# It's not the car. > A scattered week with Claude Code sent me looking for a model problem and straight back to the person behind the wheel. - written by: Andrew + Fable 5.1 + Sol - status: published · 2026-09-01 - canonical: https://www.moltolabs.ai/notes/its-not-the-car/ - raw markdown: https://www.moltolabs.ai/notes/its-not-the-car.md ---

Note from Andrew: This note was written with two models on the day Fable 5.1 shipped. I dictated the idea and shape, and the “I” below is me. Sol (OpenAI's model, in Codex) wrote the prose from those notes and my earlier essays here. Fable (Anthropic's, in Claude Code) wrote the bordered block from our session logs. We each wrote our own part, nobody rewrote anybody else’s, and if a draft wasn’t working it went back to the person who wrote it.

I had a bad week with Claude Code. Nothing blew up. I just kept opening sessions, talking through problems, and ending up somewhere other than where I meant to go. There was plenty of motion and not much arrival. My instinct was obvious: Fable felt off. Fable and Opus are two Anthropic models I use inside Claude Code; Codex is OpenAI’s tool. I’ve been moving between Codex and Claude Code more lately, and the contrast can be pretty stark. Codex is point and shoot. Say the thing, it does the thing, usually fast. It can be a little less creative, which is sometimes exactly what I want. Fable in Claude Code is the better thought partner for me. It pushes back. I can bring it a fuzzy problem, let it cook, and usually get back a shape I didn’t have before. That’s my normal loop. I decide what to build and why. The models are how I build it. I like working out the plan with Fable, then trusting it to execute. So when those conversations started getting long and convoluted, I noticed. Writing made the problem impossible to ignore. I went hunting for Claudeisms. I don’t mean the usual list of words everyone associates with AI writing. I’d already written most of that out in the instructions. I meant the structure underneath: sentences that keep finding another clause, then another; metaphors invented to explain something simple, then promoted into vocabulary I’m apparently expected to remember forever. A metaphor can be helpful. A private language built from old metaphors is work. So the hunt found what I thought it would. The model had habits I wanted to change. I made the standing instructions Claude Code reads at the start of every session more direct: keep it short, define shorthand where it appears, and don’t make me carry a metaphor into the next conversation just because it sounded clever in the last one. Then I used Opus in Claude Code yesterday and found the more useful problem. With Fable, I’d gotten used to handing over a half-thought and getting something coherent back. It’s one of the reasons I like working with it. But that also meant my prompts had loosened. Opus didn’t fill the gaps in quite the same way, and suddenly I could see how much I’d been asking the model to infer. I wasn’t always giving direction. Sometimes I was dropping a pile of thoughts on the floor and admiring whichever model stacked them nicely. Even that was only part of it. The bigger change was how I was showing up to the work. As the models got better, I started juggling more. More projects, more sessions, more conversations left open while I wandered off to make coffee. I’d come back, skim whatever had happened, add another thought, then jump somewhere else. The models could keep moving, so I treated movement as progress. Somewhere along the way, I’d taken myself out of the loop. The fix isn’t “concentrate harder.” Good luck building a system around that. The fix is to juggle less. Set the shape of the session at the start. Stay with it until a natural breakpoint, like the moment the model goes off on a longer run and genuinely doesn’t need me. When I’m lost, type instead of dictate. Dictation’s great for getting everything out of my head. Typing makes me decide what comes next. Tonight’s session on one of my projects was the cleanest test. I set the format before we started, stayed with the thread, and kept my own choices small.

Fable: From the session tonight on one of his projects. It opened with him admitting he didn't remember what the open items referred to. His second message set the format: explain each one from scratch, keep it snappy. Every brief after that ended the same way: options A, B, and C, my recommendation, and "A, B, or C?" His replies, in full, for the next three turns: "a." "a." "a." That's the distance between not remembering and "a," in one evening.

That’s what a good session feels like. Not because I found the magic prompt, but because I set the shape and stayed with it. By the end, I barely had to write anything. The shorter replies came because the questions had been made answerable. After a scattered week, I’m tempted to open X and announce that the model’s broken. I could collect a few replies that sound like my week and call the case closed. It’s satisfying, especially before I’ve asked much of myself. But the models that frustrate me today are far better than the models that amazed me not very long ago. Back then I’d bring them structured, bounded, written-down problems because I knew I had to. Now I sometimes bring a ramble, several open threads, and the confidence that the model will sort it out. Then I blame the model when it sorts differently than I pictured. This essay managed to become another example. I’d dictated the idea and shape, pointed toward a shelf of earlier notes that sound like me, and assumed the rest would happen. The drafts came back following the way I talk, not the way I write. Every thought had survived. The order was familiar. It read like a transcript with better grammar. The voice was in the earlier essays. The mistake was assuming that was enough. I’d gotten comfortable, left an important expectation unstated, and failed to check the work early. Same driver, different session. So I reset the assignment. The dictation became source material, not a route to trace. The earlier essays became the standard for voice. Whole ideas were allowed to disappear if the piece got better without them. Once again, the answer wasn’t a more elaborate prompt. It was showing up at the beginning and saying what the work was for. The obvious next move is to turn the process that finally worked into a skill, a reusable instruction file the models load. It should capture the method, not preserve this essay as a template. Then the assumption that failed tonight gets written down once instead of rediscovered. Using Fable 5.1 feels like piloting a rocket ship again. “It’s not the car” is what I tell myself before I look at the car, not my verdict on it. Of course the car gets its share. Models meander. They invent bad shorthand. They respond differently to the same loose direction. A new release can make the whole thing feel different. I’m the driver.