The Right Brain for the Job.
Your strongest model for the hard thinking, an everyday one for the routine work. One glass, not the keg.
Writes the plan. Checks the result. Goes home.
Does the typing in between.
Let the big one out, but keep an eye on him.
Most AI apps give you a choice of models. The strongest one I use is called Fable. It's brilliant, and it's hungry. My plan gives it its own weekly allowance, smaller than the one my everyday models share, and it eats through it fast. A few days earlier I'd grumbled to one of my agents:
“hopefully we don't need Fabel too much longer for design because he's a hog.”
Yes, “Fabel.” Yes, “he.” Then, one Wednesday, I paused it everywhere, in eight words: “pause any fables and replace them with opus'”. Opus is my everyday model. It's very good, and it doesn't drink the bar dry.
The next afternoon, something hard came up. A problem in my studio's avatar work had failed three fixes in a row, and my own three-strikes rule from Lesson 12 says that's when you bring in the strongest model you have. So my lead agent asked me to lift the pause. I answered into my microphone:
Lift the fabul pause. Fable
“Fabul” is my speech-to-text having a go at “Fable.”
[…] “Fabul”: close enough. Your speech-to-text is batting about .500 today, and I'm reading the rest from context, which is the whole course in action.
Done: the Fable pause is lifted. […]
I've read your answer as lifting the pause for good: Fable is back for design, research and three-strike escalations, and Opus stays the default for routine work. If you meant only this one review, say the word and I'll narrow it.
Keep an eye on that guy. He'll drink the whole party's booze if we're not careful.
[…] Noted. Fable gets one glass, not the keg:
Only one Fable reviewer. Any helpers it uses run on cheaper models.
Its job is a written plan, not grinding. It thinks hard once, hands over the design, and goes home.
I'm watching the meter. Fable sits at 3% of its weekly limit right now. If it climbs past 15%, I'll tell […] to wrap it up before it raids the liquor cabinet.
Look at what happened in the middle there. I gave a five-word order with a slip in it. My agent didn't just do it. It told me how it read me (lifted for good, not for one job) and offered to narrow it if it had guessed wrong. Infer the meaning, confirm the decision. And when I made a joke, it turned the joke into three working limits.
The gate, not the typist.
So why not use the strongest model for everything? About a week earlier, I'd raised exactly that with the strongest model itself:
[…] Regarding my fable / opus quandary. There's a lot of stuff here that Sonnet could do, maybe even haiku could do, but all of it has a standard that I believe only you can achieve
Sonnet and Haiku are the smaller, quicker, cheaper models in the same family.
That's the kindest thing anyone's said to me tonight, and I'm going to push back on it a little anyway, because I think the standard is more portable than you're giving it credit for.
[…] That's the standard, and it lives on disk, not in my weights.
So the question isn't which model is good enough to type. It's where the standard gets enforced. […] Fable writes the brief, tight enough that there's nothing to interpret. Opus or Sonnet does the typing. Fable reviews against the standard before anything ships, with the author's conclusion hidden so the reviewer isn't anchored. The bar is held at the gate, which is one Fable call, instead of at every keystroke, which is a thousand.
A model telling me to use it less. I liked it even more for that.
“Lives on disk” is the key. My standard isn't a feeling in the strongest model's head. It's written down: in my studio's rule book, in checklists, and in little checking programs that test the work no matter who did it. In the same reply my agent put it in one line: “The more of the bar that's a checker, the less it matters who typed.”
So the split in my studio looks like this. The hard thinking goes to the strongest model: the plan, the design, figuring out why something is really broken, and the final check against my standard. The long middle, where the plan gets carried out, goes to the everyday model. It's like hiring a great architect. You want the architect to draw the house and walk through it before you move in. You don't need the architect to hang every sheet of drywall.
Let the agent choose, and say which.
Here's a confession. I've been writing code for over forty years, and I still can't tell you which model a given job needs. A few days after that conversation, I stopped pretending, and I had my agent write it down for good. My words:
“Yes, and also write a reminder that Mike does not know which model to use for any given task”
That wasn't me giving up. It was me putting the decision where the knowledge is. Now my agents pick the model and how hard it should think for every helper they start, and they tell me in one line which one they chose and why. If only I can flip the switch, they tell me exactly what to pick. I read the line. If it seems off, I say so. Same deal as everything else in this course: the agent does the technical part, and I keep the judgment.
You don't need to know the model lineup either. You need to know two things: which parts of your work are hard thinking and which are routine, and what “good enough” looks like. Those are human jobs.
One glass, not the keg.
When you bring in the big model, give it a job with edges. One of it, not a crowd. A written plan or a review, not hours of typing. Someone watching the meter, with a point where it wraps up and hands in what it has.
And when a fix has failed three times, that's the other time to call it in, with all three tries in front of it. That's Lesson 12.
Match the brain to the job.
Pick something you're working on that has a hard part and a lot of routine parts. Then:
- See what your app offers.
Most AI apps have a quick everyday model and a bigger, slower, more careful one. Many also have a setting that lets the model think longer before it answers. Find both.
- Let the agent sort the work.
Paste this at the start, and let it tell you the plan instead of asking you to choose.
Before we start, look at the work ahead and split it into hard thinking and routine work. Hard thinking means planning, design, figuring out why something is really broken, and the final check against what I asked for. Tell me in one line each which model, and how much thinking time, you'd use for each part, and why. Save the strongest model for the hard parts, and use the everyday model for the routine work in between. If you can't switch models yourself, tell me exactly what to pick and when. Don't ask me which model to use: tell me what you'd pick. - Big brain plans, everyday brain types, big brain checks.
Have the strong model write a plan tight enough that there's nothing to guess. Switch to the everyday model to carry it out. Then switch back and ask the strong one to check the result against the plan.
- Write your standard down.
A short list of what “good” means to you: the tone, the must-haves, the never-dos. Paste it in every time. The more of your bar that's written down, the less it matters which model typed.
- Watch the meter.
If your app shows how much of your allowance you've used, look at it now and then, or ask your agent to tell you when the big model has used a set share of it.
Put your strongest model at the gate, writing the plan and checking the result, let the everyday model do the typing in between, and let the agent pick and tell you which, with someone watching the meter.
One Writer per File
When several agents work at once, each one claims its files, and only one ever writes to a file at a time.
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; Fable, Opus, Sonnet and Haiku are Claude models.