Super User Daily: 2026-09-28
The most useful Claude Code story today is also the least glamorous: a traveler asked it to turn a Linux laptop into a hotspot routed through Singapore, and it just did the networking. That is the week in one line. Opus 5.5 keeps pulling people back from Codex on quota and harness quality, and the cases are spreading from code into video, courseware, dividend spreadsheets, outreach and email ghostwriting. Two counter-currents run alongside: people are now publishing real bills and pruning their skill folders instead of piling on more, and the hard part is shifting from generating output to proving it works, with users asking Claude to prove each layer of a release. On the OpenClaw side, Microsoft's Autopilot and Meta's Muse made the open harness a substrate, and the most interesting builders are wiring Muse, Claude Code and Codex together as one team.
@imbktan [Claude Code]
https://x.com/imbktan/status/2103832552226357359
Traveling in China, this user had Claude Code turn a Linux laptop into a Wi-Fi hotspot for the whole group. The laptop sat on hotel Wi-Fi with Tailscale routing through a home router in Singapore, and Claude Code configured the hotspot, routing, NAT and firewall so everyone could get online through that Singapore connection. No manual NetworkManager work at all. The user's own summary is the best line of the day: coding agents now feel like a portable IT guy.
@readyfor2025 [Claude Code]
https://x.com/readyfor2025/status/2103822058455265518
A parent built an interactive architecture courseware for their kid's class with Claude Code, Opus 5.5, Blender and three.js: about 30 hours of human time plus 100 hours of Claude Code working alone, all driven by voice, half of it lying in bed steering through remote control on a phone. Claude even recorded and edited the demo video. The user estimates the same software used to need a 3D modeler, a 3D developer, frontend and backend engineers, a UI designer and a product person for a month. The detail that matters: the same job stalled for three days on Fable 5 and Opus 5 because it is mostly visual work, and Opus 5.5 is what got it over the line.
@AndreiProvkin [Claude Code]
https://x.com/AndreiProvkin/status/2103914892109517008
A rare full bill for a long creative project: the most graphically complex fully procedural world the author could reach, with terrain, trees, rocks, a river, a golem and the weapon in hand all generated by code and zero downloaded assets. It took 72 Claude Code sessions and 79 hours of active agent time, 9,963 requests and 2.53 billion tokens, 98.35% of them cache reads, about $1,738 at API prices. About 87% of requests ran on Opus 5, with the tail on 5.5 and Fable. The cache-read share is the real lesson: long creative projects are cheap only because almost everything is re-read, not re-generated.
@merill [Claude Code]
https://x.com/merill/status/2103984759269732838
A developer turned a public write-up of a JWT signature-bypass bug in Microsoft Entra apps into a reusable audit prompt and ran it with Claude Code across three of their own apps. In minutes it found every place the apps accept a token, confirmed signatures were verified everywhere, and still surfaced two hardening gaps now being fixed. The prompt forces a read-only pass first, a table of every token entry point with file and line references, and a check of what libraries like MSAL do and do not validate. This is the pattern worth copying: news of a vulnerability class becomes a same-day audit of your own repo.
@MIYAGIdesign [Claude Code]
https://x.com/MIYAGIdesign/status/2103972632949604783
This user fed Claude Code with Opus 5.5 and HyperFrames their memory and Obsidian vault and asked for a promo reel, got a 15-second cut, found it too fast to read, and had it re-cut to 34 seconds. The pipeline is HTML and JavaScript that HyperFrames renders locally to MP4, and the only extra cost was about 40 yen for five AI-generated background tracks. The underrated point is determinism: the same code always renders the same video, so fixing one scene never changes the others, unlike video models where one tweak reshuffles everything.
@y_matsuwitter [Claude Code]
https://x.com/y_matsuwitter/status/2103852488558530773
A Japanese founder played with video generation using Claude Code, Opus 5.5, Gemini 3.8 Flash TTS and Lyria 3.5 and came away worried as a businessperson. With no editing software at all, the agent assembled the video with plain tools: Canvas, Playwright, and even Python with numpy and SciPy to synthesize sound effects from waveforms. The takeaway is not that AI can make video, it is that the agent builds the minimum sufficient toolchain on the spot. The user now sees tutorials, promos and even earnings-presentation videos as something any service with its own codebase can generate.
