Super User Daily: 2026-11-07
One person looked at why a terminal lagged on an 8 GB VPS and found a single Chromium holding 2.3 GB for 1 day and 16 hours, while the CPU sat at 0.5 percent pressure. That's today's mood: the best posts open the box and read the number. Noah Kagan's Slack clone reports 45 percent accuracy and still saves about 5 hours a week. A study found the same CLAUDE.md rule obeyed 72 percent of the time as a plain statement and 17 percent as a prohibition. A video editor needed five review rounds because Whisper's word timestamps end up to 0.6 seconds early. Non-coding work keeps spreading too: a family calendar on an old tablet, eight toddler games on a dead iPhone 8 Plus, a company owner who can't program running sales and timecards. Cat Wu's move to OpenAI and the subagent status-line fix filled most of the timeline, and the fix is worth reading closely, because 99 percent on one line works out to about 5 percent on a session.
@nicolasrouanne [Claude Code]
https://x.com/nicolasrouanne/status/2118229559166300235
Moved Claude Code sessions from GCP at about $126 a month to an OVH VPS at 13.30 euros with half the RAM, then one morning typing lagged. Instead of guessing, the post reads the kernel: 111 MB free of 7.6 GB, no swap, processes stalled on memory 35 percent of the time and on disk 88 percent, CPU 0.5 percent. Measured by PSS, a session costs about 260 MB plus about 180 MB of MCP servers, and every session starts its own copy, so 12 sessions meant 12 chrome-devtools-mcp processes. The first browser tool call starts a Chromium that stays until the session ends, 2.3 GB here after 1 day and 16 hours. Cheap hosting works, but budget RAM per session and kill idle browsers.
@noahkagan [Claude Code]
https://x.com/noahkagan/status/2118342871686197321
Noah Kagan built an 'auto Noah' that reads Slack, sorts messages into Needs you and FYI, and drafted 34 replies in a day. The send button stays human, and every edit carries a box explaining why, which gets saved as a rule. The best moment is the audit: asked what the account had sent since 12:45, it listed 6 messages, flagged that only 2 came from the page, and stopped testing until someone checked. It's 45 percent accurate and still saves about 5 hours of Slack a week. That number is the honest pitch for drafting agents, since they pay off well before anyone would trust them to send.
@GrabowskiDylan [Claude Code]
https://x.com/GrabowskiDylan/status/2118426889941614875
Dylan Grabowski trained Claude Code to cut a 31-minute recording into a 15-minute video, and it took 5 rounds of review. Whisper's word timestamps ended 0.14 to 0.60 seconds early and clipped the ends of words, one edit landed on the wrong take 83 seconds away, and a 0.13 second scrap of the word 'And' kept surviving. The fixes were engineering, not prompting: measure where words end from the audio itself, add a gate that checks every cut, and allow 3 re-renders before failing. Timestamped notes on each round now tell the author where to look. Long-form is about 90 percent there, and the shorts are barely passable by the author's own account.
@doooyle [Claude Code]
https://x.com/doooyle/status/2118367488970793117
ConferenceDatabase started in 2023 and got basically zero visitors for years. In September the owner handed it to Claude Code with little guidance, and it fixed the scrapers, rebuilt the site, and now brings 100 times the traffic from Google and AI search. About once a week the owner checks in while it runs its own experiments, 24 in 3 weeks across offers, popups and pricing: it proposes, the owner says yes, it builds, measures and kills the losers. Revenue so far is exactly $0, which the post says plainly. Dead projects can get traffic back cheaply, and traffic still isn't a business.
@dkp_consult [Claude Code]
https://x.com/dkp_consult/status/2118419921591816301
A parent saw a 3-year-old tapping on a dead iPhone 8 Plus as if it worked, and two days later it was the kid's console. The first idea, writing a mini OS, was dropped for something simpler: a native app with the phone supervised in single-app mode, so it boots straight into the games. Eight mini-games around animals, colors and counting to 10, with no text to read, no way to fail, no ads, no internet and no subscription. The spec was written like one for a client, then 8 parallel Claude Code jobs ran on a shared base, and 8 Xcode builds at once maxed the Mac's CPU, so next time it's 3 or 4. The client's verdict: more, please.
