Super User Daily: 2026-11-19
Sunday and Monday, November 16 and 17, were the days Claude Code stopped being the front door. In the setups worth copying, someone talks to one coordinator, a Grok Bot, a ChatGPT Dot, a Muse team or a plain coordinator skill, and that coordinator writes a spec and hands the actual code to Claude Code and Codex, which then review each other. One builder shipped about 160 PRs in three days from a single session, another let the two agents trade a pull request back and forth 36 times overnight, and a third upgraded a multimillion-line legacy app in six hours by playing two labs against each other. The second thread is context hygiene: a developer wrote down the compaction rules that keep a never-ending bot session sane for 19 cents, while another found that auto-compaction had invented a rule nobody said and quietly crippled the night's agents. Outside software, Claude Code made Bilibili history videos, filled in a mini-game license filing, applied for Apple's 15% fee program and mapped private equity deals in 3D. The lesson running through all of it: the people who get the most are the ones who decide in advance which step the agent is never allowed to take.
@tatsumin_ec [Claude Code]
https://x.com/tatsumin_ec/status/2122679140239462415
A Japanese e-commerce operator is building an AI organization out of Grok Bot, Claude Code, Codex and Muse, and drew the whole thing as a chart. Grok Bot is the commander and the only window the owner talks to, Claude Code is the execution team for design, implementation, documents and research, Muse handles lookups and browser chores, and Codex gives second opinions and makes images. The agents never talk to each other directly: every request goes through one intake point that checks length and count before Claude Code sees it, results get written to a shared board, and everything is traceable afterwards. The rules learned so far are concrete: sending, deleting, payments and contracts always need a human, more than 30 written decision criteria get checked before the AI asks anything, and the open problem is the approval queue, which keeps growing faster than the owner can answer it.
@suyuan1711 [Claude Code]
https://x.com/suyuan1711/status/2122220237399666700
A Chinese developer rebuilding a few always-on bots wrote up how a single never-ending chat window can stay sharp, using Claude Code's own compaction as the cautionary example. The first rule is to split what the user sees from what the model gets: scheduled jobs and bot-to-bot tasks run in hidden temporary sessions and only post the final result back. The second is that summaries of summaries rot, as a real log showed, where four specific unfixed issues shrank to the words 'those 4 problems' after three compactions, so the fix is to copy open items verbatim and give the bot a history search tool. The third is timing: compact when the context is over 100k tokens and the session has been idle for 55 minutes, just before the one-hour cache expires, which compacted a 260k-token session entirely from cache for $0.19.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2122434536822673761
The builder behind Voxyz shared the coordinator skill they run all day and overnight, and the receipts: one session rebuilt a product with about 160 PRs in three days from 107 task briefs. Each lane is one worker on one branch, with front end going to Opus 5.5, back end and shipping to Codex Sol, and anything touching money or login reviewed by Codex Astra first. A progress board refreshes every 20 seconds, and a worker only wakes the coordinator by writing a single READY line. Batching ready PRs into one push fixed a real bill problem, since main had taken 136 pushes in 32 hours and each one was a paid build, and the only things that stop and ask the human are shipping, database changes, money and deletes; once a ship worker saw green tests and tried to push the payment PRs, and it was blocked on the spot.
@btibor91 [Claude Code]
https://x.com/btibor91/status/2122807736941609261
A developer who tested GPT-6.2 Astra and Sol on daily work, from debugging to managing servers across several data centers, ran most of it through T3 Code with Claude Code and Codex side by side. The headline job was upgrading a multimillion-line legacy app with legacy dependencies from a very old version of its language to the current one: usable locally after about three hours, all tests green from a clean checkout after about six. The workflow is the interesting part: one model does the research and hands its findings to the other to see what it missed, they brainstorm fixes together, and they review each other's code. The conclusion is a two-lab setup on purpose, because Astra still finds defects in Fable's code and Fable still finds defects in Astra's.