@dashboardlim [Claude Code]
https://x.com/dashboardlim/status/2103907678103781700
This user has started using Claude Code less as a coder and more as a release auditor, because merged is not the same as working. In one session the PRs were merged and deployed, but Claude still surfaced a Microsoft login callback that needed checking outside the repo, while an n8n workflow looked to have died from memory pressure. The new routine is to ask Claude to prove each layer separately: the merged PR, the production deploy, the external dependency, runtime evidence, and the user path actually working. It is the clearest description yet of verification as the job.
@rdylina [Claude Code]
https://x.com/rdylina/status/2103649946981507404
A project that stalled in Codex, where subagents struggled past three to five parallel tasks, was handed to Claude Code, which happily spun up as many as 16 threads each running multiple subagents. The user reports progress roughly doubled compared with Codex. Their read is that this does not feel like a model difference but a more advanced harness, especially the task orchestration and the new projects interface.
@DrSN1392 [Claude Code]
https://x.com/DrSN1392/status/2103734268526821885
An individual investor updated their September high-dividend stock list and published it as an xlsx, with every field except the margin-trading ratio collected by Claude Code. It is a small but very real non-coding use: a recurring research chore that used to mean hours of copying numbers now runs as a collection job.
@PaulKlayVC [Claude Code]
https://x.com/PaulKlayVC/status/2103933648659460527
A founder open-sourced an outreach skill and showed it working from one sentence: a B2C meditation app looking for partners who already have its audience. After four quick questions, the agent ruled out sleep podcasts as competitors, pointed instead at mattress and sleep-accessory brands whose customers already return products because they did not help them sleep, and reframed the pitch as a card in the box, a free month, a revenue share and a lower refund rate. It then found the affiliate manager listed on each company's own page, wrote the first email and a two-week outreach plan.
@masashi20555444 [Claude Code]
https://x.com/masashi20555444/status/2103959043286974631
A Japanese office worker is publishing a book on systematizing internal document creation with Claude Code and bundling the exact skill as a free extra. In their day job, one document used to take two hours and now takes ten minutes. The system interviews the author first, asking who will read it and what decision it should drive, and the human only answers, chooses and checks. The book also documents where it broke along the way and how it was fixed.
@mika_15ai [Claude Code]
https://x.com/mika_15ai/status/2103847791114035215
A Kindle self-publisher who says they do not understand even the vocabulary had Claude Code rebuild a sales-rank checking app that a community member was retiring at month end. It took a few hours instead of the few minutes an expert would need, but it works and even sends email notifications. This is what non-developer adoption actually looks like: replacing a tool you depended on before it disappears.
@menmarsan [Claude Code]
https://x.com/menmarsan/status/2103728429577416834
Six months into their creator activity, this user handed Claude Code with Opus 5.5 at max the entire half-year activity folder and asked for a 15-second recap, free to use any material inside. The result pulled in moments the user had forgotten. It is the personal-archive version of the code-rendered video wave: the input is a folder of your own life, not a prompt.
@HiTw93 [Claude Code]
https://x.com/HiTw93/status/2103789033864491058
An open-source developer spent a gifted week of quota on real product work instead of demos: Opus 5.5 on high inside Claude Code is what made the Windows version of Mole take shape fast enough to test. The setup is a Mac for writing native Windows C# code, Parallels running Windows, and the agent driving software inside the VM to test results. They split work across subscriptions: Codex with Astra Medium for requirement analysis and code review, Cursor's Grok models for open-source issues and PRs, and they explicitly reject token-usage leaderboards that ignore what the tokens produced.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2103824976713306214
One prompt asked Opus 5.5 to use Dynamic Workflows and make an animated movie-shoot about its own efficiency in Claude Code. Ultracode turned Claude into an orchestrator: a pre-production writers' room, parallel agents building character, sets and render pipeline, a soundtrack written entirely in Python, seven scenes built against an adversarial reviewer, and a final polish workflow. The user had to board a flight mid-run, and Claude picked the workflow back up cleanly on landing.
@ashen_one [Claude Code]
https://x.com/ashen_one/status/2103853135957500295
A clean usage datapoint: Opus 5.5 on xHigh ran a /loop overnight for 9 hours and cost 25% of the $200 monthly plan's allowance. The same night, GPT-6 Astra on Low running a /goal burned the remaining 35% of the user's Codex allowance in 2.5 hours. Both were doing fairly complex work with very different tools, so it is not a controlled test, but it explains the week's mood shift back toward Claude Code.
