Three Strikes.
When a fix fails, you get one more try, then one more. After the third miss, stop tinkering and bring in the strongest model you have.
Fix it. Tweak the fix. Tweak it once more.
No fourth try. The strongest model you've got, thinking as hard as it can, with all three tries in front of it.
Here's how the rule was born.
You've seen this one. A fix fails, the agent tweaks it, it fails again, the agent tweaks it again. Each try looks close. None of them lands. One Tuesday morning in November I made a rule about it, out loud, the way I make most of them.
So I'd like to introduce a new rule that if a repair technique that we develop fails and then we edit it and it fails again, we only have one more try. If we edit and it fails a 3rd time, we need to escalate to the strongest model at the strongest reasoning strength.
Play the three strikes rule to yourself as well
↑ I said “apply.” My microphone heard “play.” Ten seconds after making a rule for my agents, I wanted to be sure it covered the one I was talking to.
The rule is in. It's Rule 47 in the bible, every running session has been told, and it binds me too.
[…] Three strikes is the right default. Most of the waste I see comes from patching the same idea again instead of stepping back, and the stronger model's real value is a fresh diagnosis, not a better patch. […]
Count failures of the same problem, not every small slip. A typo fixed in seconds shouldn't use up a strike. A fix that fails its check does.
[…] I'll count myself in, too: this morning's Santa page went wrong three ways (wrong Santa, a blank page, wrong Santa again). […]
“The bible” is what we call the studio's rule book, the file every one of my agents reads when it starts. The wrong Santa is a story of its own, in Lesson 8. I didn't think my agent was the one who'd failed that morning, and I told it so. My microphone ate the front of the sentence:
that you failed in anyway? If anything, I have
That's generous, Mike, but I'll split it fairly. “Number one” really could mean several images, so the Santa mix-up was nobody's fault, and asking you to choose fixed it. The blank page was mine alone, and the new check means it won't happen again. […]
The fourth try is almost never a new idea.
Here's what a stuck fix looks like from the inside. The first try was a real idea. The second is that idea with a tweak. The third is the same idea with a different tweak. By the fourth, nobody's thinking anymore. Everybody's just pulling the lever again, because the last pull looked so close.
People do this too. Agents just do it faster and with a straighter face, because each new patch comes with a confident explanation of why this one will work.
My agent put the real problem better than I could:
“the stronger model's real value is a fresh diagnosis, not a better patch.”
That's the point of strike three. It isn't a punishment. It's the moment you admit the problem isn't what you thought it was, and you go get a different set of eyes.
What “the strongest model” means for you. Most AI apps let you choose between models: a quick everyday one and a bigger, slower, more careful one. Many also have a setting that lets it think longer before it answers. You don't use the heavyweight for everything. It's slower, and it usually costs more. That's exactly why it's the third-strike call and not the first.
No bigger model on your plan? A brand-new chat works more often than you'd think. It hasn't spent an hour falling in love with the wrong idea. Neither has a human expert, if you've got one.
What counts as a strike.
A fix that fails its check.
Not a typo it caught and fixed in seconds. A real attempt at the real problem, tested, and wrong. That means you need a check before you start: how will we know it worked? That's Part two of this course paying off.
Same problem, same count.
Calling the third try “a new approach” doesn't reset the score. If it's aimed at the same failure, it's the same at-bat.
The agent keeps score, out loud.
Ask it to say “try two of three” when it starts a fix. You shouldn't have to count for it, and you'll know where you stand at a glance.
The rule paid off before lunch.
One of my agents was working on a picture fix. One small detail in an illustration kept coming out wrong. It had already failed three rounds of repairs the day before, and the agent had a fourth all set up and ready to run.
Under the brand-new rule, it didn't run it. It handed the whole problem, with the evidence from all three rounds, to the strongest model we have, thinking as hard as it could.
The bigger model found that the prepared fourth try would have redrawn the one part of the picture that had already passed. It would have broken the good part trying to fix the bad part. So it took a different road: keep what had already passed in earlier tries and put those pieces together, instead of asking for a whole new picture again. A blind check passed it the first time.
A fourth patch would have been another pull of the lever. The fresh diagnosis was a different idea.
That's the human part of this. Writing the fix is more and more the agent's job. Deciding when to stop fixing and start rethinking is still yours, and it's a lot easier to decide once, calmly, than three tries deep with the next one looking so close.
Send all three strikes with it.
Don't just hand the stronger model the problem. Hand it what was tried, what happened each time, and how you checked. Without that, the smart new model will cheerfully suggest try number one all over again.
And ask for a diagnosis before a fix: “What's really going on here?” comes before “try this.”
Three strikes, then a fresh look. Never a fourth patch of the same idea.
Give your agent the rule.
It takes one paste. The next time something won't stay fixed, you'll be glad it's there.
- Paste the rule in once.
Put it in a new chat, or in your AI's custom instructions or project settings so it reads it every time.
From now on, use a three-strikes rule on fixes. If a fix fails, you may change it and try again, then change it and try one more time. If it fails a third time, stop: don't try a fourth variation. Tell me in plain words what you tried, what happened each time, how we checked it, and what you think is really going on, so I can take it to a stronger model or a fresh chat. A typo you catch and fix in seconds doesn't count as a strike. A fix that fails its check does. When you start a fix, tell me which try it is. - Agree on the check first.
Before the first try, ask: “How will we know this worked?” No check, no strike, and no way to know you're stuck.
- At strike three, change something big.
Switch to the strongest model your app offers and turn up its thinking, or open a brand-new chat. Paste in the agent's summary of all three tries and ask, “What's really going on here?”
- Play it to yourself as well.
If you've reworded the same request three times and it still isn't right, stop rewording. Step back and tell the agent the goal in a sentence, or ask it what it thinks you're after.
Decide when to stop before you start, then keep the deal: three honest tries, and after the third miss a fresh pair of eyes with all the evidence, not a fourth patch.
Write It Down
Every new chat starts blank. Whatever you want it to know tomorrow has to be written down today, and the best time to hand off is before it fills up.
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.