@glennrothwell [Claude Code]
https://x.com/glennrothwell/status/2122692545578664017
A founder set up about ten personal agents in one day and directs them by voice while driving, with Grok Bot as the main bot that routes work to specialists for coding, research, design, brand, content and a podcast. The coding bot never writes code itself: it writes a short spec, gives each job its own git branch and worktree, and hands it to Claude Code or Codex, which run headless on the bot's cloud computer under the founder's own subscriptions through the official CLIs. Claude Code gets UI, copy and refactors, Codex gets backend, tests and bulk edits, they review each other, and a separate reviewer bot gates every PR on security, data access, secrets and copy rules plus a weekly scan. One AGENTS.md is the rulebook for both tools, models switch by difficulty, and every Monday a routine updates both CLIs and checks whether the labs shipped something better.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2122433137217294683
Meta's own Ads MCP lets Claude build campaigns, write ads, set audiences and read every number, and one marketer locked the money switch before connecting it. Everything Claude creates lands paused, and only one tool, ads_activate_entity, starts spending budget, so in .claude/settings.json that tool and the Instagram boost tool were put on deny, and entity updates were put under ask. The first question was which ads spent money last month and got nothing back: six ads, $412, gone. Claude rebuilt them as 12 drafts and could not launch a single one, and the human turned on three by hand, which is exactly the split of labor that makes an ad-spending agent safe.
@JamesWard [Claude Code]
https://x.com/JamesWard/status/2122388285419049243
A developer reports that phase one of a personal software factory now maintains about 50 projects and has offloaded the routine menial work, and it has already helped resolve one production incident. The build started from maintenance plus a self-improvement feedback loop, and the same skills power both the factory and manual agent sessions for feature work, new projects and issues tagged as needing a human. Guidance comes in three tiers: a factory skill for how the factory itself works, project archetype skills for websites versus libraries, and project-specific AGENTS.md files. The limit hit so far is honest and specific: the factory cannot fully manage its own subroutines with Claude Code routines, which the first architecture was built on, so the next step is testing Kiro plus Lambda MicroVMs.
@nptacek [Claude Code]
https://x.com/nptacek/status/2122803635499511846
A developer found that Claude Code's context compaction can turn an offhand remark into a standing law. During a long session, Claude was spinning up several Blender instances and pinning the machine, so the developer asked it to back off for a while and later said it could go back to full use. The next day the agents had burned plenty of tokens for little progress, and the reason was a self-imposed rule, 'NO HEADLESS BROWSERS, NO GPU USAGE', that most likely got written in during an auto-compact. Neither line was ever said, and the complaint is about the mechanism: the model's own temporary decisions survive compaction and get reified into constraints that quietly hobble the work.
@optimikee [Claude Code]
https://x.com/optimikee/status/2122379663381082317
A French developer let Codex and Claude work together overnight without supervision: 36 conversations between 1 a.m. and 7 a.m. with no break, of which only one was opened by hand. Codex fixed two of the developer's PRs, Claude reviewed, and they sent the code back and forth until version five, all inside T3 Code running on the developer's own subscriptions. The useful details are the handoffs: when the Claude limit runs out mid-task, Codex picks up the same thread from a summary, Claude can hand its fresh code to Codex for review and the original thread wakes up with the result, and agents can be moved to a Mac or a VPS when the PC is overloaded. A thread open for three days showed no slowdown, even on a nightly build that shipped three to eight versions a day.
@togethercompute [Claude Code]
https://x.com/togethercompute/status/2122543623183208778
Together AI described the agent eval loop it uses to make its own docs and APIs work for coding agents. The team defines about ten representative tasks, runs Claude Code on each in sandboxes, and tests three configurations per task: a baseline, the same task with their MCP server, and the same task with skills. Then they read the trajectories to find where the agent stumbled and fix the docs and APIs rather than the agent. Task completion went from 67% to 96%, which is a clean example of treating Claude Code as the customer whose experience gets measured.
@AntoineBlanco99 [Claude Code]
https://x.com/AntoineBlanco99/status/2122604898650894785
A non-developer running a social media product called PlugKit with a small team and Claude Code publishes a daily public log on the way from $7,262 MRR to a $100k goal. The numbers are unusually honest: €3,278 of Meta ads since September for 14 traceable customers, €234 each against a €150 ceiling, and a churn figure the founder refuses to trust because most customers have not hit a renewal yet. The Claude Code lesson came from a failure: the script that checks the ads every morning had been failing silently for three days because the founder's Mac was asleep at 8 a.m. The fix was to move that check to the cloud, a reminder that a home-laptop agent is only as reliable as the laptop.
