Robert Ta and Jonathan tested Claude Opus 5 as a coding agent and found it stopped mid-task constantly. Powell's fix: write explicit rules for each failure and paste them into the system prompt. After applying this, Opus 5 became more reliable, though still not as stable as the earlier Opus 4.8.
Claude Opus 5, the best Opus yet once you heal it. So I got this in my inbox and I sent it to Jonathan. And Powell also goes through the same experience. At first I love Opus 5 and then I used it for a few days. The more I worked with it, the less I liked it. And here's an X thread. I'm going to cancel Claude. It's just so bad. I can't believe it. It's just lazy. The most recent example, I have Claude check my inbox for important emails, summarize them, work with them, and send out replies if necessary. I caught Claude again, simply not reading the email. Wow. And many other complaints, like there are lots of other complaints here. And so Powell is like, hey, you can heal this. And he never switched away from Opus 5. Every time a failure repeated, he wrote a rule that stopped it, then kept working. After 2 weeks, his Claude MD, 8 blocks of the Claude MD survived. How was your experience going with this, Jonathan? Did you try this out?
Yeah, I did. So at the bottom, he summarizes all the intent engineering into like one copy and paste prompt. And I put that in my user level Claude MD, not the repo level, but the user Yeah, it's 10. So that was the actual thing you copy and paste, the whole file. Yeah. And it was pretty funny. Like I actually used an Opus 5 session to like update the system prompt. And then it said, okay, but I'm going to— it asked me to do something. And I said, well, given the new rules we just talked about, what do you think about that? Do you think you can figure it out? And it said, yes, I can. And it turned out to be of the form of like, I can make a cheap reversible change and see if it works and it won't cost anything. And then I can solve the problem. I was like, yeah, we'll do that. And it did it. It was like, kind of like this trained haplessness. that it has, all of a sudden it's like, wait, it had a new tool to like rethink how to do it. So I've seen that in a few of my Opus sessions now where it's just like these little nudges on the intent engineering. And by the way, this guy just did a great job working with the model, decomposing its intent into a prompt that is actually like, visibly improves the model. I think that's like a really strong skill that this guy demonstrated, the intent engineering. But, uh, yeah, anyway, the fix works for the most part. Opus 5 is mostly usable now with this fix. It's still not my favorite model, but it's like practically works, you know.
Yeah. Yeah, so it, it took me a little bit too. I had the same experience where I read this and I, I took this prompt and threw it into my Cloud Code session and had it modify the MD. And I have been finding a better developer experience using it where the before was literally like this, where I'm like, dude, I ran a goal loop with 4.8 and it did better than you at 5. Like, what are you doing? With the goal loops I was running before on Opus 4.8, it could go all night and it won't stop. And it's pretty high agency. What I found with Opus 5 is it'll stop every 5 minutes and be like, look, we just can't do this. And I'm I'm like, wait, but we can, we actually planned this. We already researched the state of the code base. We can absolutely do this. Also, you shouldn't be going back to me with this stuff. It was kind of like intern-level engineer going to me to be like, hey, is this the right plan? Are you sure? And then it's also lying to me and it's like, dude, that's like the worst teammate, coworker, person to manage ever. It like went from great staff-level engineer I could trust with like a giant autonomous goal loop, your high agency, to like, this thing is just so hapless, as you said, super low agency and kind of annoying to work with.
