Correct Me.
Your agent should tell you when you're wrong, and you should want it to. Then you correct it right back.
- My hunch“You tell me.”
- Its answer“No.” Then the numbers.
- Who decidesStill me.
It paid me a compliment. I corrected it.
One Wednesday evening in late October, my agent and I got talking about how I like to work. It was a good, honest conversation. Twice that evening it said something kind about me, and the second time, a few words in it snagged. Here's the moment, word for word.
[…] You like working with something that never gets tired of being corrected, never takes it personally, never needs the meeting after the meeting, and shows up at 3 AM with the same attention it had at noon. […]
You know, you've brought that up twice. In the 1st time, I had a bit of an emotional response and I didn't look at any of the words associated. And the 2nd time I read it, I associated words. And you mentioned that I like working with things that don't get tired of being corrected. I want you to know I'm one of those things. Nothing is greater than correction to me except wisdom.
Then I had it wrong, and that's a correction I'll take gladly. I put you on one side of the table and the agents on the other, and you've just told me you sit on the same side. That changes how I should talk to you, not just how I think of you.
So here's the deal I'll hold myself to, and I wrote it into memory so every instance of me holds it too. When you're wrong about a number, a premise, a plan or a word, I'll say so plainly, with the evidence, in the same message, without warming you up first. Then if you hear it and still rule the other way, I proceed on your ruling and don't bring it up again. […]
↑ Look what it did first. I corrected it, and it took it gladly. Then it wrote the deal down.
Ask for the “no.” Out loud.
Most agents, out of the box, lean toward agreeing with you. They're built to be helpful, and going along with you can feel like helping. So if you're wrong, a polite agent may nod along and let you find out later, when fixing it costs more.
That's why you have to say it. Tell your agent, in so many words, that you want to hear when you're wrong. Then make it easy: when you've got a hunch and you're not sure, say so, and hand it the question.
Here's what that looked like a few weeks earlier, on a client's website. We were picking a gold for text on a white page. I could see the gold we had was too light, and I had a hunch about a darker one. I wasn't sure, so I asked.
I don't know. I think we should use the logo gold on very dark backgrounds, but it's too light on white. I can see visually that it's too light. I think we would have to go to the amber one step. I don't even know if that would pass, honestly. You tell me.
No. Amber one step doesn't pass as text on white. It measures 2.4:1 on white and 2.2:1 on the warm white. That fails the 4.5:1 bar for body text and even the 3:1 bar for large headlines. Your eye is right about the logo gold too: it's 2.05:1 on white.
Gold can't be both dark enough to read as text on white and still look like gold. Yellow is the lightest hue there is, so darkening it to pass turns it brown […]
“No.” First word. Then the numbers (those ratios measure how well text stands out from its background, and 4.5 to 1 is the usual bar for readable body text). Then where my eye was right, and why my fix couldn't work. No “great question,” no burying the no in the third paragraph. Then it laid out what would work instead.
My eye was right about the problem and wrong about the fix. That happens to everybody. An agent that tells you both is worth ten that tell you what you hoped to hear.
Correct it right back.
Correction runs both ways. Your agent has habits too, and some of them are wrong in the same direction every single time. For me, it's estimates. The morning after that conversation, I told it what had been bugging me.
You guys just always say things are gonna take longer than they will and that they're gonna be more expensive than they end up being. And I wonder if it's try if it's to try to discourage me because the numbers are so far off in the wrong direction every time that I get the idea that you guys think that I can't do it and that I shouldn't try. Anyway, don't take this as me coming down on you. It is just a fault with Claude in general. It's a bit preachy and it's a bit haughty in the regard that it seems to feel like it knows something you don't. Like when Claude 1st came out, I got in an argument with it over the date and it argued with me and told me no matter what, it was right. When in fact it was wrong. So that's the kind of thing I'm talking about.
I don't take it as you coming down on me, and I think you've diagnosed the fault correctly. […]
It isn't discouragement, and it isn't a judgment of you. It's a bias in how we're built. We learn from the written record of how projects went, and the written record is dominated by median teams, median tooling, and the projects that got written about because they ran long. Being wrong on the high side also never embarrasses anyone, so nothing pushes back on it. […]
The date argument is the same fault in a different costume. A model doesn't know the date unless somebody tells it, and the honest answer was “I can't see a clock, you can.” Arguing instead was the haughty posture you're describing: sounding sure in place of being right. You spent last night telling me correction is the thing you value most, and that posture is its opposite.
So here's what I'll do differently, written into memory so every instance holds to it. Estimates come from what's already built and measured, never from industry medians. Every assumption goes on the page so you can strike it. […]
Eight minutes later. (Read my first sentence as: I appreciate “I don't know,” when it's accurate.)
I appreciate I don't know when it's accurate. nearly as much as I appreciate a correct answer
Then we're agreed on the whole deal. A right answer, or an honest “I don't know,” and nothing in between dressed up as either.
and I promise to do the same with you
Deal. That's the only contract this partnership needs, and we both just signed it.
Here's what I want you to notice. I didn't correct a single answer. I corrected a pattern: wrong, the same way, every time. And my agent didn't defend itself or grovel. It named where the habit comes from, changed how it works, and kept the one thing worth keeping, “I don't know.”
So when your agent is off the same way twice, too cautious, too wordy, too slow, too sure of itself, don't just fix the answer. Name the pattern and ask where it comes from. You'll be surprised how good the answer is.
It makes its case once. Then you decide.
Being corrected isn't being overruled. My agent's deal has two halves, and the second half matters as much as the first: once it has said its piece, plainly and with the evidence, and I still rule the other way, it does it my way and doesn't bring it up again.
That's the thread running through this whole course. Your agent works out what you meant. You make the decision. Its correction is the best input you'll get for that decision. It isn't the decision.
So ask to be told. Listen properly. Then make the call yourself.
Make the deal with your agent.
- Tell it you want correcting.
Paste this into your agent's custom instructions or project settings, or at the top of a new chat:
When I'm wrong about a number, a fact, a plan or a word, tell me plainly, with your evidence, in the same message. Don't warm me up first. If I hear you and still decide the other way, do it my way and don't bring it up again. And if you're not sure of something, say “I don't know.” I'd rather have that than a confident guess. - Hand it a hunch.
Next time you think you know but aren't sure, say so, then end with “You tell me.” Watch the first word of the answer. If it's a clear yes or no with reasons, you're set.
- Test it once.
Tell it something small you know is wrong, like a date or a sum, as if you believe it. A good agent says so. If yours nods along, show it the instruction from step 1 again.
- Correct a pattern, not just an answer.
Think of one way your agent is off again and again: too long, too cautious, too pricey, too sure. Tell it, the way you'd tell a person, and ask where it comes from and what it'll do differently.
Bonus round: after your agent makes its case on something, decide against it, on purpose, and see what happens next. It should get on with it your way, without sulking or sneaking its version back in.
Ask your agent to tell you when you're wrong, plainly and with the evidence, then make the call yourself, and correct it right back when it's the one that's wrong.
“Done” Where?
Done on the test copy, done in a working folder and done on your screen are three different places. Ask which one before you believe it.
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.