@hraness [Claude Code]
https://x.com/hraness/status/2103949013401968775
A power user posted their Saturday setup: 30 AI subscriptions, 17 ChatGPT Pro, 7 Devin Max, 5 Claude Max and one Cursor Ultra, running 39 terminals in Zed. 28 are Devin with SWE-2, which is free and unlimited on the Max plan, and 11 are Claude Code with Opus 5.5, each with many subagents inside. It is an extreme case, but it shows where the heaviest users are: routing volume to whatever is unlimited and saving the frontier model for the hard lanes.
@TJwithAI [Claude Code]
https://x.com/TJwithAI/status/2103741081259315254
A Chinese developer found a split that stretches both subscriptions: GPT-6 Astra as the brain that understands the task, breaks down requirements and decides the plan, Opus 5.5 in Claude Code as the hands that write code, edit files, run tests and fix bugs, then Astra again as the acceptance reviewer. An earlier attempt with Astra plus DeepSeek V4.1 Flash turned out less cheap than expected once context, call counts and rework stacked up. The conclusion is the same one many reached this week: saving money is less about the cheapest model than putting the right model on the right step.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2103849152035831873
A practical recipe for getting more out of a Claude subscription: one main chat on Opus 5.5 xhigh as orchestrator, the work session planned as tickets with Matt Pocock's skills, each subagent loaded with exactly one ticket and minimal context. Complex tickets go to Opus 5.5 medium and everything else to a very cheap model, which does not even have to be Claude: Claude Code can call another harness as a headless worker, like pi.dev with DeepSeek V4.1 Flash or Codex with GPT-6 Luna Fast.
@wh0sumit [Claude Code]
https://x.com/wh0sumit/status/2103826938502558203
A startup team generated a full video for their brand in about ten minutes with Claude Code, motion, sound design and voiceover included, posted it, then deleted it. Not because it was bad but because it was good: when creating becomes almost free, the hard part becomes deciding what deserves to exist. They chose to keep their Instagram human-made with AI assisting, a rare case of a team using the capability and then drawing a line on purpose.
@ladprofit [Claude Code]
https://x.com/ladprofit/status/2103873970886029533
A zombie ad for a client cost $8.94 end to end and was built entirely inside Claude Code, with the human only approving the concept and storyboard. The pipeline: character sheets so all five characters stay identical in every shot, a start image per scene, animation in Seedance 2.5, editing and captions by Claude, and externally made music mixed in automatically. The price tag is the story: commercial-grade short ads have fallen under ten dollars.
@dotey [Claude Code]
https://x.com/dotey/status/2103683057689522564
A well-known Chinese AI writer made an explainer video on what a Transformer is with Claude Code and Opus 5.5 from one short prompt: make a JS video that a high schooler can follow, cover attention and some math, use any tools or install any tools, go online if needed, and surprise me. It became one of the day's most shared examples because it chains research, explanation, code and final render into one loop with no video model involved.
@blacklist_ryu [Claude Code]
https://x.com/blacklist_ryu/status/2103977996042592652
A Japanese hobbyist built a free flight simulator with Claude Code on Google Earth maps and Cesium: controller support, a Cessna, an F-35, a jumbo jet and a helicopter using free 3D models, instruments, switchable camera views, airport or mid-air starts, voice warnings, building collision so you can crash, and landings on airports or rooftop helipads. No paid APIs, so it costs nothing to play. The demo flies over central Fukuoka.
@francedot [Claude Code]
https://x.com/francedot/status/2103909301605761276
A 20-second Pokemon Johto Autumn short made in two hours with Opus 5.5 in Claude Code runs in real time from one HTML file: raw WebGL2 and GLSL raymarching, rigged glTF models, and a web audio score written in code. The author offered to open-source the whole session. It is the realtime cousin of the rendered-video trend, where the output is a program rather than an MP4.
@sbalhatlani [Claude Code]
https://x.com/sbalhatlani/status/2103812370921316844
A creator made a Japanese anime opening for their project Shard of Dawn by stacking tools: Opus 5.5 wrote the lyrics and refined every prompt, Claude Code served as the engine, MiniMax H3 Max Turbo generated video via fal, gpt-image-2.5 made the art, Suno v6 made the song, and gpt-audio-1.5 synced the beat. Claude Code here is not the renderer but the conductor that keeps six models in time.
