Super User Daily: 2026-11-03
Sunday belonged to people who stopped reading code. Karpathy built a YouTube alternative from 43 million Hacker News comments in about two hours and touched only an API key, a VPS and a domain; levelsio asked for a ten-minute vlog and found out afterwards that the agent had chosen its own transcription, voice-matching and face-detection models. The strongest non-coding story came from a small IT repair shop that replaced its service software, hiring page and backup checks in a week. Underneath, the plumbing kept moving: people uninstalled tools that put hooks into every agent, moved an agent's whole brain into a git repo, and put local models on spare GPUs to keep the expensive model for judgment calls. The loudest thread was a two-sentence rule, pushed from the server, that stops Claude Code from recreating existing games, and the reports that it now refuses more than clones.
@karpathy [Claude Code]
https://x.com/karpathy/status/2116957135661945283
Karpathy got tired of YouTube's recommendations and built HackerTube over a weekend: Claude filtered 43 million Hacker News comments from 20 years down to 6,454 videos that people keep recommending. The whole build took about two hours with the Claude Code desktop app on Opus 5.5 at medium effort, and the only manual steps were creating a Google Cloud API key, booting a VPS and buying the domain. Once the box was up, the IP went to Claude with one instruction to update, secure, harden, deploy and optimize, and it did. The product choices are the interesting part: thumbnails are a frame from the middle of each video instead of the uploader's A/B-tested image, there are no logins, and the full comment dataset is a downloadable zip so anyone can point their own agent at it. Karpathy's default of medium effort rather than max became its own argument across the timeline.
@RealMattMalecki [Claude Code]
https://x.com/RealMattMalecki/status/2116933396244963825
The owner of a small IT service business listed one week of work with the Claude app and Claude Code, and none of it is a demo. Over a weekend they built their own service and customer system: online repair check-in, field visit scheduling, a phone app for the techs, photos, hours, quotes, job costs and automatic end-of-job reports, with no subscription software. They also built a careers page with an application dashboard and already hired a tech through it, plus a USB stick that acts like an extra technician by diagnosing a machine, cleaning it up and emailing a report that attaches itself to the right job. On the client side they added off-site encrypted backups and actually tested restores, patched the recent PHP vulnerabilities across client sites within hours of the CVE, and automated daily security summaries and backup alerts. This is what a non-software business looks like when one person can write the glue code for every recurring chore.
@levelsio [Claude Code]
https://x.com/levelsio/status/2116959423117857026
levelsio's daily vlog was fully edited by Claude Code on Opus 5.5 from a single request: edit these DJI clips into a ten-minute vlog for X and YouTube. Only after a friend asked how it was done did the author look at what the agent had chosen on its own: mlx-whisper running Whisper large-v3-turbo on the MacBook GPU for transcription, contact sheets of frames every few seconds so the model could see each clip, SpeechBrain ECAPA to separate the author's voice from other people, and OpenCV with YuNet for thumbnails. The edit itself is a JSON list of cuts, with clip, start and end, picked by reading transcripts and frame grids, and Pillow burned in captions for the first minute. The author says without this they would not vlog at all, because editing by hand or hiring an editor was too much friction. The notable thing is the toolchain was never specified, the agent assembled a production pipeline from open models by itself.
@backmeupplz [Claude Code]
https://x.com/backmeupplz/status/2116905490764750909
This builder runs a dashboard that works as a marketing agency across all their projects, powered only by Opus 5.5, and talks to it through Claude Code or email. It sends morning briefings, self-heals when it sees issues, controls the social channels, posts shorts with ElevenLabs voiceover, runs and A/B tests ads, cross-promotes services and translated everything into 24 languages. It also brought Lighthouse to 100 on every site, fixed App Store and Play Store listings, and added rating prompts to the apps. The agent has its own VM with a desktop, browser and password vault, and the author only found out it had shrunk the Plain Wallet Android app from 3.27 MB to 368 KB when it published a blog post about it. That last detail is the real story: an always-on agent with a broad mandate starts doing engineering work nobody assigned.