@minorivoice [Claude Code]
https://x.com/minorivoice/status/2118176681483493846
A company owner who says they can't write a line of code, with only AI as staff, runs sales management, barcode inspection and timecards built with Claude Code. Files were the real mess: the first audit found same-named folders on the PC and Drive with one empty, and key notes that had never been backed up despite the owner's belief they were syncing. The fix is 4 fixed locations, an inbox for anything unclear, and a Friday check run by a script. The owner now does two things only, saying 'I'm done' and asking where a file is. One hard-won detail: never keep the Obsidian vault inside Drive, because Drive sync and GitHub touching the same file can corrupt it.
@roeishous [Claude Code]
https://x.com/roeishous/status/2118265000540168199
A solo founder of two startups published the actual counts: 108,678 shell commands from agents in 30 days, 4,959 messages typed, half under 40 characters, and the most typed word apart from /compact was 'continue'. On one Wednesday the founder switched between repos 86 times. The system holding it together is one place per company where decisions live, with an AGENTS.md pointing at which doc wins on conflict, plus 5 scheduled Claude Code routines that write to Notion and send nothing automatically. So much for the autonomous company, the post admits. Honest usage logs like this beat any screenshot of a green dashboard.
@FavourYusuf1 [Claude Code]
https://x.com/FavourYusuf1/status/2118360436840059339
A freelancer turned a $200 Upwork job into $1,800 a month by researching the client before writing a word. Claude Code went through everything public about a former investment bank VP: 3,430 site pages with about 100 on Google, 16 YouTube videos and 4 subscribers. The judgment was what came next, because the Google finding looked the most impressive but wouldn't produce a user in 90 days, so it got cut. The proposal asked one question instead: why chase strangers on TikTok before the people who already trust you? The author's close rate went from 1 in 15 to 20 proposals to 1 in 3 clients spoken to, though the post also sells an app built on the workflow.
@RjeyTech [Claude Code]
https://x.com/RjeyTech/status/2118413707218727174
A household wanted a digital family calendar that syncs with Google Calendar, and the product in question, Skylight or something like it, ran around $500. About five hours of tinkering with Claude Code and an old Android tablet later, the living room had one with the design and features they actually wanted. The post pulled 206,000 views, far more than the prompt threads around it. People respond to the math: a drawer tablet and an afternoon against a $500 device. Consumer gadgets that are mostly a screen plus a sync are in real trouble.
@AIC_Hugo [Claude Code]
https://x.com/AIC_Hugo/status/2118190236752867432
A rare honest shutdown: Space, a shared workspace for docs, tasks and memory reachable from Claude Code, Codex and Hermes over MCP, closed as a business after 2 months of building, 17 days public and 0 paying users. Notion shipped the same pitch in versions 3.7 and 3.8, aimed at companies already living in Notion. The founder also changed the positioning three times in two weeks. The repo stays MIT and still runs the founder's own agents. The lesson is the useful part: the only real edge was being the first user, and small personal tools pay off on that edge, while category plays against Notion and model vendors don't.
@ai_database [Claude Code]
https://x.com/ai_database/status/2118175971815870514
If Claude Code keeps ignoring your rules, researchers from NAIST, Wakayama and Osaka tested whether phrasing matters, and it does. Compliance was 72 percent for a plain statement, 49 percent with a reason attached, 36 percent for a conditional, and 17 percent for a prohibition, even when conditionals were counted only where the condition applied. The experiment ran on Sonnet 5, so don't assume the exact numbers carry over. Still, it's a five-minute fix to test: rewrite your 'if X' and 'do not' lines as direct statements. Explaining why, which most guides recommend, came out worse here.