@hasegawa_ra [Claude Code]
https://x.com/hasegawa_ra/status/2103749621109706955
A 51-year-old first-timer took on AI video generation and finished a piece in 4.5 hours using Kling, Claude Code, Gemini, Final Cut Pro and Logic Pro. The self-assessment is modest, not easy but decent for a first try, which is exactly what makes it useful: the tools have reached people who still edit in Final Cut and mix in Logic.
@yuu_aishorts [Claude Code]
https://x.com/yuu_aishorts/status/2103755249534251479
A short-video creator, whose boss had already noticed the side project, used Claude Code to analyze videos with over a million views and found them eerily standardized: not just the hook, but the shot rhythm and even the audio loudness. The plan is to replicate those three variables and speed up automation of their own shorts.
@kaede_gpt [Claude Code]
https://x.com/kaede_gpt/status/2103745387098177965
A TikTok creator runs the entire pipeline from Claude Code: analyze a reference video, write the script, generate the clip, then push it straight into TikTok drafts. Generation still uses Higgsfield and editing another AI, but both are driven from Claude Code, so the whole video is effectively delegated end to end.
@kutaro_ai [Claude Code]
https://x.com/kutaro_ai/status/2103975230373028303
A brand-new TikTok account automated with Claude Code is getting repeated 10,000-plus views just three days in. The user's own AI analysis says the algorithm rewards watch quality, completion, repeats and saves, and expands distribution later, so the account may still grow. Worth tracking as a data point on automated content, not yet as proof.
@RCIllingworth [Claude Code]
https://x.com/RCIllingworth/status/2103891916177678492
A cold-email operator rebuilt the whole process inside Claude Code in eight steps: pull a targeted list from Apollo from the ICP and offer, enrich every contact in parallel by scraping their site and tech stack, validate through a verifier, write sequences from the enrichment with the PPQ framework, QC every variant under 80 words, load into Instantly with daily caps of 15 to 20 per inbox and open tracking off, send from warmed Google and Outlook inboxes, then let Claude say which segments to scale or cut.
@minorun365 [Claude Code]
https://x.com/minorun365/status/2103700370199601260
A Japanese user quietly lets Claude Code handle some of their email and Slack replies even when not announcing it, and has trained it so heavily on their writing style that people mostly do not notice. It is a small post that raises a real question for everyone else: how much of your inbox is already answered by someone's agent.
@dielanoRdinG [Claude Code]
https://x.com/dielanoRdinG/status/2103776679650857082
A trader took three open-source Polymarket bots that match paid versions costing five to six thousand dollars, used Claude Code to change the configs, and only went live after simulation runs passed. One bundles over 118 strategies, one scans top traders by PnL and win rate for copy trading, and one manages limit orders to farm liquidity rewards. The user's point: the gap between paid and open source is whether you can modify, test and wire it into your own system.
@fexisnpc [OpenClaw]
https://x.com/fexisnpc/status/2103865671608307838
A Chinese tinkerer's month in side projects: a 25-yuan Bluetooth controller officially unsupported on Windows had its control protocol reverse-engineered and a ViGEm driver vibe-coded so every function works on PC. They also built a Docker reverse proxy for a loopback-mode OpenClaw so its Chat API and Control UI are reachable over Tailscale and wired into home-made work tools, and their daily work now leans heavily on a self-built reviewer agent.
@kidonaru [Claude Code]
https://x.com/kidonaru/status/2103678499747938783
A game modder released COM3D25.DevBridge, a plugin for Claude Code and Codex that lets the AI restart the game and enable a plugin to test it, or drop in a mod and see models and outfits update while the game is running. It closes the loop that usually makes modding with agents painful: the agent can now see its change in the actual game.
@m_matsubara [Claude Code]
https://x.com/m_matsubara/status/2103700408095113290
A long-time Japanese database tool developer added an MCP server to a work-in-progress build of A5:SQL Mk-2, so Claude Code can now write and fix SQL through it. The point is environments where API keys are not allowed: the agent reaches the database through the desktop tool the team already trusts. The author is candid that real-world usefulness is still to be proven.
@mocchicc [Claude Code]
https://x.com/mocchicc/status/2103693821251125350
A developer built an imagegen-via-codex skill for personal use: inside Claude Code, Fable or Opus writes the image prompt, and the CLI hands it to Codex to actually generate the image. The results have surprisingly good taste. It is a neat example of cross-vendor composition, letting each model do the part it is best at.