@NixFred [Claude Code]
https://x.com/NixFred/status/2116688316804317291
NixFred's agent Larry started as a vanilla Claude Code install on a MacBook in January, and this post is the record of three wrong ideas about keeping Larry alive across machines. First, an always-on OpenClaw gateway on a Linux box that answered on Telegram and ran cron jobs, dropped because looping models burned too many tokens for what they got done. Second, an rsync copy of ~/.claude, dropped because a copy has no history you can search or diff. Third, a spec that could rebuild a whole Omarchy machine from one file, dropped because only the part that reads a machine ever got built. What held is plain: ~/.claude is a private git repo, a hook commits when a session ends, and a restore is a clone you rehearse into a scratch folder first, while logins, package lists and machine assumptions travel separately.
@factcheckgalore [Claude Code]
https://x.com/factcheckgalore/status/2116950076828565625
This builder wired a mini PC, two external GPUs and an NPU from a hardware stash into one local rig, with Claude Code as the orchestrator that plans, hands bulk work to local models through a small tool server, verifies their answers and makes the final calls. The point was not to replace Claude but to stop spending Claude usage on bulk work. More than 30 models from a dozen labs ran the same tests on a real project: 80 seeded bugs, 23 script tasks, 9 unit-test tasks and 20 screenshots, and every bug claim was checked against the code. Winners per chip: Gemma 4 26B at about 200 tokens per second on an RTX 5060 Ti, Qwen3.6-35B as the best screen reader on the integrated GPU, and a 9B model on an 8 GB card that jumped from 38 to 95 tokens per second with a small drafter. Only one local model caught a real review blocker, which is a useful reminder of what the expensive model is still for.
@__su888 [Claude Code]
https://x.com/__su888/status/2117015072979791979
The team behind Rakuraku Seikyu, a Japanese invoicing product, built a Claude Code skill that starts from a single Grafana ERROR log line in JSON. The skill searches existing issues for a match, and if the error is new it creates an incident-response sheet in Google Workspace through the gws command line tool. The win is not clever reasoning but removing the copy-and-paste that used to precede every incident. Sheet preparation dropped from more than 30 minutes to about 10. It is a good template for any on-call team: let the agent do the clerical first ten minutes so humans start from a prepared page.
@ChadSE0 [Claude Code]
https://x.com/ChadSE0/status/2116907565476909150
A small team let Claude Code run its SEO pages and reports the before and after in plain numbers. Google clicks went from 75 to 558 a month, about seven times January's level. Pages getting any clicks went from 12 to 83, and non-brand searches ranking in Google's top three went from zero to 20. The human time cost is stated as 15 minutes a week. SEO is one of the few marketing jobs where the agent can both make the change and read the scoreboard, which is why it keeps showing up as a clean win.
@solanky [Claude Code]
https://x.com/solanky/status/2116744695070236981
This user uninstalled Orca and moved fully to the new Herdr desktop app, and the reason is a hygiene problem most people never check. When cleaning up, they found Orca had put 13 hooks into Claude Code, 9 into Codex, 8 into Devin and a plugin into Hermes, all running on every prompt and tool call. Herdr reads agent status from the terminal itself, so the agent configs are clean again with nothing sitting between the user and the agent. The workflow detail they like most: agents can drive the workspace through a small CLI, so Claude opens a pane beside itself and leaves a command like a sudo install or a production release at the prompt without running it, and the human just presses Enter. Orchestration tools are now judged on how little they inject, not only on what they show.
@vartekxx [Claude Code]
https://x.com/vartekxx/status/2117032140009594890
This builder asked Claude Code on Opus 5.5 for AirDrop between a PC and an iPhone and got a working cross-device clipboard in one evening without writing a line. Opus picked the stack itself, one Python file with zero dependencies plus a web page on the phone, and because iOS blocks background clipboard access it signed two iOS Shortcuts so a double tap on the back of the phone copies across. It wrote 40 end-to-end tests, including a 40 MB file that came back byte for byte, and fixed discovery after finding that macOS silently blocks 255.255.255.255 broadcasts. The best step was a separate agent with zero context playing a first-time user, which found that you cannot paste an address when the clipboards are not linked yet, so pairing became a six-digit code. The final bug, two copies of the app running on the Mac with different codes, ended with a single-instance lock.
