Claude Code Learns to Design, Cursor Hosts Your Code & Everyone Wants to Replace Git
Claude Code & Anthropic Updates
/design lands in Claude Code as a research preview
Nate Parrott announced it (2,096 likes, 75 replies, 341.1k views):
"Today we're releasing an early preview of the /design command in Claude Code! from CC Desktop or CLI, try something like '/design a few options for {feature}' before you build — pick your fave artboard, edit it and implement."
The ClaudeDevs framing is that this brings Claude Design's artboard workflow into the CLI, built on artifacts. Boris Cherny passed it along with a plain "Let us know what you think!", and trq212 was blunter about it: "go into CC and type /design
The useful detail is in the replies. Asked whether the artboard survives into the build step or you end up re-describing the layout in a prompt, Parrott confirmed:
"carries over! (Remember to press Save for now; we'll get rid of the need to press it soon). Ofc it's up to the model to adhere to the design but they're getting better and better at this"
The best critique came from Rasmus Andersson (47 likes):
"Would be neat if this just opened Figma! I feel like any design made here would be locked up inside there, with a limited life and thus limited usefulness."
It's a research preview and it shows — multiple people report /design still printing Usage: /design consent | /design revoke after updating, and artboards rendering weirdly in the CLI. Also unanswered so far: whether it works on CC web, and what the recommended flow is when you're designing a new screen inside an existing large project rather than starting from zero.
The boring perf fix that moved p99
ClaudeDevs (3,872 likes, 151 replies, 270.1k views):
"Perf win of the day: Claude Code CLI now uses 2x less CPU at p99. Bun's garbage collector was running on a fixed timer, so it would kick in mid-turn and steal CPU right when Claude Code was busiest. Now it waits until the process is idle."
Ziang Gao drew the generalizable lesson for anyone building a harness: "run GC when it is least disruptive, such as while the process is idle or waiting for the model response, rather than simply every N seconds." Boris Cherny added: "Small quality of life improvements like this add up. More on the way."
The reply section, meanwhile, was almost entirely people asking for a usage-limit reset so they could test the fix — plus a request to drop the 50% Fable 5 cap for Max subscribers ("I think people are exhausted").
Source Control & Code Hosting
Cursor's Origin is live — during a GitHub outage
Cursor (20,570 likes, 1,285 replies, 19.2M views):
"Origin, our code hosting platform, is now live. It's fast, easy to use, and deeply integrated with Cursor. Get started by syncing your repos from GitHub."
swyx's entire commentary was "is live!". Armin Ronacher was less easily impressed (post):
"People are saying Cursor has great timing to launch Origin but given how often GitHub is down these days I don't find the timing to be too impressive :P"
He spent the day mining the outage for material — "Didn't we all ask for a bit of friction and back pressure? GitHub is providing it!" — and used the moment to point at his tangled account (tangled.org/mitsuhiko.at).
The replies are the real signal on whether Origin lands: it's asking for payment before you can move repos over, which people compare unfavorably to GitHub being free (1, 2); $200 Ultra subscribers report not having access; and the recurring question is simply whether Cursor integration is enough to move anyone (Kien Pham: "We have GitHub, then GitLab...and now Origin"). Several people concede the outage helps: "After what happened with GitHub today, I suspect more people will give this a shot".
Zed's DeltaDB: source control where the conversation is part of the commit
Theo has been arguing that "GitHub is dying and git is not the right primitive," and explicitly asked someone else to go fix source control while he fixes clouds for agents. Zed took the invitation (1,146 likes, 126 replies, 183.3k views):
"We've got you @theo. (you're invited). We spent our psychosis building Delta on a different primitive: every edit recorded as an ordered delta, together with the conversation that produced it."
Early access at zed.dev/deltadb. Theo's response was three eyeball emoji (129 likes) and, separately, "This is a really good tweet". The thread itself is mostly Zed staff hand-issuing invites in the replies, so there's no design detail beyond the one-liner yet — but "the diff and the prompt that produced it are the same object" is the most concrete answer anyone has given to the post-git question, and it's a natural fit for a world where most edits are made by an agent.
Agentic Coding & Agent Harnesses
"I'm tired of terminals"
Theo posted a video under exactly that title (505 likes, 62 replies, 40.3k views) — the pitch being T3 Code, his open-source browser-based harness, over TUIs. He's been leaning hard on the remote-first angle all day: T3 Code on a bad network is "legit trippy... You forget the connection is terrible until you alt-tab back to your browser and everything is slow again," prompted by someone reporting it working fine on Delta's non-Starlink wifi.
He also conceded the field to Codex on remote — "Codex is far ahead of the other labs with their remote stuff" — while arguing T3 Code goes further on project creation, multi-PC management and remote configuration. That was in response to Federico Viticci's notable claim (quoted in Theo's post) that Codex Remote for iOS is the only remote-control product with a full start-from-iOS flow (Claude and Cursor make you start from a Mac), near-desktop parity, and a voice mode that can dispatch to individual threads, load desktop context and plugins, and reopen threads on screen. Corey Quinn, no horse in the race, called T3 Code "freaking transformational".