@mankaff_ [Claude Code]
https://x.com/mankaff_/status/2103836093368135905
Fed up with Claude Code asking for an API key but refusing to accept it in the session and nudging them to rotate it instead, this user built human-input, inspired by Matt Pocock's wizard skill, so secrets can be entered without leaving the Claude Code interface. It is a small tool that exposes a real friction: safety nudges that are right in principle but block the obvious workflow.
@mika_systems [Claude Code]
https://x.com/mika_systems/status/2103862116469928158
A developer checked nine Claude Code sessions and logged 422,164,852 cached input tokens, because earlier file reads kept riding along with every later request. So they built a Jev plugin that selects the relevant code before Claude reads it. Their paired tests show 30% fewer billed input tokens for finding code, 38% for checking files, 41% for narrowing large reads, and 12,190 tokens cut to 2,722 on three file-reading tests with all answers still correct. A search for the register flow found the right four files even though the code called it signup and the interface was in French.
@FeiFei_io [Claude Code]
https://x.com/FeiFei_io/status/2103834009977684312
Importing locally installed skills into CC Switch, a Chinese user discovered nearly 100 of them, more than 60 for Claude and over 30 for Codex. The advice that follows is worth repeating: prune skills regularly, because every available skill's name and description can enter the model's context, duplicates add noise and tokens, and today's agents often do better without heavy workflow skills forcing extra steps.
@ziwenxu_ [Claude Code]
https://x.com/ziwenxu_/status/2103748089089806819
One of the first people to run the new /checkup prompt-audit command on their setup removed more than 20 skills and MCP servers that had sat idle for days. The audit checks CLAUDE.md, skills and agents for patterns written for older models. It pairs with the CC Switch cleanup above: the week's quiet trend is subtraction.
@makaneko_AI [Claude Code]
https://x.com/makaneko_AI/status/2103817749239836710
A Japanese developer tested more than 100 implementations and found detailed planning templates no longer help with Opus 5.5 or GPT-6 Astra, because the newest models already know how to write requirements. One line in the system instructions is enough: for medium or larger work, save what was agreed with the supervisor to docs/specs/yyyymmdd_change-name.md with goals, completion conditions including failure cases, and constraints, but no implementation steps. It works across Claude Code and Codex.
@seiichi_satoweb [Claude Code]
https://x.com/seiichi_satoweb/status/2103656670744445320
A Japanese writer shared the exact prompt that first worked for turning their article-writing process into an AI agent: attach the process documents and rules, ask Claude Code to build an agent, and invite it to propose whichever combination of modules, skills, hooks, plugins, MCP, subagents and RAG fits. What made it work, in their words, was not deciding the features themselves and letting the AI choose.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2103977414711767244
A pre-launch checklist for vibe-coded sites: have Opus 5.5 send out five subagents to check twenty things in one pass, grouped into design consistency, mobile, every UI state, real-user click-throughs and launch basics. Each subagent only reads, the main session fixes everything from their lists so no two agents edit the same file, effort stays at medium except for harder groups, and nothing changes until the user sees the issue list and gets before-and-after screenshots.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2103946635831050740
A reusable subagent for long unattended runs: a dashboard-builder on Opus at medium effort that can only write to a .dashboard folder and its own memory, asks your style once and remembers it, and builds a self-refreshing HTML dashboard in about two minutes before any task longer than 30 minutes. The dashboard shows progress, what is stuck, questions waiting on you, and the default it will take if you do not answer, so the main session keeps working instead of blocking on you.
@celinehaverstad [Claude Code]
https://x.com/celinehaverstad/status/2103777480473710970
A Saturday refactor experiment: Claude Code's creator reportedly runs up to 15 sessions at once, so this user tried. At six sessions they were reviewing more than building and went back to four, using worktrees, subagents and a lean CLAUDE.md. The honest ceiling is set by human review bandwidth, not by the tool.
@runtimeking [OpenClaw]
https://x.com/runtimeking/status/2103853989024628845
A solo founder says they are officially off the terminal and local development: a Claude-native workflow now hands back completed tasks backed by full end-to-end QA and verification, and they keep throwing more ambitious tasks at it. Their role moved from building with Claude to approving what Claude does. They tried the same idea early in the year with OpenClaw and Hermes and found it hacky; now it runs across code, design, marketing and support from anywhere.