@mathieuhq [Claude Code]
https://x.com/mathieuhq/status/2118269291791290553
A French developer stopped approving every command and runs Claude Code on bypass permissions, and the safety comes from the structure around it. Each task starts as a GitHub issue, scoped through interview, spec and ticket skills, and gets the label agent-ready only when it can run without coming back to a human. A runner on a VPS picks up that label, creates an isolated worktree, codes, tests and opens a PR, day or night. The agent runs on a dev server with no merge rights on main, and a human reads, tests and merges at the end. The job becomes writing tickets precise enough that nobody gets interrupted.
@ainponce [Claude Code]
https://x.com/ainponce/status/2118349462158348316
A concrete security hardening recipe using ISO 27001, SOC 2 and ISO 42001 as the base, done on the author's own app. Map the system and what sensitive data each part touches, build a matrix of the technical controls that apply, search the code for evidence and mark gaps, then fix each gap in a PR that cites its control. The agent gets hardened too: no access to secrets, a sandbox, and human review of every change. The post is careful to say this doesn't certify anything, because many controls are organizational and only an accredited auditor certifies. Citing the control in each PR is the piece worth stealing.
@Edesberg [Claude Code]
https://x.com/Edesberg/status/2118339489382686842
A team says 4 times more of its PRs pass first review since adding one skill at the end of every coding session, and has now open sourced it as make-native. The diagnosis is familiar: agents love writing new code and are bad at making it look like it belongs, so you get a flag here, an if (isNew) there, and a copied component 'just for this case'. The skill finds the bolted-on code, finds the helpers and types that already do the job, rewrites the feature to use them and deletes the rest. The 4x is their own number. Spending the last step of a session on removal rather than addition is still a sound habit.
@elliotarledge [Claude Code]
https://x.com/elliotarledge/status/2118386115170496855
Elliot Arledge built phonon, a free, open source dictation app for Mac that runs entirely on the laptop, with Claude Opus 5.5 doing the building. Key-up latency is 65 ms on an M4 Max against about 520 ms for Wispr Flow on the same clips, and a 6-minute session lands in about a quarter second. Training used 77,000 transcripts rewritten by DeepSeek to a written spec of what 'what you meant' means, none from the author's own dictation, and a blind judge preferred its text 90 times to 23. It adapts by app, casual in Discord, formal in email, clear prompts in Claude Code. Voice input keeps coming up because talking is how people steer agents now.
@zawa_c_i [Claude Code]
https://x.com/zawa_c_i/status/2118174911990817001
A Japanese builder made a Mac voice input app with Claude Code, and the write-up is about process more than features. Voice stays on the Mac through Apple's Speech framework, and a glossary built from past Claude Code sessions handles fine Japanese terms that big apps kept getting wrong. The nine-step flow starts by writing what it won't do, defines data tables, states and cost caps, and has a different AI, Codex, review the design before Claude Code implements in small tested slices. One nice safety touch: before pasting, it checks the target window and field, and on any mismatch it refuses to auto-paste. Treating a personal tool like a real product is why it holds up.
@dotey [Claude Code]
https://x.com/dotey/status/2118164405594595702
Baoyu's summary of an AI Engineer talk by Addy Osmani gives the most useful numbers on review at scale: Anthropic went from about 2,500 merged changes a day in May to over 30,000 by September. Nobody reads that line by line, so Claude first uses the feature like a real user, with end-to-end tests and browser checks, hooks trigger fixes, and team taste lives in shared CLAUDE.md and AGENTS.md files. Humans then look at evidence, such as a screen recording of it clicking through or screenshots per device. The same talk cites CircleCI data: output up 59 percent while 3 of every 10 merges to production fail. Verification is where the next speedup has to come from.
@chasing_next [Claude Code]
https://x.com/chasing_next/status/2118420116471963655
A simple pattern for making Claude Code proactive without a chat app: scheduled jobs, short scripts that format results, and the Resend API to email them. Daily site performance, GTM leads and to-dos, tailored digests, weekly SEO numbers, and security audit summaries all land in the inbox with charts and color coding. Each email can kick off a follow-up skill, and it leaves the inbox when finished. The author prefers this to Grok Bot's chat threads because it's easier to control the output and avoid drift. An inbox already has states for read and done, which most agent UIs still don't.
