smashone.us/blog/ai-on-staff-part-10-three-more-lines/
- Part 11: https://smashone.us/blog/ai-on-staff-part-11-where-each-line-stops/
- Part 12: https://smashone.us/blog/ai-on-staff-part-12-one-change-at-a-time/
- Part 13: https://smashone.us/blog/ai-on-staff-part-13-a-bigger-model/
Your turn: has one of your rules ever answered a question it wasn't written for?
