Update Claude Instructions for Opus 5 Model

We keep adding instructions to make AI smarter, but sometimes the smartest move is to remove the outdated ones. As models evolve, our prompts and workflows should evolve too. Less clutter. Better systems. Better results. I learn alost form Charlie Hills 🦩

Anthropic deleted 80% of Claude instructions. It got better. I built a (free) skill that updates yours: "Every six months, delete your CLAUDE.md. Delete your skills. Delete your hooks." That's exactly what Boris Cherny (creator of Claude Code) recently said about the latest Opus 5 model. Because every instruction in your setup was written to fix a problem an older model had. The model moved. Your instructions should too: - It reads Anthropic's live docs for your model. - It checks your CLAUDE.md, skills and hooks. - It tells you every line: delete, keep, rewrite. - It hands you the implementation plan. You just have to read and approve it. Grab the skill here → https://lnkd.in/eM89acTy Repost ♻️ to help someone in your network. P.S. When did you last update Claude?

The (Death of) The Unrecorded Moment Item No. 22 “We press record and then we view our own lives as directors and actors, but, perhaps, not often enough as ourselves.” What is the unrecorded moment? We’re not even sure what an unrecorded moment is! That’s because for the most part in a social media driven world, we’re incentivized to share pretty much everything all the time with everyone. So it's hard to remember when we didn’t simply just record all our moments. It's hard to remember an unrecorded moment. We used to remember. With our brains. Now we press record and live our own lives as directors and actors. We perform for an audience instead of living for ourselves. Our stumbles, our goofy forgettable moments meant for no one but us, as we learn to live our fallible lives in a fallible world, are disappearing. We forget what its like not to perform. Isn't it in the unrecorded, the unreplayable, the private moment, where we stumble and learn, that we find poetry in living? The poetry of being. Being #human, that is? What if we didn’t record all our moments? Would we pay closer attention to our own lives? To the lives of those we love? Is (y)our memory full (or is that just (y)our digital storage that’s full?) Do you, do we, remember when…? (See (Future) Catalog entries for: Forgetting, Privacy, Boredom.) Learn more about the project at the following links: 🔗🔗 https://lnkd.in/dW4aY7Tj

A few weeks ago, I first tested Claude Fable 5 with codeveloping a series of analyses with TidyTuesday datasets, then a series autonomously. The results were impressive, especially considering visual analysis and ‘judgement’ is something LLMs have historically struggled with. Given Opus 5 seems to achieve parity with Fable 5 over many benchmarks, I decided to update the repo and site with a session of six self-selected and autonomously analysed datasets from Opus 5. They’re different datasets, so there couldn’t be peeking, but hopefully help show the contrast in both capabilities and style. (Link below) Maybe the main difference seems to be that Opus 5 tends to work harder, though not always smarter. You can see this, for instance, in how much more wordy Opus 5 pieces were than Fable 5 pieces, pretty much no overlap in the distribution. Arguably, Opus 5’s visualisations and analyses, though very competent, are less creative than some of Fable’s. (As perhaps the most impressive example of this, look at Fable’s visualisations and analyses of digits in pi, especially its mapping of digits to polar space to demonstrate lack of bias.) Though it might not be the most business friendly comparison (at least no business I’ve worked at!), a useful heuristic for thinking through the different approaches is: Fable 5: “Guru Claude” Opus 5: “Cocaine Claude” Both are clearly very capable. And for now both seem useful complements to each other. Just be aware, when you’re engaging either model, of the kind of ‘personality’ you’ll be dealing with!

A benchmark can reward a successful finish while ignoring review time, reversals, and mental overhead. That’s my problem with the current Opus 5 conversation. Anthropic markets Claude Opus 5 as close to Claude Fable 5 at half the price. In mixed daily use, I don’t get that feeling. It also doesn’t feel dramatically different from Opus 4.8. Some explanations are more verbose without becoming more useful. The output creates a second job: editing the model before I can use it. In March, METR reported that roughly half of SWE-bench Verified PRs that passed automated tests would not have been merged by maintainers. METR isn’t saying benchmarks are useless. It’s showing why a score is one piece of evidence, not a proxy for production usefulness. My scorecard needs to include time to a usable result, cleanup burden, reversals, cost, latency, and run-to-run variance. This is an early mixed-use impression, not a controlled head-to-head. My Opus 5 verdict: I’m not a fan yet. It may be better on paper. It hasn’t earned a better place in my workflow. What does your eval measure that the leaderboard misses?

“The Age of Uncompressed Thought Sharing: Chapter 10” A clear look at how base text and preserved thought structures relate, revealing how passages connect to recorded reasoning over time. From a Medium note: https://ift.tt/0un8RXk

Anthropic deleted 80% of Claude's prompt. It got smarter. Anthropic does this every model release. They delete the entire system prompt, then add it back one line at a time. Claude Code's creator explained which 80% went. Most of those lines were patches for things older models kept getting wrong. Opus 5 handles them on its own. Their own finding: the model is a little more intelligent without them. His advice: Every 6 months, delete your CLAUDE /md, your skills and agents system prompt.

Does peer review need evolution or revolution? - a question asked by Dmitry Kochetkov in the Scholarly Kitchen. https://lnkd.in/ekZ4u4vy I do like the Trust Markers model quoted at the end of the article with a "four-tier trust ladder with optional modifiers".

Opus 6 probably isn't close. There's a rumor going around that it ships in the next few weeks. Most capable model yet, six times cheaper, 2M context window. I went looking for something behind it and came up empty. No Anthropic mention. No model IDs. Nobody with a track record. It traces back to one post from today and copies of that post. So here's what I'd actually expect. Not in the next month. I'd put that around 10%. Inside three months is closer to one in three, and I think that's the honest range. I wouldn't expect those specs to survive either. The 2M context and the six times cheaper come from the same unverified post. Nobody else has confirmed a number. I wouldn't bet on the name either. Anthropic went 4.6, 4.7, 4.8, then Opus 5, all between February and July. Four releases in six months. Point bumps, not whole numbers. What I would expect is another model inside the quarter, because that's been the rhythm all year. But my point is this. None of it changes what's actually slowing you down. If your data is a mess and nobody owns the workflow, a cheaper model just runs your problems faster.

"The Age of Uncompressed Thought Sharing: Chapter 9" Constructing a thought structure through dialogue tracing how records become a provisional understanding with confirmations over time. From a Medium note: https://ift.tt/pBR3m9I