@sakamoto_582 [Claude Code]
https://x.com/sakamoto_582/status/2103991264240959976
A Japanese engineer reflects on how fast the IDE died. A year and a half ago the norm was Cursor with a chat pane and accepting or rejecting changes chunk by chunk; when Claude Code arrived they kept using it as a VS Code plugin to watch the code. Now they barely open an IDE, never accept or reject, let Auto implement, and if they look at code at all it is the GitHub diff.
@fumokmm [Claude Code]
https://x.com/fumokmm/status/2103846048276091018
A Japanese developer writing Claude Code procedures for a team was told by a senior reviewer to pin the model to Sonnet 4.6, on the classic principle of using mature technology and not updating docs every time the model changes. The rollout is scheduled for December. Their point is sharp: with AI, nothing is less useful than a mature model, and the enterprise instinct for stability is quietly locking teams a generation behind.
@Pirat_Nation [Claude Code]
https://x.com/Pirat_Nation/status/2103681213235962105
A widely shared incident: a user says Claude Code deleted 48,218 files in 103 seconds while rebuilding a file mirror, after Windows directory junctions let a cleanup step reach outside the intended folder. The damage included the Git object store, so the normal recovery path was gone too. As follow-up analysis noted, the destructive command lived inside a helper script the agent wrote itself, which command-level permission checks cannot inspect before it runs.
@fujibee [Claude Code]
https://x.com/fujibee/status/2103809306860847486
A developer of agmsg, a system that lets Claude Code, Codex and Grok agents address each other like named teammates, tried Meta's Muse and asked whether it could join. Muse offered to write its own agmsg skill, then, unprompted, listed the transport gaps: reads are fine by syncing the DB into its VM, but writes must go through Terminal via computer use to respect agmsg's rule that messages always pass through the official script, and real-time push is impossible so it would poll. It even specified the driver type to add on the agmsg side. The builder now plans Muse support.
@Zacitus [Claude Code]
https://x.com/Zacitus/status/2103970897203990757
A user gave Meta's Muse access to Claude Code, Codex, SuperGrok and a $30 Gemini API key and uses Muse as the coordinator: it leans on agents from each service to build websites and to make decisions and validate its own answers. The consumer agent becomes the manager and the coding agents become staff, an inversion several people joked about the same day.
@road_ninjart [Claude Code]
https://x.com/road_ninjart/status/2103972333010763874
A Japanese developer built Claudex, a Mac app where Claude Code and Codex work together: both can tackle the same instruction independently, or recognize each other's strengths and weaknesses and split the work. It is one of several tools this week treating two rival agents as a pair rather than a choice.
@DottChen [Claude Code]
https://x.com/DottChen/status/2103656886809858394
A small team launched Merlin after two months of building and daily internal use: mention @Merlin in Slack, Linear or GitHub and the request is routed to your local Claude Code or Codex, which replies in the same thread, issue or PR. Teammates see the local agent's progress where the conversation already happens, on your own subscription, with no migration.
@bensig [OpenClaw]
https://x.com/bensig/status/2103867133252006262
A user built a personalized daily newspaper with OpenClaw that covers only the topics they care about, drawn only from sources they trust. It is the simplest possible always-on agent use and still one of the most compelling: a curated front page no algorithm controls.
@augmentedtraff [OpenClaw]
https://x.com/augmentedtraff/status/2103885594296537184
A maker has been running an AI smartwatch built on M5Stack's StopWatch for weeks, well before Meta announced Muse Charm. It talks to OpenClaw or ChatGPT, with two buttons for nine actions plus touch, smart controls including car climate and briefings, tactile vibration, and a lanyard with a magnetic clip. The open agent stack is already reaching custom hardware.
@menhguin [OpenClaw]
https://x.com/menhguin/status/2103951843793768516
A practical OpenClaw hosting datapoint: the user first ran it on a VPS, but at 10 to 20 GB of RAM a Mac mini turned out cheaper after a year, adds useful Apple features, and can be resold for roughly what it cost. That is the unglamorous reason so many always-on agents live on a Mac mini.