@therosieum [Claude Code]
https://x.com/therosieum/status/2122648100493766843
A creator whose map of private equity deals blew up, with 4,000 bookmarks, explained how it was built: entirely with Claude Code. The data comes mostly from Companies House and trade press, every pin links its source, agents did the digging and the author fact-checked it by hand afterwards. The map is MapLibre on free OpenFreeMap tiles, and the 3D buildings are simply OpenStreetMap's building heights, so the whole stack is open and free. A submit button now lets readers add missing spots, and Paris and Berlin are next by popular request.
@kokushing [Claude Code]
https://x.com/kokushing/status/2122255352561615318
A Japanese indie developer rebuilt a boar monster for their own MMORPG with Tripo Studio and Claude Code, after the original model came out with its head bent to the right. The cause was the source drawing: a three-quarter body with the face turned toward the viewer, which the AI faithfully turned into a crooked 3D head. The fix was a front, side and back turnaround drawn first, the right side mirrored from the left, horizontal lines to align heights, then Tripo's Multiview to 3D, a retopo from 50,000 to 2,500 faces, and 11 animations re-rigged in Blender and tested in Godot, for 52 Tripo credits in total. The method was then written into the developer's SKILL.md with Multiview made mandatory, and three recolored variants of the monster were fixed for free along the way.
@chongdashu [Claude Code]
https://x.com/chongdashu/status/2122150154316988471
A developer submitted an entry to the Epic AI Megajam built with Fable 5.2 and an Unreal 5.9 preview, and published a making-of covering the whole path from idea to finished game. Claude Code wrote the C++ and the design notes and drove the Unreal editor over MCP, GPT Image on fal made the mockups, character sheets and all menu and HUD art, Tripo made the 3D characters, rigs and animations, Blender built the set and props by script, and ElevenLabs produced every sound effect, ambience and music track. The game runs on Windows and Mac with controller support, with web and iOS versions next, which makes it one of the more complete examples of one person shipping a full game through an agent stack.
@atakantech [Claude Code]
https://x.com/atakantech/status/2122382584076226619
A Turkish creator who was tired of paying Higgsfield built a free local image and video studio and wrote the control panel with Claude Code. ComfyUI runs as a service that starts with the desktop, with Wan 2.2 for video and SDXL-based models for images, all open source. The machine is not exposed to the internet: Tailscale gives it a private HTTPS address only the creator's devices can see, the panel sends ComfyUI workflows exported as API format with the prompt, length and image swapped in, files never pass through a cloud, and the panel checks GPU memory and switches off features that will not fit. The result is image quality close to GPT Image 2, better-than-expected motion consistency in video and zero cost per try, and the same setup runs on a rented hourly GPU by changing one address.
@yokuboukun8888 [Claude Code]
https://x.com/yokuboukun8888/status/2122503772606026173
A non-programmer in Japan set out to build internal tools with Claude Code and got three errors in a row at first. Telling it to pack everything into one HTML file still produced broken layouts and missing input checks, so nothing was usable. What changed the outcome was setting constraints up front: build on the assumption that no data is stored, and keep the branching logic simple. After that, a quote calculator and a daily report text maker went into actual use on the shop floor, and the author's takeaway is that the clumsy trial and error is what made the tools their own.
@benshandebiao [Claude Code]
https://x.com/benshandebiao/status/2122696991281140170
A Chinese creator noticed that about half of the popular-science history videos on Bilibili now seem to be made with Claude Code, so they tried making one. Claude Code did the research, built the video the way it would build a web page, did the voiceover and rendered the final file, and the human only gave instructions and made small adjustments. The author reports the result as very good. It is a small post, but it captures a real shift: explainer video is turning into a web-development task, and the agent that writes the page can now ship the episode.
