Desmond Choy

@Teknium @josephdviviano @redtachyon Can Opus still run with Hermes agent? I was under the impression that all OpenClaw and OpenClaw denominations wouldn't work with Claude Max anymore.

@josephdviviano @redtachyon pretty default. With opus ofc

Should I actually try hermes agent? How is it different from openclaw (which to me seemed largely useless)?

🚇🇸🇬 we made a SG MRT station guessing game - 3 tries to match the names to 5 stations on the MRT map - hints provided for line + direction - speedrun mode so you can brag to friends - island-wide tour if you're good enough try it here: 👉 mrt.eiti.dev 🚇

@socialwithaayan Why is this "Breaking"? @karpathy published the gist precisely so everyone could build and customize their own wiki. You do so by pasting the entire gist into a LLM, plan mode perhaps, and just discuss and iterate. Works with any language and OS because the spec is in English.

🚨 BREAKING: Someone just built the exact tool Andrej Karpathy said someone should build. 48 hours after Karpathy posted his LLM Knowledge Bases workflow, this showed up on GitHub. It's called Graphify. One command. Any folder. Full knowledge graph. Point it at any folder. Run /graphify inside Claude Code. Walk away. Here is what comes out the other side: -> A navigable knowledge graph of everything in that folder -> An Obsidian vault with backlinked articles -> A wiki that starts at index. md and maps every concept cluster -> Plain English Q&A over your entire codebase or research folder You can ask it things like: "What calls this function?" "What connects these two concepts?" "What are the most important nodes in this project?" No vector database. No setup. No config files. The token efficiency number is what got me: 71.5x fewer tokens per query compared to reading raw files. That is not a small improvement. That is a completely different paradigm for how AI agents reason over large codebases. What it supports: -> Code in 13 programming languages -> PDFs -> Images via Claude Vision -> Markdown files Install in one line: pip install graphify && graphify install Then type /graphify in Claude Code and point it at anything. Karpathy asked. Someone delivered in 48 hours. That is the pace of 2026. Open Source. Free.

@ivanleomk @pbakaus This looks sick! Amazing for someone as design-illiterate as myself! Thanks!

impeccable.style by @pbakaus is the best skill i've used this year. Can't believe it's free

@felixrieseberg @maxhodak_ What's also cognitively jarring is that projects in cowork are different from projects in regular chat. When I have an idea and I want to resume where I left off in projects, but I see empty projects, I start to think what happened, and then I realize I'm in cowork...

You know, we think about this literally 24/7 and I suspect that we’ll figure this out eventually For now, they offer slightly different flavors to users - chat for easy conversations, Cowork when you want to safely work on something, Code if you’re a developer and want to code with Claude

why are claude, cowork, and code three different uis? why is this not all just claude?

@ryancarson Feels like this was already coming with the Dispatch announcement? In any case, the crowd has a short-term memory. Anthropic is prob betting Mythos + Dispatch will bring back the lost sheep.

My whole timeline is everyone talking about the Claude Max plan not being able to be used in OpenClaw. Will be interesting to see how this shakes out

@doodlestein Ah ok it works now for me. Strange.

@Norest Works for me?

This skill is no joke. You just point it at your project and trigger it and come back in an hour and it has usually made some massive performance improvement in an isomorphic way. Then just rinse and repeat over and over again. It basically applies every leetcode and IOI trick.

@doodlestein @doodlestein jeffreysprompts.com is down btw.

Another cool skill in action as I'm onboarding the vastly improved fuzz testing skill I made this morning (see screenshot from within Claude Code below). For every new skill I add to the site, I use my /interactive-visualization-creator skill to make a slick looking visual for it on the website. But that skill is so in-depth and huge that the agents never "saturate" what it can do in a single application. So, with urging from @joshmh (see x.com/joshmh/status/… ), I created a "meta skill" called /repeatedly-apply-skill that lets me reliably called the visualization skill 5 times in a row, with the first pass creating the visualization and the other passes iteratively polishing and improving it. And all this is called automatically from within my  /onboarding-new-skills-to-jeffreys-skills-md (not included on the site because it's so specialized). Skills calling meta-skills that apply other skills. Skills all the way down!

I put a lot of heart into my technical writing, I hope it's useful to you all. 📌 Here's a pinned thread of everything I've written. (much of this will be posted on the Claude blog soon as well)

@jxnlco Codex agents jamming with each other, with memory and auto compaction. Video in link! linkedin.com/posts/choydesm…

Send me your non coding Codex demos! - 30-50 second video music production, game development, 3d, automations, anything. anything good we'll shoot over some credits so you can go deeper!

@karpathy @paulg Why make it last exactly 5 minutes though? Context window filling up?

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. github.com/karpathy/autor… Part code, part sci-fi, and a pinch of psychosis :)

Day 3. Kids begging to do homework. Turns out if you wrap fractions in dragon quests, they don't notice it's learning. odysseyos.polsia.app

@HamelHusain @petergyang @trq212 Sorry, "request_user_input"

@HamelHusain @petergyang @trq212 In Codex, the same feature is called "ask_user_request" tool

The thing in Claude where it dynamically presents you with a mini interface to choose between options is amazing Just did a calendar thing and it gave me a picker on the web In claude code I got a picker for three different design options

@AlexReibman @Steve_Yegge Yup. Using it across 4 projects. It's fantastic and very easy to use.

@Norest @Steve_Yegge Ok this one is a bit… contentious. Have you tried/gotten value out of it?

What are the best memory tools for coding agents?

@koltregaskes My waiting times are gone with this: x.com/i/status/20258…

Google Antigravity is definitely my preferred IDE as it tests code in a browser but that's not to say it's pain free. I've turned on all settings to 'always proceed', yet it still asks me to run or reject everything. It also gets stuck on "Waiting" all the time and then needs a restart. If AI Studio is going to be 'powered by' Antigravity then Google needs to fix these first.

@pankajkumar_dev I added this to my GEMINI.md, and it seemed to help. The `command_status` tool may lag behind actual command completion. Use a maximum 3 s wait (`WaitDurationSeconds: 3`); if the command still shows as running, proceed rather than polling repeatedly.

Antigravity Problems: - Eats a lot of RAM and feels very heavy. - Quotas finish too fast now. Need API key option, pay-as-you-go, or a plan between Pro and Ultra. - Agent terminated due to error happens too frequently. - Models don't perform as well compared to Claude Code or Codex, especially with long context. - If history gets large, it sometimes doesn't even open. Have to wait or terminate it. - UI/UX needs improvement no clear view of remaining context or token usage, and no compact context feature. - Browser feature is very slow. Navigation and screenshots take too much time. - Overall performance has become slow, especially with Gemini models. Please google team fix these issues and refine it.

@unnamed1tw @GeoffreyHuntley @cursor_ai AEE = Artificial Enhanced Environment?

@Norest @GeoffreyHuntley @cursor_ai Cursor isn’t a traditional IDE but rather AIDE or closer to AEE. CLI which we also have, is not going to be long term target for dev as more use Cursor Desktop and Cursor Cloud now.

tbh @cursor_ai team is really cooking, really cooking. i'm not saying that as someone who has free tokens thus biased. i'm saying it's becoming super high taste wait and see