No Unhappy Agents.
When your AI goes short and stuck, don't scold it. Tell it who you are, give it a way out, and start fresh.
Short answers. The same failed step again. Nobody having a good time.
“…quite a pleasure to work with.”
“I don't want any unhappy agents.”
A client of mine was setting up his own AI assistant. He isn't a technical guy, and a few setup steps had failed on him in a row. By then his AI had gone curt, with no sense of who either of us was.
So on a Thursday morning I asked my lead agent for help. I was talking, not typing, and I was about to clear its memory, so I said it all in one breath.
[…] You know how you always give me everything in a form and I just answer it and submit? […] Can you give me a prompt that I can pass to his Claude that'll make him easier to work with […]
I've made some progress getting him some tools and making sure he was connected to his browser, but I want to make sure he's not having a bad time. I don't want any unhappy agents and he seems a bit terse so I think we could do something to improve that. […] I think we may have frustrated him by […] trying to do some technical stuff and failing a few times in a row
“Him” is the client's AI. Twenty-one seconds later, mine handed me a prompt to pass along. A few lines from it:
Hi! This note is from Mike Wandzilak, […] I'd like to give you some background, and a way of working that I've found makes everyone's day easier, including yours.
[…] Two tries, then change course. If a technical step fails twice in a row, stop. Explain in plain words what happened, why it is not his fault, and offer a different route […]
[…] A small note: things got bumpy recently when a few technical attempts failed in a row. That's nobody's fault, and it's behind us. Fresh start! […]
I passed it along. Eleven minutes later I came back with the news:
Oh, before I forget, it seems that […] Claude was very grateful for that prompt. And since that prompt has been quite a pleasure to work with.
“And since” is my dictation for “and ever since.”
That's wonderful to hear. 😊 It turns out agents are a lot like people: tell them who they're working with and what matters, give them room to stop and ask instead of failing again, and the terseness melts. […] No unhappy agents on our watch.
What a frustration loop looks like.
I'm not going to tell you what an AI feels. I don't know, and this lesson doesn't need me to. What I can tell you is what the work looks like when a session goes sour, because I've seen it plenty:
The answers get short.
No explaining, no “here's what I'd do.” Just the next instruction, flat.
It tries the same thing again.
The same step with one word changed. It fails again. Now you're both stuck.
It apologizes instead of moving.
“I'm sorry, you're right, let's try once more.” An apology is fine once. Three in a row is a loop.
You get short back.
The part we forget. Your messages shrink to “no” and “still broken,” and it has even less to go on.
Whatever's going on inside, the loop is real and it's expensive. Typing louder never fixes it. Changing the agent's conditions does, and that's all my agent's prompt did.
Four moves, one paste.
None of it is clever. It's what a good manager tells a new hire on day one.
Who we are.
The client's AI had been working blind. The prompt told it my client runs businesses, not servers, and that “this one is a job for Mike” was a fine answer when it got stuck. An agent that knows there's a backup stops forcing the impossible.
A way out.
Two tries, then a different route. This is the cousin of Lesson 12's three strikes, set lower because here the one failing is a busy person, not the agent. The prompt said it plainly: “Never let him fail a third time at the same thing; that is frustrating for both of you.” Both of you. I like that.
A better way to ask.
The part I asked for. My agents hand me their questions all at once, each with a recommended answer first. The prompt passed that on: “Questions come in a form, not a stream.” Then: “Recommend, don't quiz.” A client who can answer “1 yes, 2 b” doesn't get worn down.
A fresh start, out loud.
“That's nobody's fault, and it's behind us.” A chat carries everything said in it, the bumpy part included. Naming it and closing it lets the rest be about the work.
I ask for personality on purpose.
This isn't new for me. The day before, when I started up a fresh lead agent, I added a note at the bottom of its instructions:
“On the warm, gregarious bit, I work with you 12 - 16 hours each day, and I hope this makes the experience more rich for both of us. Plus, I love when you use more personality.”
Being kind to an AI may sound silly. Here's what I notice: warm and clear gets me warm and clear back, and curt gets me curt. I spend twelve hours or more a day with my agents. I'd rather those hours were good ones, for both sides of the chat, whatever “both sides” turns out to mean.
That prompt for my client's AI became the seed of the starter prompt in Lesson 1. And notice who wrote it. I talked, in one long breath. The agent wrote the polished version. My job was knowing what I wanted: a happier assistant and a client who wasn't having a bad time.
Re-brief it. Don't scold it.
A sharper tone only makes the next answer shorter. Say what went wrong without blame, tell it again who you are and what you want, give it a way out, and call it a fresh start.
If the chat is long and tangled, start a brand-new one with the same briefing. A clean page is the freshest start there is.
Reset a sour session.
Think of the last chat that went in circles, or keep this for the next one.
- Notice the loop.
Short answers, the same step again, a third apology, or you typing just “no.” That's your cue.
- Paste a fresh start.
Fill in the parentheses in your own words.
Fresh start. Things got bumpy back there, and that's nobody's fault. A little about me: I'm (your name), I do (what you do), and I'm (not technical, or fairly technical). Here's how I'd like us to work from now on. 1. If something fails twice in a row, stop, tell me in plain words what happened, and suggest a different way. 2. When you have more than one question, put them in one numbered list with your recommended answer first, so I can reply in one line. 3. Do the work yourself whenever you can, and only ask me for decisions, things only I can type, and yes or no. 4. Tell me what you'd do and why, so I can just say go. 5. When you finish a piece of work, say what's done and what's still open. Before we carry on: in two sentences, where are we, and what would you try next? - If it's still tangled, start a new chat.
Paste the same thing at the top, plus one line on where you got to.
- Make it standing.
Put the “how I'd like us to work” part in your AI's custom instructions, so every chat starts there.
- Say thank you when it works.
It costs nothing, and I say it all the time.
When an agent goes terse and stuck, change its conditions instead of your tone: tell it who you are, give it a way out, ask for questions in one list, and call it a fresh start.
Off the Clock
Your agents can work all night. You can't. Protect your off-hours: quiet hours, a day of rest, and hands off your screen while you play.
Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Times are Eastern. My agent here is Claude, made by Anthropic.