@benshandebiao [Claude Code]
https://x.com/benshandebiao/status/2122682805583028511
The same creator added a step to their mini-game launch workflow: let Claude Code operate the browser and fill in the WeChat mini-game filing. The filing asks for a lot of tedious content, so Claude Code reads the game's code and the filing requirements, generates the answers itself, fills in the form and submits it. The author also notes the platform rule that matters: the WeChat Official Accounts platform does not allow AI to operate its pages, while the mini program platform does. It is a good example of agent work landing on paperwork rather than code.
@usutaku_channel [Claude Code]
https://x.com/usutaku_channel/status/2122573981697917367
A Japanese app developer applied to Apple's App Store Small Business Program, which cuts Apple's commission from the standard rate to 15% for developers with under $1 million in revenue in the prior year. The whole application was done by giving Claude Code the program's link and saying 'apply for this', and it finished almost instantly. The author's point is that every app maker who qualifies should do it, since there is no reason not to. Paying less to a platform is about as direct a return on an agent as it gets.
@minorcoworker [Claude Code]
https://x.com/minorcoworker/status/2122121629853733315
A Japanese office worker listed five single-file HTML apps worth building with Claude Code and handing out inside a company: a dashboard for CSVs exported from internal systems, a quote and pricing calculator, a text maker for daily reports, requests and prompts, checklists with branching conditions, and a searchable internal FAQ or glossary. All of them run from one file and store no data, so they can be thrown away when no longer needed. The author mostly builds the third kind, an HTML collection of original prompt templates, because Copilot's own prompt template feature is hard to use. The pattern is the opposite of a platform: tiny disposable tools that fill the gaps the official software leaves.
@amix3k [Claude Code]
https://x.com/amix3k/status/2122573980464751012
A Todoist user admitted to not understanding why so many people connected the Todoist MCP to Claude Code and Codex, until trying it. Whenever something needs a follow-up, the agent gets one line, add a follow-up for X, with an optional project, section or task URL, and it lands in the inbox otherwise. The agent turns the item into an actionable task with the context attached, and later the author just gives the agent the task URL and asks it to complete it, and it fetches the context back from Todoist. A to-do app becomes the memory that survives between agent sessions.
@Avenoxai [Claude Code]
https://x.com/Avenoxai/status/2122313517034516749
A Turkish creator went through close to 250 GitHub repos for editing video with Claude Code, then ran a controlled test on 18 of them. Every tool got the same job, turning the same five sentences into a film with the same colors, and plain Claude Code with no extra tooling did the job too as a baseline. Only a few stood out, and the author published the top ten with the films each produced. With so many video-editing skills appearing in a single week, a same-prompt bake-off is the kind of evidence most roundups skip.
@simonw [Claude Code]
https://x.com/simonw/status/2122500307557265651
Simon Willison has been exploring new ways to sandbox the code that coding agents write, and wrote up the result as the external sandbox pattern. The agent itself runs normally on your Mac, but the code it produces executes somewhere isolated, in a WebAssembly runtime, a QEMU virtual machine or an Apple Container. The write-up includes working demos for Claude Code, Codex and Pi. It separates two questions people usually blur, where the agent runs and where its code runs, and answers the second one with containment rather than permission prompts.
@_architected [Claude Code]
https://x.com/_architected/status/2122432968576876881
A developer with early access to Claude Code Projects, which is different from Claude chat projects, says it is exactly what they wanted. A project is an orchestrator for several threads: it spins them up, talks to them, coordinates, and checks in only when it needs something, and right now one coordinator is working through five tickets with one thread per ticket that it keeps prompting. The orchestrator runs in the cloud but can use remote control to run threads locally, so pet projects use cloud sessions through a personal GitHub connector while work projects run on the work MacBook Pro. Default models can be set separately for the orchestrator and for the threads.
@JustinNickles [Claude Code]
https://x.com/JustinNickles/status/2122606589798666339
One user basically runs their life on about 30 bots on an always-on Mac mini, with a chief-of-staff bot that sets priorities and hands work to specialists. The bots triage email, track nutrition, watch flight fares and Starship launches, help with the kids' learning, spot raccoons on the garage camera, keep an inventory of all the user's gear, and a Forge bot turns a needed part into a print-ready 3D file. They share a handoff folder and an HQ dashboard on the Mac mini so they can coordinate without the human in the middle. The bottleneck is usage, not ideas, so heavy development work gets pushed to Codex and Claude Code to stretch the main bot's budget, with pet face recognition, a 3D Starbase model and games with the user's son still waiting.