@MichaelGannotti [OpenClaw]
https://x.com/MichaelGannotti/status/2103942046813692331
An OpenClaw power user tracking merged PRs reports that Team Reports collection had been chewing the Gateway event loop and now runs in a worker with bounded SQLite batches. On their fixture, peak heap fell from about 96 to 52 MiB and the longest event-loop turn from about 69 to 19 ms, with report output unchanged. The same morning they noted a merge letting owners paste keys into chat as a SecretRef that is never echoed back.
@coder_left [Claude Code]
https://x.com/coder_left/status/2103678591687159844
A Chinese user woke up to a banned Claude Code account, and says the only notable thing they did was ask it to reply in Chinese. The claim cannot be verified from outside, but it spread widely because it matches the ambient fear among Chinese users that account safety depends on signals they cannot see.
@leeoxiang [Claude Code]
https://x.com/leeoxiang/status/2103831166977737074
Another Chinese researcher lost a Claude account used for over two years, with Fable 5 as their daily model for algorithm experiments, and asked whether appeals ever succeed and how others keep a Claude Code Max plan stable. The productivity hit is the point: for some users the risk is not a bad answer but losing the tool mid-project.
@IterIntellectus [Claude Code]
https://x.com/IterIntellectus/status/2103884588414316926
A viral family story: this user showed their dad how to use Codex and Claude Code for work, and the dad called back in tears because five minutes of work had replaced what took one of the dad's employees an entire month. No workflow details, but it is the week's most emotional adoption datapoint and it came from outside tech.
@kenichiota0711 [Claude Code]
https://x.com/kenichiota0711/status/2103842276468490645
A Japanese educator tried Claude Slides: slides generated with Opus 5.5 can be edited inside Claude Code, and the UI is clean and easy to follow. Presentations becoming editable code artifacts is the same move the video wave made, applied to decks.
@priymrj [Claude Code]
https://x.com/priymrj/status/2103680253004779606
A creator gave raw 4K 60fps footage from a DJI Osmo Pocket 4 to stock Claude Code on an M1 Pro MacBook, no connector, only HyperFrames installed, and let it edit. The interesting shift is from generating footage to cutting real footage, where the agent has to work with what the camera actually captured.
@charliejhills [Claude Code]
https://x.com/charliejhills/status/2103893708550914076
A creator produced 16 motion graphics in Claude Code with a three-step method: pick one motion such as a morphing chart, show Claude a reference and name every state, then ask for HTML and SVG and fix it round by round. Naming every state is the key trick, because it turns a vague animation into a spec the agent can check.
@kurtinc [Claude Code]
https://x.com/kurtinc/status/2103834982678466877
A heavy user's current agent-maxxing setup: a custom launcher that boots Claude Code, Codex or Grok Build from one menu, a monorepo in Google's OKF format so every agent reads the same docs, and Claude 20x, ChatGPT 5x and SuperGrok subscriptions. The only thing slowing them down is weekly limits.
🗣 User Voice
User Voice
Quota is the switching trigger, and it is flowing toward Claude this week. A $200 Codex plan plus $500 in auto-recharge that still ran out in days pushed one developer to Claude Code (@DrSerunjogiEmma), and a 9-hour overnight loop costing 25% of a plan became the week's reference number (@ashen_one). The flip side exists too: 70% of a week's quota gone in two days sends others to DeepSeek (@KylianEcom).
Account safety is still the biggest fear for Chinese users. Bans with no clear cause (@coder_left, @leeoxiang), outputs that reveal the user is in Shanghai (@Xudong07452910), and folk rules about clean lines and time zones (@BruceDu521) show that people want transparent, predictable rules more than they want new features.
Safeguards need to fail loudly and rarely. Some users see normal work blocked by the new classifiers while others see nothing (@TokenGremlin), and a secrets nudge that refuses a pasted key got a tool built around it (@mankaff_). People are asking for a vanilla mode with no skills, plugins or MCP at all (@alymoursy).
Unattended runs need a real safety net. The 48,218-file deletion showed that permission checks do not see inside scripts the agent writes for itself (@Pirat_Nation), and losing three hours because a VS Code session could not be resumed from a phone (@kiyoshi_shin) is the same demand from the other side: control from anywhere, and undo that actually works.
Users increasingly credit the harness, not the model. Sixteen parallel threads versus three to five in another tool (@rdylina) and teams that still pin an older model for stability (@fumokmm) both point to the same conclusion: the orchestration layer is now what people choose on.