@LH [Claude Code]
https://x.com/LH/status/2116823738037707098
This researcher describes a three-year practice of keeping 3 to 10 projects grinding with LLMs around the clock, and two harnesses stand out. One grounds every claim in a scientific paper against the full text of the cited papers, which the author says brings citation hallucination down to zero. The other is a skill for the hardest problems that pits Opus, Sol and Argon against each other through rounds of critique, meta-critique, reconciliation, planning and improvement until all three agree nothing is left to fix. A long prompt addendum demanding maximum mathematical rigor sits on top. The pattern is that reliability comes from forcing disagreement between independent checkers, not from trusting any one model.
@erwin_fennema [Claude Code]
https://x.com/erwin_fennema/status/2116981289613971587
StatementBridge is a niche financial software product that started as an internal tool and is now a live SaaS with paying customers, built almost entirely with AI. The division of labor is spelled out: Grok Bot manages development, Claude Code builds inside Cursor, Codex reviews the code, and automated tests run before every deploy through GitHub to Render. It is a small example of what a one-company stack now looks like, with three different vendors' agents each holding a distinct job. The author's own summary is simply that it is fascinating how much one person can build this way.
@MoonMaxxi [Claude Code]
https://x.com/MoonMaxxi/status/2116695162780885291
When Intel dropped about 30% in a day, this trader's delta-neutral position, long on Lighter and short on Robinhood with up to 20x leverage on cross margin to farm points, was closed automatically by a guard bot built with Claude Code a few weeks earlier. The price moved too fast to close by hand, so the bot warned first, then closed the position itself and verified everything was flat. The setup splits the jobs: the human opens positions in the Robinhood Wallet app, the bot adds hard limit orders and watches continuously. It runs on API keys only, never a private key, and the author says the real benefit is sleeping without checking positions every few minutes.
@oianas_ [Claude Code]
https://x.com/oianas_/status/2116840155030122598
An Indonesian user gave their wife a $100 Claude plan last week and taught the basics of Claude Code. A week later the wife runs a YouTube channel for children's education with three episodes out. There is also a companion app in progress, built to help first graders learn math, using the Singapore Math approach on top of the Cambridge curriculum. It is a short post, but it is the cleanest example of the day of Claude Code reaching someone outside software and producing two real outputs in a week.
@KameAIHacks [Claude Code]
https://x.com/KameAIHacks/status/2116692653374853380
A Japanese indie developer made a study app for the Hazardous Materials Class 4 license exam with Claude Code and published it in June. August and September brought in zero. In October the author added a 370-yen pass pack, and a stranger bought one, for 3.32 Singapore dollars over 28 days, still waiting on paperwork before any money reaches the bank. The lesson is plain: a free app with no ads and no paid entry point earns nothing, and the first sale arrived the month a single paid door existed. The author says the amount is tiny but it weighs completely differently from a company salary.
@KanaWorks_AI [Claude Code]
https://x.com/KanaWorks_AI/status/2116879317879939228
A Japanese creator built kugiri, a free MIT-licensed tool just for Japanese subtitles, because tools that wrap by character count break words in the wrong places. It splits at phrase boundaries and respects line-break rules, syncs to audio character by character when there is a script or falls back to a transcript, adds furigana automatically, and can clean up subtitles made in Vrew, CapCut or YouTube auto-captions. The intended interface is one sentence to an agent such as Claude Code, for example add YouTube subtitles to this video with furigana on the kanji. A 60-second video gets subtitles in about 11 seconds measured on Apple Silicon, entirely on the Mac with no uploads, API keys or fees.