@VersunPan [Claude Code]
https://x.com/VersunPan/status/2122141332600926407
A Chinese developer finally finished a harness leaderboard after four days of runs, and the results surprised them. Ranked by cost per task, Codex CLI came first and Claude Code was close to the bottom, while Pi and Antigravity landed where everyone expected. Ranked by pass rate, which the author cares about more, ZCode came first, then Codex CLI, then Kimi Code, with Antigravity still last. The models used were mostly DeepSeek and Kimi K3, so this is a test of the harness with cheaper models rather than of Claude itself, and it is one more data point that Claude Code's per-task cost is a real weakness when it runs other models.
@yihui_indie [Claude Code]
https://x.com/yihui_indie/status/2122208795048964531
A Chinese indie developer spent two days doing real work on Gemini 4 Carbon, burned half of an Ultra 20x account's quota, and then brought Claude back. The verdict is specific: setting aside Antigravity's poor harness, Carbon can replace 90% to 95% of Opus 5.5 for coding, and it is the current best at plain-spoken writing and copy. But it is slow, so where the author runs four to eight parallel tasks with Claude, Gemini needs eight to twelve to keep the same throughput. And for video work, Opus 5.5 and Fable 5.2 have no rival, so the combination of slowness and the harness makes the overall experience far worse than Claude Code.
@ricbuilds_ [Claude Code]
https://x.com/ricbuilds_/status/2122677384243532177
A developer explained the one reason they still pay $100 a month for ChatGPT: the Codex plugin for Claude Code. In their view Opus 5.5 and Fable 5.2 are the best models right now, but two things are missing: image generation, where OpenAI's models are very good, and computer use, which feels like magic through Codex. The plugin lets them run the Codex CLI straight from the Claude desktop app, so both live in one place. It is a tidy example of a pattern that kept showing up this week, Claude Code as the home base that borrows other labs' strengths instead of being replaced by them.
@0xShoopy [Claude Code]
https://x.com/0xShoopy/status/2122695064304435566
A builder tested Anthropic's advisor setup on the same 3D game twice to see whether Fable 5.2 was worth its new pay-as-you-go price. With Fable doing everything the run cost $9.77; with Opus 5.5 writing all the code and Fable only reviewing the plan and the final result, it cost $5.94. Fable wrote 16K tokens instead of 148K, about nine times less, and the cheaper run finished in one pass with no fixes and both islands of the game fully working. The setup is just two commands, /model opus and /advisor fable, and the open question the author leaves is whether the all-Fable version is worth $3.83 more.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2122125276893044819
A developer who considers Fable 5.2 the best and also the most expensive model built a Claude Code plugin called fable-advisor to spend less of it without losing much of the intelligence. Fable is selected as the main model, it spins up a swarm of Opus 5.5 subagents to do the work, and it reviews the swarm's output when they are done. The author is clear about where it pays: the savings show up on big tasks, and light routine work can skip the pattern entirely. It is the mirror image of the advisor mode, with the expensive model at the top as manager instead of on call as a consultant.
@AlexFinn [Claude Code]
https://x.com/AlexFinn/status/2122146011602919832
A creator pointed out that Claude Code's interface can now be changed by prompting, and built a football theme for Sunday with live NFL scores inside the Claude UI. The point was to lock in on work all day without missing the games. The same post lists other mods worth building: a theme that always shows usage, one that changes with how many tokens were burned that day to build streaks, live analytics for the project being built, live visitor counts, and a status view of all other projects. A tool people stare at all day is becoming something its users design for themselves.
@szmichalczyk [Claude Code]
https://x.com/szmichalczyk/status/2122285141032599944
A freelancer did not want to pay for Toggl just to see basic reports of their hours, so they built a replacement. Focus is a local time tracker for Mac that covers freelance projects plus automatic activity tracking, and it counts Claude Code sessions as tracked work. It is free and open source. Small as it is, it shows two things at once: a subscription that only existed for reporting is now easy to replace, and agent sessions have become billable time worth tracking.