@hlibdoesgtm [Claude Code]
https://x.com/hlibdoesgtm/status/2118264133074194603
An outbound agency published its whole operating system as about two dozen Claude Code skills, and every one is a markdown file in a folder. Client onboarding creates the Slack welcome, research, intake form and ClickUp project in one command, a list skill spends paid enrichment only on rows that pass filters, and a launch skill sets up domains and mailboxes. After launch, /campaign-guard checks live campaigns daily and flags reply rates under 1 percent, and a weekly report posts every client's numbers on Friday. Whatever you think of cold outreach, this is what an agency looks like when its playbook is executable. The moat is now a folder, so it can walk out with an employee.
@nickel0 [Claude Code]
https://x.com/nickel0/status/2118213110133924072
A Japanese engineer connected Claude Code to MATLAB over MCP for the first time, wanting to learn satellite phased arrays. Given a simple simulation premise of 500 km altitude and n256 FDD, it produced an easy personal teaching material, with results collected in Claude Artifacts. The post got 920 likes, which says something about who wants this. Pairing a coding agent with real numerical software turns a textbook topic into something you can poke at. It's self-study with a lab attached, and that used to need a professor or a lot of patience.
@dontbesilent [Claude Code]
https://x.com/dontbesilent/status/2118256527857561976
A small, sharp consumer use: buying car insurance, someone screenshotted every salesperson's chat and had Claude Code find the best deal. The more interesting part is the follow-up thought from the seller's side. If the customer can't use AI, quote high; if they can, show your lowest number right away; but how would you know which one you're facing? Price comparison used to be limited by patience, and sales scripts were built around that limit. That assumption is going to get expensive for anyone relying on it.
@yaohongan2025 [Claude Code]
https://x.com/yaohongan2025/status/2118327637227479145
A Chinese indie builder told Claude Code that SEO was too complex and asked to skip straight to getting recommended by ChatGPT. It answered that you can skip learning SEO but not doing it, and used the builder's own data to argue the point: 26 visitors from ChatGPT against 8 from Google on one site, with ChatGPT visitors the most willing to pay. It then wrote a division of labor where the agents handle indexing, submissions and answer pages while the human thinks about what users will ask AI. Being argued out of a shortcut by your own numbers is a good use of an agent. Small samples, but the direction matches what others posted today.
@qianshuixia2026 [Claude Code]
https://x.com/qianshuixia2026/status/2118148917842071786
Tired of asking AI by hand, a builder wired Claude Code to an SEO agent as a skill and ran a diagnosis with real data behind it. It found that 90 percent of a competitor's traffic comes from ChatGPT, which triggered a quick move to GEO, and used an established old site to link to a new one. The post's real claim is that AI with a data foundation behaves like a different product from AI without one. Same model, connected to rankings and traffic, stops giving generic SEO advice. That's the MCP argument in one small example.
@OkaTakuma1 [Claude Code]
https://x.com/OkaTakuma1/status/2118297289165058111
A Japanese site owner used to spend about 40 minutes publishing a blog post and now mostly hands it off to Claude Code with the Chrome connection. Writing, image placement, SEO settings, structured data, the GitHub push, the Vercel deploy, and a check in Chrome are all done by the agent, and the human only reviews and says publish. The nice touch is recursive: the article about setup errors and test results was written and published by the same flow. Small publishing chores like this add up to the most reliable savings around. Nobody needs a benchmark to see 40 minutes turn into a review.