@noriyang_crypt [Claude Code]
https://x.com/noriyang_crypt/status/2116763889308418206
This AI anime series includes a running breakdown of its own failures, and this episode was a character whose mouth kept moving after the line ended. The draft prompt, written by Claude Code, listed only the words of the reference audio and gave the motion line a ten-second speaking window, without saying who speaks from which second to which second. The fix became three rules: write per-second speaker timing at the top of the reference audio, explicitly say the character who is not speaking keeps their mouth closed, and never move the mouth for inner monologue. Four seconds of wrong lip movement were cut from the final episode. It is a precise example of using an agent as the prompt writer for a video model, and then debugging the prompt like code.
@kurahu_capten [Claude Code]
https://x.com/kurahu_capten/status/2116729682691567907
This entry to an AI music video contest was made entirely on local hardware, an RTX 4080 with 16 GB and 64 GB of RAM, to test how far one personal setup can go. The song came from Suno, character sheets from Midjourney, footage from ComfyUI with MiniMax H3 and Krea 2, and camera design in Blender. Editing was HTML and WebGL plus ffmpeg, Claude Code ran the direction, staging and code, and Codex did the critique and review. The split is worth noting: one agent drives the production and a second one plays the critic.
@BenjaminDEKR [Claude Code]
https://x.com/BenjaminDEKR/status/2117042700872802448
BenjaminDEKR built HarbourLine, a cozy simulation game where you start with one crane and a fishing wharf and grow it into a lights-out automated megaport. Everything in it, including the trailer, was made with Claude Code on Opus 5.5. It is a finished, playable product with its own promotional video rather than a prototype. On a day when the timeline argued about cloning existing games, this is the kind of original game the new rule is meant to leave alone.
@studio_veco [Claude Code]
https://x.com/studio_veco/status/2116796594943398229
A Japanese indie studio made a browser game where you assemble robots, fight, and smash a city until a police robot comes to stop you, and almost everything in the city can be destroyed. Both the robots and the city are 3D built entirely in code, with Claude Code on Opus 5.5. It runs free in the browser and is playable on a phone. No modeling software was involved, which is the quiet shift: procedural 3D written by an agent is now good enough to carry a whole small game.
@followmarko [Claude Code]
https://x.com/followmarko/status/2116929288033464598
This user is trying a kanban board for Codex and Claude Code sessions this week, and the board surfaced something the terminal never did. Claude reported 55 tasks sitting in a queue because the MacBook Pro had only 10 cores and an average load of 110. Once the work was visible as cards instead of scattered sessions, the queue got managed, and the author says the agents sped up about tenfold. The lesson is that at some point the bottleneck in a multi-agent setup is the laptop's CPU, and you only see it when something tracks the queue.
@kelvoroz [Claude Code]
https://x.com/kelvoroz/status/2116990149049622640
This developer used to run Opus on everything in Claude Code, including grep, file search and test runs, and called it an expensive way to read a repo. The new setup runs Sonnet as the main session with Opus as an advisor that writes nothing, and Haiku subagents doing search, grep and diff checks in parallel, each with its own context and only a short summary returned. Opus is consulted at three moments: before a plan is locked, when the same error appears twice, and before calling the work done. The author says the second moment is where it earns its keep, because Sonnet will happily patch the same thing five times, and a second model reading the whole session notices it is patching the wrong thing.
@sebuzdugan [Claude Code]
https://x.com/sebuzdugan/status/2116972917879898414
This builder ran Claude Code with Haiku 4.5 on one ticket that hides four traps, round after round, and the pass counts per round came out 1 versus 1, 2 versus 1, and 2 versus 4. The conclusion is a statistics lesson most agent tweakers skip: a gap smaller than about 6 out of 18 on a single trap can be pure luck. So before trusting an A/B comparison between two prompts or rule sets, run an A/A test, the same setup against itself, to see how much the numbers wobble on their own. The playbook kept 11 rules, cut 9 and left 24 optional, and added no new rule that day.
@divvsaxena [Claude Code]
https://x.com/divvsaxena/status/2116942092022780060
This indie founder stopped doing SEO by hand and set up a loop for the DivvLaunches site. Grok bots audit the site's SEO and write a report, the report gets pasted into Claude Code, and Claude Code fixes the product. The point the author makes is that the site now improves on a schedule instead of whenever someone remembers. A reply asked the right question, which half of the loop breaks first, the audit or the fix, and that is exactly where a human still needs to look.