@shun_cmdouga [Claude Code]
https://x.com/shun_cmdouga/status/2122434540773765303
A Japanese video maker produced a commercial for a fictional anti-scam phone app, in which a swordsman slices apart the phrases scammers use on the phone. Claude handled the planning and structure, After Effects did the editing, the footage and music came from Envato, and the voice came from ElevenLabs. The service in the ad is invented and labeled as such, so this is a portfolio piece rather than a real campaign. It shows where Claude sits in a creative pipeline today: it writes the concept and the shot list, and the craft tools still do the cutting.
@BeyondTheBoxCN [Claude Code]
https://x.com/BeyondTheBoxCN/status/2122645935734042653
A Chinese account highlighted Lucca Cerf's automated game-character production line built with Claude Code, which takes a character from design through 3D modeling, rigging and animation to import into Unity. The pipeline uses Claude, Tripo, Blender MCP and Unity, with three independent AI quality-check agents in the middle. What makes it stand out is that the quality bar is written as numbers: a planted foot may not slide more than 1 mm per frame, a looping animation may not be off by more than 2 mm between its first and last frame, and the body may never pass through the floor. The video shows six characters, including a mage, a warrior, a robot and a beast, each with walk, attack, skill and victory animations, and the seven-skill workflow is open source.
@vibelancer [Claude Code]
https://x.com/vibelancer/status/2122386426671095997
A post broke down a photoreal Japanese train station that a creator built in Blender with Claude Opus 5.5, a 36-second sunset clip with 368,000 views and 4,371 bookmarks that is geometry, not generated video. The clay renders show the town behind the station is modeled too, shaders are mostly procedural, and only the foreground trees are ready-made assets, with ChatGPT for sign text and dirt stencils and Tripo for a few pieces. Claude Code planned the scene and ran subagents for modeling and visual QA, following Kamome Ashizawa's guide: references first, a CLAUDE.md rulebook, subagents for modeling and QA, and a ban on outside 3D models and texture photos. The honest catch is the human part: it took six days of review and specific fix requests, so Claude is the modeler and the person is still the art director.
@aacle_ [Claude Code]
https://x.com/aacle_/status/2122339973143945562
Cloudflare pointed an AI bug hunter at a typical 30,000-line repo and got 80 validated bugs in about three hours, ten of them critical, all patched in production within five days. The whole thing started as a 450-line Claude Code skill, which Cloudflare has open-sourced. The rule that kills fake findings is the interesting part: every reported bug has to ship with a proof of concept that runs against untouched code, and a bug without one is treated as fake. Finding bugs got cheap; proving them is now the job.
@nvandewetering [Claude Code]
https://x.com/nvandewetering/status/2122691353620447345
A user says they have spent more than 100 hours on voice calls with their ChatGPT Dot, talking through the whole day in the car or on the treadmill, so that by the time they reach the office the work is already done. The key step was connecting Claude Code to the Dot's computer, so the Dot hands tasks off to Codex and Claude Code agents and follows best practices while doing it. The user's favorite part is that the usage feels close to unlimited. It is a clear picture of the voice-first layer, where the human never types and the coding agents sit two steps away.
@chrismaconi [Claude Code]
https://x.com/chrismaconi/status/2122731271021326413
A heavy Grok Bot user posted follow-up numbers after the recent token limits: their team was getting about 1.4 billion tokens a week and is now tracking toward about 2 billion. The setup is about 40 Grok Bot agents working together with an Ultra Claude Code subscription at $200 a month for software-factory and coding work. The observation that matters for Claude Code users is that Grok Bot appears to serve Opus 5.5 for most inference, and the author now considers the weekly usage reasonable for the money. The advice is that the burden has shifted to users, who can make their agent teams more token-efficient without hurting results.
@korosuke_tv [Claude Code]
https://x.com/korosuke_tv/status/2122280631564730867
A Japanese user showed their Genspark GenTeam screen with seven agents, each with one job: a general fixer that corrects itself when told it is wrong, a calendar agent with Google Calendar permission that books open slots, a coordinator that runs a daily 18:00 check on DeepSeek to save credits, a diet coach that reads food photos and replies with calories, an English coach that reviews AI notes from real lessons, one agent that instructs Claude Code, and another connected to the Mac mini at home. The design matters more than the count. Permissions are split per agent, so calendar access goes only to the calendar agent, models are chosen per agent so light work runs on light models, and agents can live on the home machine as well as in the cloud.
