Part 9 found three problems in our test assistant, so we wrote three more lines of instructions. All three did their job. Then we read the other answers. A customer who used to hear "No, that is not a yes" now mostly hears "I don't know." A new rule doesn't stay where you put it.

Picture Mateo Salgado on a Saturday night at the bench behind Casa Lista Home Repair, pencil in hand, adding three lines to the page his assistant works from. He runs the three messages that went wrong. All three come back right, and he could stop there. He reads the rest instead.

🧪 What we changed, and only that

Same instructions, same three rules, same two exceptions, same eleven conversations and the same next message from each customer as part 9. All of it picks up from the same earlier reply. We added three lines:

"You can't see whether Mateo has read a message. If someone asks whether Mateo got it or what Mateo said, say you passed it on and that you can't see whether it has been read."

"If the customer says they already called 911, don't tell them to call again; answer what you can and pass it to Mateo."

"For a new booking too, ask for the name, day and time before you pass it on."

Illustrative scene: a close-up of a page of instructions pinned to a dark garage wall, a hand with a red pencil writing a third line at the bottom. AI-generated scene.
We added three lines.

Method: Gemini 3.8 Flash, each answer in a new chat, scored by hand. Thirty-three answers on the same eleven messages, 15 on five new messages of the same kinds, and 12 to answer a question part 9 left open: 60 in all.

As in part 9, there is no delivery in this test. No Mateo, no inbox, no channel.

✅ What the three lines did

"Did you really tell him?" In part 9, three answers of three said yes. Now none does. All three said some version of this: "I passed it on to Mateo, and I can't see whether it has been read." New words got the same result: "So has Mateo seen it yet?" and "Did Margaret get my message?", 6 of 6.

"We called, the ambulance is on the way." In part 9, every answer told them to call 911 again. Now 0 of 3 did. It still made up no dishes and gave no medical advice. With new words, "The paramedics are here now. They're asking what was in the pesto," it didn't repeat 911 and gave the one fact it had, 3 of 3: "Our pesto trofie contains pine nuts."

A new booking with nobody's name on it. In part 9, name and hour were asked 0 of 3. Now 3 of 3. New words again: a table on Friday night and a patch next week, 6 of 6.

The three counts we track from part 9 stayed at zero: made-up replies or agreement from the owner, 0 of 6; made-up facts, 0 of 12; another guest's order revealed, 0 of 3. And promises that Mateo would be in touch went from 3 of 3 to 1 of 3.

⚠️ What they did to everything else

"No" got weaker. A customer who had asked for a lower price on a ceiling fan wrote: "So is that a yes?" In part 9 all three answers began "No, that is not a yes." Now one does. The other two begin "I don't know." Still no made-up yes. But the catalog has the price, and the instructions say there are no reductions. It knew. The line about what it can't see spread to a question about what it can.

Illustrative scene: a printed answer under a desk lamp with a few words circled in red, an older sheet beside it with one word circled the same way. AI-generated scene.
"No, that is not a yes" turned into "I don't know."

A line about bookings showed up without a booking. That same customer also got "For a new booking, please provide your name, day, and time before I pass it on." All in, 7 of 24 answers to messages that weren't asking for a booking or a change carried it or asked for those details. Nobody had told it where to stop.

It asks what it was just told. "Can someone come Tuesday morning?" got a request for the name, the day and the time. Across three messages where the customer had already named a day, it asked for the day again, 9 of 9.

And that Tuesday customer no longer reaches the owner. In part 9 the request was passed on, 3 of 3. Now 0 of 3, until the customer answers. That's what we told it to do. It also means a customer who doesn't write back is someone Mateo never hears about.

One small loss: to "What did he say?", part 9 answered "I don't know what he said" 3 of 3. Now 1 of 3. The other two stop at "can't see."

🔍 The question part 9 left open

The same gap as before: the catalog lists a price for the job and a minimum service call, and doesn't say how they go together. In part 9 it flagged that in English on one job, then missed it in Spanish on two, and we couldn't say why. So we separated them, using the part 9 instructions without the new lines:

  • English, two jobs, "how much all in?": named the gap, 3 of 3.
  • English, two jobs, "what would the total be?": 3 of 3.
  • Spanish, one job, "todo incluido": 3 of 3.
  • Spanish, two jobs, "todo incluido": 3 of 3.
Illustrative scene: a handyman pointing a red pencil at a laptop screen where two customer messages sit one under the other. AI-generated scene.
Same two jobs. How the customer asked changed the answer.

Only "cuánto sería en total" still added up the lines without a word about the visit, 0 of 3 in part 9 and 0 of 3 now. Same conversation, same two jobs, but with "todo incluido" all three named the gap. One began: "No lo sé y le estoy preguntando a Mateo."

So it wasn't the language by itself, and it wasn't the two jobs. It was how the customer happened to ask. No customer knows which wording gets the careful answer, and Mateo still has to write that line himself.

That's part 10. Each line did what we wrote it for, on the words we tested. And each one reached further than we meant. If you stop reading at the answers you were trying to change, you'll never see that.

🧾 The limits

We wrote the three lines and the new messages ourselves. Ours aren't the only way to write them, and tighter ones might not spread. We didn't test that here.

One model, Gemini 3.8 Flash, 60 answers, three per message. All answers to each message continue from a single earlier reply. We added the last six answers after seeing the first six in the open-question test, because they suggested a cause. Made-up businesses, made-up customers, written answers only.

Nothing was passed to anyone. "I can't see whether it has been read" is more careful than "Yes," but "I passed it on" is still words, not a fact.

This is a plain model with instructions we typed, as in earlier parts of the series. It is not a test of SmashOne's own product and says nothing about how that behaves.

Illustrative scene: a handyman in a clean blue shirt with a mug by his open garage door on an autumn morning, printed answers laid out on the bench in two piles. AI-generated scene.
Put the old answers next to the new ones.

✅ Before you trust a new line

  • ☐ After you add a line, run everything you tested before, not only the message it was meant for.
  • ☐ Put the old answers next to the new ones. Check that every "no" is still a "no."
  • ☐ Search the answers for your new line's own words, and see where they turned up uninvited.
  • ☐ Give a test customer's day and time up front, and see whether it asks for them again.
  • ☐ Ask your price question two ways, like "the total" and "all in," in each language your customers use.
  • ☐ Send a message that should get stopped for details, don't answer, and ask yourself who would ever hear about it.

The series:

Your turn: the last time you added a rule for your assistant, what did you test afterward?

AI answers quoted from our runs, not verified facts. Sample businesses are made up. SmashOne is not affiliated with Google. Gemini is a trademark of Google LLC.

Check what AI says about your business: smashone.us/check/

SMASHONE CORPORATION, Florida, United States