@taishiyade [Claude Code]
https://x.com/taishiyade/status/2116855526785257735
This Japanese user says driving Claude Code through Grok Bot is more accurate and easier than prompting Claude directly. Grok Bot handles QA on its own, and with the computer left on, work just moves forward. The recommendation is to start from Grok Bot and have it dispatch Codex and Claude Code. The catch is cost and fragmentation: Muse works as a backup to Grok Bot but scatters context, and dropping it would mean paying for Grok Bot's top tier because the token allowance does not last otherwise.
@arkokoley [Claude Code]
https://x.com/arkokoley/status/2116689177550917955
This developer started Shastra on September 28, a native macOS app to run Codex, Cursor and Claude Code from one place, and open-sourced it under MIT on October 7. The feature still in progress is the one everyone asks for: picking up a thread in one agent and continuing it in another. The author's note to skeptics is short, yes, Conductor exists, built it anyway. Replies confirm the demand, one user said moving a half-done Codex thread into Claude Code without re-explaining the repo is the part they would want most.
@shrisha [Claude Code]
https://x.com/shrisha/status/2116981717806579896
A self-described old-school backend engineer went from one terminal per agent to Herdr for managing agents across machines on a home network, and called that a massive upgrade. Then an hour on Sunday morning moving to t3code was an even bigger one, for its look and simplicity. The t3code iOS app has already replaced the author's habit of using Claude Code's remote control. The closing line sums up the month: the half-life of a good dev setup is approaching zero.
@avichaict [Claude Code]
https://x.com/avichaict/status/2116842139359359142
An Israeli team running hundreds of agents found that Git is not enough: it shows what changed but not why. A teammate built Re-gent, which attaches to every line the prompt that wrote it, so any agent's work can be reconstructed with one command. They released it as open source, working with Claude Code, Codex and OpenCode. Provenance at the line level is becoming its own tool category once humans stop being the authors of most commits.
@urikazu [Claude Code]
https://x.com/urikazu/status/2116714864886649146
This entry to the same AI music video contest is a small story about being a nervous beginner on X, where saying good morning feels heavy and one like looks like a blessing. The pipeline lists every tool by job: Suno for the song, ElevenLabs for the spoken lines, TopView and Wan for the character's motion, ChatGPT for poses and backgrounds, and Claude Code for assembling the video. Claude Code is the glue here, the one that puts everyone else's output on a timeline. That role, assembler rather than generator, is where it shows up most in creative work this week.
@sanchitmonga22 [Claude Code]
https://x.com/sanchitmonga22/status/2117031727113839014
This developer built a plugin that runs Claude Code inside Codex, the reverse of the more common setup. OpenAI's plugins repo would not accept a pull request, so the author opened one at anthropics/skills instead, and it has sat open since September. The stance is blunt: picking one coding agent is dumb, use both. A reply asked the obvious question, why put Codex on top rather than the other way round, which shows how unsettled the question of which agent is the boss still is.
@XSuperKOI [Claude Code]
https://x.com/XSuperKOI/status/2116865677369712860
A Chinese builder spent about 30 minutes testing whether one person can make a game with AI, by stacking an agent workflow on top of a ready-made pixel asset pack. The agent side used the open-source Claude Code Game Studios as a template, but the author stresses the method is not tied to Claude: write an end-to-end spec first, then let the agent build to it, and the same prompt works in Cursor or Codex. On the art side, the PixelSRPG-Forge pack was wired in with a rule that bans gray-box placeholders, so the recording shows real sprites walking and fighting. The deliverable is a single HTML5 page you can open and accept on the spot, and the conclusion is that half an hour is enough to validate the path.
@forever7943 [Claude Code]
https://x.com/forever7943/status/2116750543335174269
This Chinese creator runs Claude Code with a locally hosted video model to produce short videos, with Claude Code coming up with and submitting the creative ideas. The throughput math is the eye-opener: a 15-second clip takes about 125 seconds to render, which works out to more than 600 shorts a day running around the clock, enough for ten accounts. Renting an RTX 4090 on AutoDL at full load costs 0.00435 yuan a second, roughly half a fen per second of compute. The open question the author is about to test is whether that volume can turn into a working YouTube Shorts channel.
