From subjective experience, Sonnet now feels just as slow as Opus. Maybe not in raw tokens, but in outcomes. I’d love to see model labs start measuring this more.

I'm afraid I'm going to be reposting this every day until the far side of the singularity. Work on the things you want to see more of! Work on flourishing! Work on things that are freeing!

This is so true and I'm happy it turned out this way. Humans + AI both benefit. The reason why it's MORE important now is that LLMs forget everything while humans learn from trial and error. We need to compress all that learning into documentation.

This is so awesome. The fact that I can turn my expert coding agent into a really good teacher by dropping some files in a folder is just wild.

Ooh, I like this direction. I’ve been experimenting with keeping thinking disabled for straightforward tasks, mainly for speed.

I really wish model labs would start posting benchmarks on speed and token efficiency. Can’t we just measure the time/tokens to complete existing benchmarks? Especially interested to see if smarter models can be faster on certain tasks by needing fewer tool calls.

Let’s make some virtuous software today

stranger: what do you do for work? me, shamefully: computer

For all you digital gifters, I made a heart-shaped QR code generator 💝 Can also be used for links or secret love notes! 🤫💕 (only tested scanning on iphone) qheart.vercel.app

For all you digital gifters, I made a heart-shaped QR code generator 💝 Can also be used for links or secret love notes! 🤫💕 (only tested scanning on iphone) qheart.vercel.app

always so funny to me when people are like “how am I supposed to explain Gay to my kids?” my kid asks me way harder shit all the time. “why did they build the road this way? why are some people bad?” I don’t know. please ask me about Gay

Sitting in a cafe telling the computer what to do all day is a new level of work in public

I’ve been getting so many chores done while Claude is working 🧽

Agentic coding is starting to feel cost-bound again for me. I use API-based billing (bc work) and I’m no longer sure that what I build is necessarily worth the price tag. $20 for a feature is not much, but also not negligible. And I don’t like that I have to consider that before starting.

I think CLIs are winning precisely because they are *not* very open-ended. You’re fully constraining yourself to text input/output, and that’s great for LLMs.

Made a self-updating skill for scaffolding React projects. It runs a script so I get a fast and consistent setup every time, but uses the LLM to update the script based on shadcn docs. github.com/nimobeeren/s...

I've been having way more (good) interactions on here for the last few weeks! I was added to a starter pack but I feel like it can't just be that. Probably For You has something to do with it? 💖

Trying to explain nondeterminism in LLMs: it's like talking to clones of the same person but they're in a slightly different mood each conversation

Rambling into a recording device like it’s the 70s, except I’m not causing psychic damage to the person who is asked to transcribe it

The whole “prompt engineering” thing still has people believing you need to speak some arcane language to make good use of LLMs. In reality, you just need to be unambiguous, structured and complete. No tricks necessary.

Sonnet is definitely due for an upgrade! Opus is actually faster for me in most cases because it calls the right tools instead of messing around.

Yeah, like please invest that effort into turning your app state into local Markdown and JSON (or SQLite) files so I can use my good AI tools to work with it

The other big win with agents is that they let people do things that were completely out of reach for them before. So, not a productivity win, but a capabilities win.

While Claude is definitely useful for boring, large scale refactors, it’s most fun when it lets you do something you just couldn’t do before. For me that’s complex algorithmic optimization.

Moltbot is a pretty strong signal to me that software security is going to get a lot worse before it gets better. We're in for a wild ride.

I wish models didn't leave code comments like this. It only serves to highlight the current change, but has no value to readers of the code in the future. Haven't been able to prompt this out yet.

I'm blasting some ~lofi~ and having a beer, gotta maximize those vibes while Claude codes ✨

The default behavior when sending a new message in an AI chat app should be to scroll new message to the top and then STAY THERE. I don’t actually want my text to auto scroll and be unreadable until streaming stops. Sucks that this is hard to implement though!

Apparently it's still really hard to make a scroll container stick to the bottom while content is being added (i.e. every AI chat app ever). Have there been any developments to make this easier in web apps?

Subagents just landed in Cursor 2.4! I think these will be especially useful for tasks at scale, e.g. create 100 components/pages/test suites in parallel. They also ship some subagents which run automatically. This looks like a nice optimization to avoid context bloat.

Today’s models feel very strongly biased towards keeping existing code instead of rewriting. Almost like the sort of laziness from GPT-3.5 times. I think I need to work on my tests and make a habit of deleting large parts of implementation.

Even if LLMs can’t write “production-ready” code to your standards (and you are sure those standards are still serving you), they let you iterate and explore so much faster. Even the best programmer doesn’t always know what users want. Prototypes help you get that feedback.

Is there a future where TypeScript types could ever be read at runtime? It’s the one thing I miss coming back from Python

Knowing about a large collection of quality software to be shamelessly inspired by is such a superpower. From simple UX patterns to solving entire categories of problems, you just have to know what’s out there.

Claude just complimented my incredible prompting skills 😌

I know we're in the "horseless carriage" stage of augmented development & it's too soon to know what the "real" name is of this new activity, but... ...the genie suggested "wishcraft" & I don't hate it.

my competitive advantage is that i'm having fun

This is an alternative hypothesis: External Forceful Depantsing. I will evaluate it. Evidence against this hypothesis includes the subject's relaxed posture and lack of visible injuries. It is improbable that an assailant would selectively remove only the pants. My original analysis stands.

Claude + GitHub CLI is so useful for finding out the actual status of a long issue thread. Just had it navigate through 3 connected issues to find out I can remove a workaround in my code after upgrading some dependencies 🎯

Building the NYT Strands solver has been such a fun project since I started nearly 4 months ago. It was addicting to see the progress with each new idea. Experiencing the advancement of LLMs/coding agents was awesome too. I simply could not have done this without those tools. Here's how it works 🧵

Calling a side project “done” is a hard but liberating act (I reserve the right to call it “undone” at any future moment)

Finally published strands-solver 1.0.0 🚀 It solves NY Times Strands puzzles using a backtracking algorithm and word embeddings. Try it with `uvx strands-solver solve today` (free tier GEMINI_API_KEY optional) github.com/nimobeeren/s...

I personally haven't run into the issue of Claude bailing out before finishing its work (at least not recently), but the principles and great examples make this valuable to anyone using coding agents. I'm glad someone put this together and did it so well 🧡

We love predictable optimizations. Easy puzzles got a bit slower, hard puzzles got a lot faster 🚀 This was the result of integrating strand crossing checks into the search, rather than post-filtering them.

Genuine question for Claude Code users: what makes you prefer it over more visual/IDE alternatives like Cursor or GH Copilot? Is it overall quality of results, ergonomics, specific features, pricing?

I was experimenting with some battle-tested SAT solvers but found them slower than my Claude-rolled Python (!) implementation. Surely this is a skill issue? Unless my optimizations are too specific for these solvers to find. I figured offloading to optimized C++ code would be faster, but no!

This might become a problem in a world of hyper-personalized software. If you’re the only one using an app, you won’t know it’s broken until it happens to *you*. There is safety in using software that millions of others use.

I was messing around with Opus to generate an evals app for voice AI apps. I asked it to generate some mock audio files and it kinda bops 💿

I want my coding agent to run in the cloud for security/sandboxing but also want to interact with the code and run it manually. Maybe it’s time to do the full cloud IDE thing?