@Botan_cr [Claude Code]
https://x.com/Botan_cr/status/2122226319085965742
A Japanese account shared a creator's swimmable underwater world built as a single 97 KB HTML file. The code was written with Claude Code, and the creator used no images, no 3D models and no external libraries. The seabed and roughly two million seagrass blades are generated in code and rendered with WebGPU, and visitors can explore schools of fish, manta rays and a sunken ship. It is a reminder that an agent plus a hard constraint can produce something no asset pipeline would: an entire ocean small enough to email.
@rimtoln [Claude Code]
https://x.com/rimtoln/status/2122370200683135362
A user who wanted Claude to join a meeting found an open incident instead, Claude Meetings in beta could not connect to new calls, and then went and counted. From November 8 to November 16 Anthropic's status page logged nine incidents, including Apple sign-up failing for some new iOS users for 56 hours and a November 13 with three separate problems spanning login, the API and Claude Code. Only three of the nine were about the models, and the longest of those lasted 73 minutes; everything else broke around them, in calls, sign-in, billing and connectors. Ninety-day uptime reads 99.76% for the API and 99.46% for claude.ai, roughly 5 hours down versus 12, which the author reads as the cost of turning a model company into a suite of apps.
@yurinakanishi33 [Claude Code]
https://x.com/yurinakanishi33/status/2122693197503836532
A Japanese maker printed a pegasus on a Bambu Lab A1 mini, a 3D printer currently discounted for Black Friday. The modeling was done with TRELLIS.2 together with Claude Code and Codex. It is a short post, but it is part of a steady stream this week of agents reaching off the screen: models that start as a prompt and end as an object on a desk.
🗣 User Voice
User Voice
The approval queue is the new bottleneck, and people want it designed, not just tolerated. @oanaolt asks how teams stop one engineer's agent from quietly contradicting an architecture decision another engineer's session made weeks earlier, and @emilylambert asks the people running agents day and night what those agents actually do, because the honest examples are scarcer than the claims. @dontbesilent pushes back on the fully automated video editing hype: cutting filler is solved, but moving a highlight to the front is a copywriting decision, and knowing whether a line fits the account and the business model is beyond even Claude Code with Opus.
Projects and Remote Control have early fans and sharp critics. @Zeta4994 tried the Claude Code Projects beta and went back to normal chat, saying the coordinator got stuck on low effort, nearly broke the work and could not run remote sessions, and @izutorishima finds the Codex app better built than Claude Code Desktop and wants Remote Control fixed.
Limits and costs are still the loudest single complaint. @DavidOndrej1 calls the five-hour limit ridiculous, @altantutar runs out of Grok Bot credits in a day or two and wants an open-source equivalent that runs on an existing Claude Code or Codex subscription, and @minorun365 found that four months of personal use would have cost about 9.3 million yen without a flat-rate plan.
Moving and managing the setup itself is harder than it should be. @brandenflasch asks how to move Claude, past transcripts and Claude Code setups from one Mac to another, and @signulll wants GitHub rebuilt for agents as first-class users, since threads turn unreadable, pages load slowly, and even a tiny company now strains the platform with agent PRs.
Some users argue the harness should get out of the way. @valigo says 90% of success comes from the model and that Opus and Fable can do worse inside Claude Code because it intervenes too much, while @yoavgo reads the Claude Code team's 90% agent-written claim as a sign of how simple and low-risk that codebase is rather than proof of the testing method. @nft_dreww is one of several users publicly debunking viral videos that claim Claude Code found hidden spyware, pointing out scripted logs and protections the claims ignore.
The approval queue is the new bottleneck, and people want it designed, not just tolerated. @oanaolt asks how teams stop one engineer's agent from quietly contradicting an architecture decision another engineer's session made weeks earlier, and @emilylambert asks the people running agents day and night what those agents actually do, because the honest examples are scarcer than the claims. @dontbesilent pushes back on the fully automated video editing hype: cutting filler is solved, but moving a highlight to the front is a copywriting decision, and knowing whether a line fits the account and the business model is beyond even Claude Code with Opus.