The replies push back usefully. Dima: "what's the alternative you actually want, not another wrapper that reopens a terminal underneath". Josh McK thinks the missing piece is "a more semantic terminal protocol that avoids making cells the unit of interaction, rather than trying to bake apps on top of the 50 years of spec tower of babel." And Madeline Fox gave the most honest version of the argument:
"Moving away from living in the terminal has honestly been one of the bigger mindset shifts I've had to make with agentic coding... my reliance on my terminal was increasingly making me the bottleneck. I still use it and use it most days, but far less."
Daniil sketched the likely end state: a split between hands-on steering work (cognitively heavy, burns you out if it's all you do) and fully offloaded background work you check via task trackers and PR review, time-boxed instead of context-switched.
"One dev, many terminals" as an awkward interregnum
Matt Pocock quoted Jared Palmer — who hasn't set up his laptop for local development since joining Cognition and works with Devin in Slack and the webapp instead — with:
"It me. We'll look back on 'one dev, many terminals' as an awkward interregnum"
Best exchange in the thread: told this only works "until your provider silently reroutes you to a weaker model," Matt replied "What if your provider is you". One reply pushed back on the whole ADE-wrapper category — forking a webapp harness like T3 Code and molding it to your needs beats adopting Herdr, since "there doesn't seem to be anything I can do on herdr that I can't on webapp" — and another thought "awkward interregnum" was generous: "more like the dev setup dark ages."
Meanwhile Gergely Orosz summarized the era in one line: "If you've not built your own AI coding harness by now, are you even a serious tech company?"
Podcast: State of Agentic Coding #9
Armin Ronacher and Ben Vinegar are back for episode 9: how agents are getting more autonomous, "the subsidies will continue until morale improves," Herdr and VC-funded terminal apps, and vibe coding from your car.
Codex & OpenAI
Tibo asks what OpenAI is obviously missing — 4,728 replies later
Tibo Sottiaux (3,731 likes, 4,728 replies, 510.7k views):
"What is an obvious thing that we should do with Codex, API or our models that we should just do but haven't yet? What is 100% within reach, but we just seem to be missing?"
Filtering out the reset-my-limit jokes (there are thousands), the substantive asks cluster into a few themes:
- Unsupervised reliability above everything. ephe explorer: "please make sure that all models that are released are good at running unsupervised. every time i need to steer or check if the model is doing okay, or not doing needless things is time wasted."
- A "slow mode." zsilver92 (487 likes): the inverse of fast mode — fire a prompt off before bed, don't care whether it takes a minute or an hour, and charge a fraction of normal usage the way fast mode charges 1.5x. Easily the best new product idea in the thread.
- Kill the pickers. Colin Devroe (271 likes): "I should never have to pick a mode or a model. ChatGPT/Codex should just know what I am trying to do and choose everything for me to be as efficient as possible." (Compare Grok Bot's no-model-picker stance from yesterday.)
- Worktrees from
origin, not stale local main. Dafydd: new-worktree-from-main usually branches off an outdated local copy. - Tell me what the update button updated. 1kuaikuaide (85 likes): "After I click that blue update button, let us know what has been updated. Most obvious one."
- Remote is painful. Chris Laupama (241 likes) describes the app's remote system as a mission to reach, slow to connect, and constantly disconnecting — worth reading against Viticci's glowing iOS review above. The ChatGPT desktop app stuttering on high-end hardware (716 likes) got the second-most traction of any reply.
Tibo also posted a scorecard for Codex — "Almost 100% reliable / Occasional resets / Open-source / (will have Astra)" — which Theo amplified.
Video: agents beyond code authoring
Databricks published a conversation with OpenAI's Peter Steinberger and Tibo Sottiaux, hosted by Databricks co-founder Patrick Wendell, on what happens as agents move past writing code into observability, operational work, and long-running responsibilities across software and data systems. Databricks' pull quote: "With coding agents, you're putting this jetpack on and going 10,000 miles an hour, doing things you were never able to do before." Full webinar.
Skills, Standards & the Human Loop
The ladder for eliminating agent corrections
lauren (React Compiler core contributor, now at Cursor) posted the most reusable thing of the last two days (751 likes, 49 replies, 36.4k views):
"every time you intervene and correct your agent, you should think about how to eliminate it entirely. in order of value:
- categorically eliminate the problem through better architecture or choice of data structures
- turn it into a lint rule or test so CI catches it
- turn it into a skill or rule
- have humans review the code to catch it (ngmi)"
Note the ordering: skills and rules are third, below architecture and CI. Asked whether a memory layer with similarity recall would help instead, lauren's answer was the line of the day:
"your codebase is a form of memory too"
Theo pulled out a related quote from her: Cursor's internal framework "actually bans useEffect outright and only exposes it indirectly through framework-provided hooks. Highly recommend setting something like that up for yourself!" — level 1 of the ladder, applied to React. Same thread also confirms Cursor has largely completed its Solid → React migration, converted all its scss and Tailwind to StyleX, and that the new agents window is 99% React; the motivation was perf and maintainability, since "it is actually quite easy to create perf footguns for yourself in signal based approaches when reactivity causes a large amount of accidental fan out in large and complex apps."