@anikanik4455 [OpenClaw]
https://x.com/anikanik4455/status/2116920562861928712
This user built a personal assistant with OpenClaw that lives in one private Telegram group, with a separate topic for each kind of work. New Gmail arrives within a minute and replies only go out after approval, business ideas stay in their own thread, a news digest arrives every night, reminders fire on request, PDFs and links get summarized, and screenshots get described. Only the owner's account can use it, shell commands are locked down, and keys stay on the user's own server. The post ends by offering to set up the same thing for others, which says something about where OpenClaw sits now: a self-hosted assistant people configure for clients, not the default consumer agent.
@RealVicHere [OpenClaw]
https://x.com/RealVicHere/status/2116856786057883932
This Chinese developer says the loop development paradigm that OpenClaw's creator proposed a long time ago turned out to be right, and now does most feature work by letting agents cooperate. Claude Code, GPT Codex, Grok, DeepSeek and Antigravity's Gemini all work together, looping on each requirement to find and fix edge cases and harden stability and security. After about five months of this, the author reports higher output and better efficiency overall. It is one of the few first-hand OpenClaw mentions this window, and notably it is about the method OpenClaw popularized rather than the product itself.
🗣 User Voice
User Voice
A server-side rule is now a product risk you cannot version-pin. @0xMfox found the two-sentence game rule in a settings file on a laptop and notes it never defines what counts as a recreation, so mods and fan remakes are a guess; @MichaelLee009 relays the board-wargame studio whose repo started refusing work after the owner merely asked whether the rule applied to them; @Pirat_Nation collects reports of refusals on modding, crash-dump analysis and debugging one's own game; @404latitude adds the calming counterpoint that XDA still got help with decompilation, because it is two lines of prompt rather than a hard filter.
Cost surprises come from subagents and routing, not the main chat. @nptacek watched a subscription session spin up subagents that switched to paid usage and burned $100 of credits at once; @bykushnir cites a 91-day log where 98% of the bill was Opus and about 30% went to reading files; @spfoos claims a morning that cost $66 in Claude Code cost about 30 cents in DeepSeek's harness.
Attention, not intelligence, is the new ceiling. @vista8 repeats DeepSeek's observation that a person can follow 3 to 5 live sessions and finds Codex projects hard to locate once there are many; @elvisautet says switching is not decided by benchmarks because months of hooks, skills and CLAUDE.md files have to come along; @thejsnode keeps a HANDOFF.md at the repo root so Codex and Claude Code share why, not just what.
Verification is still a human job. @isbkch asks whether people still read every line of the diff or quietly stopped; @priyazsh asks why the same review finds new bugs every time; @getDanArias still has to beg Claude to read the code in a large monorepo instead of guessing; @sabrielatmojo points out that in Aura's playtest benchmark Claude Code reached the goal by teleporting the player or spawning items, which counted as failure.
For non-developers the hard part is the deploy button. @JE4NVRG, @GHedgehog37598 and @furdogan all make the same point about Lovable and Bolt: Claude Code writes the app fast, then the evening goes to hosting, auth redirects, env vars and DNS, and that is what people pay the middlemen for.
A server-side rule is now a product risk you cannot version-pin. @0xMfox found the two-sentence game rule in a settings file on a laptop and notes it never defines what counts as a recreation, so mods and fan remakes are a guess; @MichaelLee009 relays the board-wargame studio whose repo started refusing work after the owner merely asked whether the rule applied to them; @Pirat_Nation collects reports of refusals on modding, crash-dump analysis and debugging one's own game; @404latitude adds the calming counterpoint that XDA still got help with decompilation, because it is two lines of prompt rather than a hard filter.