Quota is the switching trigger, and it is flowing toward Claude this week. A $200 Codex plan plus $500 in auto-recharge that still ran out in days pushed one developer to Claude Code (@DrSerunjogiEmma), and a 9-hour overnight loop costing 25% of a plan became the week's reference number (@ashen_one). The flip side exists too: 70% of a week's quota gone in two days sends others to DeepSeek (@KylianEcom).
Account safety is still the biggest fear for Chinese users. Bans with no clear cause (@coder_left, @leeoxiang), outputs that reveal the user is in Shanghai (@Xudong07452910), and folk rules about clean lines and time zones (@BruceDu521) show that people want transparent, predictable rules more than they want new features.
Safeguards need to fail loudly and rarely. Some users see normal work blocked by the new classifiers while others see nothing (@TokenGremlin), and a secrets nudge that refuses a pasted key got a tool built around it (@mankaff_). People are asking for a vanilla mode with no skills, plugins or MCP at all (@alymoursy).
Unattended runs need a real safety net. The 48,218-file deletion showed that permission checks do not see inside scripts the agent writes for itself (@Pirat_Nation), and losing three hours because a VS Code session could not be resumed from a phone (@kiyoshi_shin) is the same demand from the other side: control from anywhere, and undo that actually works.
Users increasingly credit the harness, not the model. Sixteen parallel threads versus three to five in another tool (@rdylina) and teams that still pin an older model for stability (@fumokmm) both point to the same conclusion: the orchestration layer is now what people choose on.
📡 Eco Products Radar
Eco Products Radar
Codex (128 mentions): The default comparison and the most common partner: many split planning, review and execution between Codex and Claude Code
OpenClaw (97 mentions): Now the substrate under Microsoft Autopilot and the design ancestor of Meta's Muse
Cursor (51 mentions): Mostly the tool people say they left, or the IDE they stopped opening
Muse (45 mentions): Meta's consumer agent, increasingly used as a coordinator over coding agents
Jev (35 mentions): Cheap typed decision model used as a gate, router and context filter around Claude Code
Hermes (22 mentions): Alternative agent harness users move between and OpenClaw
OpenCode (20 mentions): Open-source coding agent listed alongside Claude Code in most multi-agent tools
GPT-6 Astra (18 mentions): Used as planner and reviewer around Opus 5.5 execution
Autopilot (13 mentions): Microsoft's persistent agent built on OpenClaw, with fixes contributed upstream
DeepSeek / Kimi (11 each mentions): Cheap headless workers and fallback models when quota runs out
Docker Sandboxes (7 mentions): Cloud microVMs that keep agents running after the laptop closes
Obsidian (5 mentions): Home of second-brain setups driven by Claude Code
Hindsight (4 mentions): Open-source agent memory engine plugged into Claude Code
HyperFrames / Remotion (3 each mentions): Code-rendered video pipelines behind the promo and showreel wave
Paperclip (3 mentions): Open-source org chart and budget layer for teams of agents
Codex (128 mentions): The default comparison and the most common partner: many split planning, review and execution between Codex and Claude Code
OpenClaw (97 mentions): Now the substrate under Microsoft Autopilot and the design ancestor of Meta's Muse
Cursor (51 mentions): Mostly the tool people say they left, or the IDE they stopped opening
Muse (45 mentions): Meta's consumer agent, increasingly used as a coordinator over coding agents
Jev (35 mentions): Cheap typed decision model used as a gate, router and context filter around Claude Code
Hermes (22 mentions): Alternative agent harness users move between and OpenClaw
OpenCode (20 mentions): Open-source coding agent listed alongside Claude Code in most multi-agent tools
GPT-6 Astra (18 mentions): Used as planner and reviewer around Opus 5.5 execution
Autopilot (13 mentions): Microsoft's persistent agent built on OpenClaw, with fixes contributed upstream
DeepSeek / Kimi (11 each mentions): Cheap headless workers and fallback models when quota runs out
Docker Sandboxes (7 mentions): Cloud microVMs that keep agents running after the laptop closes
Obsidian (5 mentions): Home of second-brain setups driven by Claude Code
Hindsight (4 mentions): Open-source agent memory engine plugged into Claude Code
HyperFrames / Remotion (3 each mentions): Code-rendered video pipelines behind the promo and showreel wave
Paperclip (3 mentions): Open-source org chart and budget layer for teams of agents
Comments