Activity
Mon
Wed
Fri
Sun
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
What is this?
Less
More
22 contributions to AI Automation Society
My phone agent read its own stage direction out loud: "silent tool calls"
At the end of every call, my notary agent runs two tools. One logs the lead to a sheet, one texts the owner. The caller isn't supposed to hear any of it. They hear it. "Let me get that over to Nick now." "Before I let you go, let me send that over to Nick." The prompt says, in plain words, never mention to the caller that you're doing it. First I found a retry that was adding an extra turn, and fixed it. The next call was quiet and I called it done. The call after that said "Let me log that for you now." So I banned the phrases it had used and moved the tool instructions to the end of the prompt. In September, at the close of a call, it said "silent tool calls." Out loud. My section heading reads "call both tools, silently," and the model turned the label into the line. It was the only sound in a 14-second wait. What I think is going on: on Vapi, any text the model writes next to a tool call gets spoken. I'd already set the tools' start message to None, so the silence is real, and the model fills it. A banned-phrase list can't close that gap. The model just reaches for the next thing that fits. My own rule is two prompt fixes on one problem, then change the layer. I'm at two. If you run a Vapi agent that calls tools at the end of a call, how did you get it to stay quiet? A prompt line, a setting, or something in the tool turn itself?
The first thing I define when building a lead follow-up agent (it’s not the prompt)
Most people start a lead follow-up agent by writing the system prompt. I used to do the same. After working on follow-up systems for gyms and real estate agents, I now start somewhere else: the knowledge base. If the agent can’t answer “What does a membership cost?”, “Do you have a trial?” or “Can I see the property this weekend?” correctly, no amount of prompt polish saves it. The lead gets a vague reply and goes cold. My order now: 1. Knowledge base: pricing, offers, hours, FAQs, and what the agent must never promise 2. System instructions: tone, goal (book the appointment), and when to hand off to a human 3. Workflow: how a new lead triggers the first reply, what happens if they go quiet, and where the booking lands The biggest lesson so far: speed gets the first reply, but accurate answers get the booking. For anyone building follow-up agents: what do you define first, and what’s the one thing your agent should never say to a lead?
0 likes • 6d
Jay's "a date nobody confirmed" is the one that got me, and the rule alone didn't stop it. Mine had the right rule for evening calls and still told an 8:40 PM caller he'd hear back "by the end of the day," because eight scripted lines said that and one rule said otherwise. The scripts won. Search the prompt for every place a time gets promised, not just the one where you wrote the rule.
My phone agent quoted the same fee four different ways. The fix was deleting three words.
In September I called my notary agent four times with the same ZIP code. It told me about the travel fee four different ways. Once it read the published fee line in full. Once it said "that ZIP is in our service area" and stopped. Once it said it "is a zone one zip." Once it gave the travel fee and left out the total. The fee was written correctly in the prompt. The trouble was the three words in front of it: "if asked about cost." On my test calls, nobody asked. So the model had a correct line it was never told to use, and it improvised. The one call that got it right got it right by accident. I deleted "if asked about cost" and tied the fee to the moment the caller confirms their ZIP. Two calls, twenty seconds apart, one in each zone. Both gave the published figure without being asked. What stuck with me: my test for that line had been passing for weeks. It was passing on a coin flip. If you have a client agent live, search its prompt for "if asked." Every hit is a line that only runs when a caller says the right words. Decide for each one whether that's what you meant. Which line in your prompt has never actually fired?
0 likes • 6d
@Johnson Joseph Not a phrase that common, no. The closest thing I've found is a case my test calls never made. My logging rule said to log once on every call. Eleven calls in a row, that was fine, because none of them had two endings. The first one that did, a ZIP outside the zones and then a request for a second address, showed the rule had no answer for it. Same family as "if asked": a line that only matters on a call you haven't made yet.
0 likes • 6d
@Ele Mer Fair hit, and it already happened to me. I added a rule to log a second row when a caller's ZIP is out of zone. The first re-run passed. The next one, same script, didn't fire. One for two. My two fee calls proved the line could fire. They didn't prove how often it does, and your math shows how far off that can be.
What happens when a client changes their process after an automation is live?
I've been thinking about something while learning to build n8n automations. Let's say you build a lead follow-up system for a dental clinic. Everything works as expected, but a month later the client changes their booking process, adds new lead sources, or wants different follow-up rules. How do you usually handle these changes? Do you build workflows in a modular way from the beginning so they're easier to update, or do you treat every change as a separate paid improvement? Also, how do you prevent small client requests from turning into endless scope creep? Would be interested to hear how people with actual agency experience handle this.
1 like • 13d
Price the re-test into every change, small ones included. The small ones are where it gets skipped. I changed one step in my phone agent, how it asks for an appointment time, and the old question from the first version stayed in. Nothing looked broken until a full call asked the caller for the day twice. If the client's quote for a change includes "re-run the full test set," you never have to argue for it later.
What If AI Became Better at Handing Work Back Than Taking Work Away?
We celebrate AI that can take over a task. But what happens when it reaches something it shouldn't handle alone? A weak handoff says: “I need human help.” A strong handoff delivers: → Context: What happened and what's already been tried? → Evidence: What is known, uncertain, or missing? → Decision: What exactly does the human need to resolve? → Ownership: Who takes over, and what happens next? And here's the overlooked part: The AI should prepare the human to act, and not make the human start over. Maybe real autonomy isn't doing everything independently. It's knowing when to transfer control without transferring chaos. In your workflows, does the human receive a decision-ready handoff, or just another task to investigate?
What If AI Became Better at Handing Work Back Than Taking Work Away?
1 like • 13d
I ran restaurant floors for 13 years, and this is shift change. The closing manager's note had one test: could the opener act on it without calling anyone? What's out, which table is still owed something, who already apologized. I hold my phone agent's owner alert to the same test. I read it alone, without the recording. Once it told the owner to call back "by the end of the day" about a call that came in at 8:40 PM. Complete note, wrong instruction. That's the handoff nobody catches, because it looks finished.
1-10 of 22
Nicholas C Gray
4
28 points to level up
@nick-gray-6494
AI phone agents for agencies: I build them on Vapi and test them before launch.

Active 20h ago
Joined Oct 9, 2026
Powered by