Matt Pocock's course, and Uncle Bob on fundamentals
Matt launched his AI Coding Crash Course ($199 this week) with a framing that's more diagnostic than salesy — can you get it to build the thing you wanted, stop it making the same mistakes every time, schedule huge chunks of work, "or are you hanging on by your fingernails?" Contents: the smart zone/dumb zone, compaction vs clearing vs handoff, steering with AGENTS.md and skills, and building specs and tickets to tackle huge projects. Link. An early student reported cutting starting context 40% to 16k tokens off the fundamentals section alone.
The companion argument: "If your code sucks, your agent will too. So, software fundamentals matter more than ever." — which is why he's interviewing Robert C. Martin live on YouTube this Wednesday.
He also added a /handoff command to Sandman for spawning new Discord threads.
Adjacent, from Armin Ronacher, on why conventions calcify: "People come up with elaborate explanations of why things are like they are, but then they are just because of some weird decision" — his example is the C habit of cargo-culting common prefixes on struct fields, which only exists because those fields used to be globals. Worth keeping in mind before you write it into a standards file.
Models, Local Inference & Creative Work
Qwen 3.8 27B, day two
Simon Willison's review is still the best writeup ("I can't remember the last time I've had this much fun playing with a local model that runs on my own computers"), and he added a nice coda: he pointed Pi at Qwen 3.8 27B and had it build a script to convert its own .jsonl transcripts to Markdown — then used that tool to share the transcript of it being built.
LLMJunky added the hardware angle:
"the fact that a 27B model finds itself squarely between Opus 4.6 and Opus 4.7 Max Reasoning on Artificial Analysis is insane... On a single 5090 or RTX 6000 Pro, you can get around 200 TPS... As these models improve, consumer hardware will become increasingly difficult to buy. There's not going to be a cheaper time to get hardware than right now."
His practical tip matches Simon's overthinking complaint: use Medium reasoning, for everything.
Coding models are eating diffusion's creative lunch
trq212 updated his priors:
"all of the recent proc gen art, video editing and 3d game demos recently have made me update towards LLM coding models being better at a lot of creative work than diffusion models"
His reason is the one that matters for tooling: "the nice thing about code is that it is easier to edit and nudge in the directions you want and export to work with existing tools." A live example the same day — LLMJunky is building an agent-integrated video tool (mockups, custom backgrounds, text, audio, 3D models, transitions, multi-clip) precisely because you can tweak the agent's output in a UI instead of accepting the slop: "GPT 5.6 Sol made me this video in seconds."
Other Bits
- Kill My SaaS closed at 69 submissions. swyx's $10,000 solo-organized hackathon wrapped; there's now a visual field-notes index of all 69 competitor-built conference products with their AI tools, tech and build economics. Conor Bronsdon's verdict: "Confirmed: you can vibe code your way to replacing a SaaS" — his entry, an open-source speaker/CFP manager, is at callboardhq.com (repo).
- Two AI Engineer videos worth the time. Ronak Malde of Trajectory on scaling continual learning — why GRPO isn't enough, why they had to go on-policy, and the mess that follows when you do; swyx rates their taste highly. And a context engineering workshop on compaction, memory and cost, built around an open-source AI tutor: the two root problems (finite window, stateless model) and the full compaction toolkit — truncation, trimming, tool-result clearing, summarization, offloading to files — with when each actually helps.
- "Ideas are the new bottleneck." A summary of Latent Space's interview with OpenAI's Akshay Nathan (productivity pillar — ChatGPT Work and Codex): once anyone can build, the scarce inputs are ideas and taste, and the old proxies for productivity stop telling you anything. The warning for managers is that AI makes activity nearly free while progress still costs the same discipline — "the motion trap."
- YouTube is changing what counts as a view (from 8/24, the moment playback begins, no minimum watch time). Theo doesn't like it — "YouTube view counts are the only view counts that mean anything" — but gets why: he lost a brand deal to someone half his price "getting 70,000 views" on X, where a view counts as an API response with your post ID.
- Slop of the day. LLMJunky: "30 agents and 22,000 lines of slop — all i asked was to center a div". Adjacent, from Jarrod Watts: "POV: You try Codex Ultra mode (143 agents have spawned and only written tests so far)".
- Framework authors got AI-pilled early. trq212 noticed that the creators of Django (Simon Willison), Flask (Armin Ronacher) and Rails (DHH) were all early adopters — "it says a lot."
Notes on sourcing: nitter.net intermittently returned "user not found" HTML for several accounts and needed retries; xcancel.com now requires RSS-reader whitelisting and is no longer usable as a fallback. Karpathy had no posts in the window.