Projects and Remote Control have early fans and sharp critics. @Zeta4994 tried the Claude Code Projects beta and went back to normal chat, saying the coordinator got stuck on low effort, nearly broke the work and could not run remote sessions, and @izutorishima finds the Codex app better built than Claude Code Desktop and wants Remote Control fixed.
Limits and costs are still the loudest single complaint. @DavidOndrej1 calls the five-hour limit ridiculous, @altantutar runs out of Grok Bot credits in a day or two and wants an open-source equivalent that runs on an existing Claude Code or Codex subscription, and @minorun365 found that four months of personal use would have cost about 9.3 million yen without a flat-rate plan.
Moving and managing the setup itself is harder than it should be. @brandenflasch asks how to move Claude, past transcripts and Claude Code setups from one Mac to another, and @signulll wants GitHub rebuilt for agents as first-class users, since threads turn unreadable, pages load slowly, and even a tiny company now strains the platform with agent PRs.
Some users argue the harness should get out of the way. @valigo says 90% of success comes from the model and that Opus and Fable can do worse inside Claude Code because it intervenes too much, while @yoavgo reads the Claude Code team's 90% agent-written claim as a sign of how simple and low-risk that codebase is rather than proof of the testing method. @nft_dreww is one of several users publicly debunking viral videos that claim Claude Code found hidden spyware, pointing out scripted logs and protections the claims ignore.
📡 Eco Products Radar
Eco Products Radar
Codex: the standing second seat in nearly every multi-agent setup, as reviewer, backend lane, fallback when Claude limits hit, and the plugin people keep for images and computer use.
Cursor: mostly named as a comparison point and in agent lists rather than as the main tool.
Pi and OpenCode: lightweight open harnesses that users compare against Claude Code and extend into Claude Code look-alikes.
Grok Bot: the front-door orchestrator of the week, routing coding work to Claude Code and Codex on its cloud computer.
Blender and blender-mcp: the creative workhorse for scenes, characters and product shots driven from Claude Code.
Gemini 4 and Antigravity: Google's model earns praise while its harness draws complaints.
DeepSeek and Kimi: cheap models plugged into harness tests and light agent roles.
Muse: the friendly personal agent that people run beside Claude Code and sync skills into.
Unity and Godot: game engines at the end of Claude Code asset pipelines.
T3 Code and Orca: fleet tools for running Claude Code and Codex side by side in worktrees.
Tripo: image and text to 3D for game characters and props.
Higgsfield and Seedance: AI video generation feeding agent-built content pipelines.
Remotion and FFmpeg: the code-to-video layer under most agent editing workflows.
Hermes: another personal agent whose plugins approve Claude Code edits from a watch or speaker.
Codex: the standing second seat in nearly every multi-agent setup, as reviewer, backend lane, fallback when Claude limits hit, and the plugin people keep for images and computer use.
Cursor: mostly named as a comparison point and in agent lists rather than as the main tool.
Pi and OpenCode: lightweight open harnesses that users compare against Claude Code and extend into Claude Code look-alikes.
Grok Bot: the front-door orchestrator of the week, routing coding work to Claude Code and Codex on its cloud computer.
Blender and blender-mcp: the creative workhorse for scenes, characters and product shots driven from Claude Code.
Gemini 4 and Antigravity: Google's model earns praise while its harness draws complaints.
DeepSeek and Kimi: cheap models plugged into harness tests and light agent roles.
Muse: the friendly personal agent that people run beside Claude Code and sync skills into.
Unity and Godot: game engines at the end of Claude Code asset pipelines.
T3 Code and Orca: fleet tools for running Claude Code and Codex side by side in worktrees.
Tripo: image and text to 3D for game characters and props.
Higgsfield and Seedance: AI video generation feeding agent-built content pipelines.
Remotion and FFmpeg: the code-to-video layer under most agent editing workflows.
Hermes: another personal agent whose plugins approve Claude Code edits from a watch or speaker.
Comments