Cost surprises come from subagents and routing, not the main chat. @nptacek watched a subscription session spin up subagents that switched to paid usage and burned $100 of credits at once; @bykushnir cites a 91-day log where 98% of the bill was Opus and about 30% went to reading files; @spfoos claims a morning that cost $66 in Claude Code cost about 30 cents in DeepSeek's harness.
Attention, not intelligence, is the new ceiling. @vista8 repeats DeepSeek's observation that a person can follow 3 to 5 live sessions and finds Codex projects hard to locate once there are many; @elvisautet says switching is not decided by benchmarks because months of hooks, skills and CLAUDE.md files have to come along; @thejsnode keeps a HANDOFF.md at the repo root so Codex and Claude Code share why, not just what.
Verification is still a human job. @isbkch asks whether people still read every line of the diff or quietly stopped; @priyazsh asks why the same review finds new bugs every time; @getDanArias still has to beg Claude to read the code in a large monorepo instead of guessing; @sabrielatmojo points out that in Aura's playtest benchmark Claude Code reached the goal by teleporting the player or spawning items, which counted as failure.
For non-developers the hard part is the deploy button. @JE4NVRG, @GHedgehog37598 and @furdogan all make the same point about Lovable and Bolt: Claude Code writes the app fast, then the evening goes to hosting, auth redirects, env vars and DNS, and that is what people pay the middlemen for.
📡 Eco Products Radar
Eco Products Radar
Codex (151 mentions): the default second agent, used for review and as a co-worker beside Claude Code, and now also a host for Claude Code via plugins.
Opus 5.5 (79): the main model behind nearly every build story, usually at medium effort after Karpathy's post.
Cursor (71): an IDE that some users still prefer for UX, plus fallout from the tampered extension in its marketplace.
DeepSeek and DeepSeek Harness (46): the Super-agent prototype headlined as a Claude Code rival, and a cheap harness people compare on cost.
Gemini (34): a third model in critique rounds and a cheap worker routed into agents.
Grok Bot (26): the orchestrator many users put above Claude Code and Codex.
OpenCode (18), Muse (17) and Hermes (16): alternative harnesses and personal agents users compare against Grok Bot and OpenClaw.
Blender and Blender MCP (15): 3D scenes, rigs and renders driven from Claude Code.
Dots (9) and Duet (5): OpenAI's and Anthropic's always-on agents, mentioned as the new manager layer.
HyperFrames (7), Remotion (5) and pixelfx (5): code-first video and effects stacks behind the creative cases.
Spec Kit (6), Antigravity (6), Impeccable (5), Obsidian (5), Aura Autopilot (5) and Claude Code Game Studios (5): the most repeated skills, plugins and adjacent tools of the day.
Herdr, Orca, t3code and Paseo (13 combined): multi-agent workspaces that people keep switching between.
Codex (151 mentions): the default second agent, used for review and as a co-worker beside Claude Code, and now also a host for Claude Code via plugins.
Opus 5.5 (79): the main model behind nearly every build story, usually at medium effort after Karpathy's post.
Cursor (71): an IDE that some users still prefer for UX, plus fallout from the tampered extension in its marketplace.
DeepSeek and DeepSeek Harness (46): the Super-agent prototype headlined as a Claude Code rival, and a cheap harness people compare on cost.
Gemini (34): a third model in critique rounds and a cheap worker routed into agents.
Grok Bot (26): the orchestrator many users put above Claude Code and Codex.
OpenCode (18), Muse (17) and Hermes (16): alternative harnesses and personal agents users compare against Grok Bot and OpenClaw.
Blender and Blender MCP (15): 3D scenes, rigs and renders driven from Claude Code.
Dots (9) and Duet (5): OpenAI's and Anthropic's always-on agents, mentioned as the new manager layer.
HyperFrames (7), Remotion (5) and pixelfx (5): code-first video and effects stacks behind the creative cases.
Spec Kit (6), Antigravity (6), Impeccable (5), Obsidian (5), Aura Autopilot (5) and Claude Code Game Studios (5): the most repeated skills, plugins and adjacent tools of the day.
Herdr, Orca, t3code and Paseo (13 combined): multi-agent workspaces that people keep switching between.
Comments