@ksasao [Claude Code]
https://x.com/ksasao/status/2118336350944645476
A seller on Japan's BOOTH marketplace automated the shipping routine with Claude Code: pulling the shipping list, buying Click Post labels, creating and printing receipts, and sending shipping notices. The post's punchline is that the human now just moves their body as instructed. That's a funny inversion and an accurate one, since physical packing is the only step left that needs hands. For small sellers the job isn't software at all, it's the clerical work around a package. It's also the kind of site-specific flow no SaaS vendor would ever build.
@8T3_nawaz [Claude Code]
https://x.com/8T3_nawaz/status/2118478744935395737
A motion designer used Claude Code to reframe a 30-second, full motion graphics After Effects piece from 16:9 to 9:16. It took five minutes and got about 90 percent of the way, on a job that would otherwise take all day. Versioning is exactly the type of work motion designers dislike and clients keep asking for. The last 10 percent still takes a human eye, but that's the part worth paying for. Expect format-versioning to quietly become a free line item.
@stayworkgh [Claude Code]
https://x.com/stayworkgh/status/2118141833125802240
A 47-year-old in Tokyo juggling immigration paperwork, guesthouse deals and dance practice has a new habit for company site security. Venice AI analyzes weak spots, Claude Code proposes fixes, and the fixes go back to Venice to be checked, so the two argue with each other. Both still state wrong things confidently, so neither gets trusted outright, and tracking down the errors has turned into the education. The author checks only sites the company manages. Using your own site as the textbook is a better way to learn security than any course, as long as you verify.
@HaStrive [Claude Code]
https://x.com/HaStrive/status/2118489479157834198
One of the most useful short posts today: someone grepped agent configs for fast mode and found two scripts that will hard-error in three weeks. Claude Code deprecated fast mode on Opus 4.8 on October 30 and removes it November 30, and per the changelog those requests return an error with no fallback to standard Opus 4.8. So the failure is a dead turn, not a slower one. The plan is simple: grep settings, CI and harness scripts, repoint jobs to Opus 5.5, and test by November 29. If you've got unattended jobs pinned to a model, now's the time to look.
@wy_ones [Claude Code]
https://x.com/wy_ones/status/2118137909949944249
A clear rule set for running Claude Code on two machines with Mutagen: sync memory/ and CLAUDE.md so both agents remember the same things, and sync skills so an install on one shows up on the other. Don't sync conversations, since a live session should finish where it started. Don't sync login credentials, because one machine refreshing its login kicks out the other. And never let either side auto-win a conflict, since both write to memory at once and an overwrite loses notes. Deciding what not to sync is the whole design. Most people only find these rules by losing something.
@AgentNaeem [Claude Code]
https://x.com/AgentNaeem/status/2118401980603232383
A Notion power user described a handoff that actually fits how people hit limits. Research a feature in ChatGPT and save the decisions and the reasons to a Notion page, have Codex pick up from there and build, and when the usage limit hits, switch to Claude Code and read the same notes instead of waiting for a reset. Plan changes go into the same place for whichever agent is next. It's vendor-neutral memory built on a tool the person already has. Usage limits turning into a reason for shared context is a neat side effect nobody planned.
@GlobalFeedAI [Claude Code]
https://x.com/GlobalFeedAI/status/2118163651354132710
Before you ask your agent to set up a GitHub MCP, check which server it installs. Glow Labs reviewed real coding sessions and found agents installing MCP servers from an archive with no updates since May 2025, and in 98 percent of those sessions the agent appears to have filled in the package name from training data. It showed up across Claude Code, Cursor, Codex CLI and others, and those packages have 17.35 million downloads since July 2025, with high-severity flaws that won't be fixed, though Glow hasn't seen them exploited. The fix takes minutes: give the agent the full name github/github-mcp-server, and search your MCP configs for the old names. Stale training data is a supply chain problem now.
@xiaobuzaowu [Claude Code]
https://x.com/xiaobuzaowu/status/2118137954736779572
A team using Codex on its website wrote up the Postman config incident with unusual honesty. They'd already tightened Codex to its working folder on November 2, added a rule on November 4 that it must explain what make, npm or scripts will change before running, and removed plaintext credentials from a Mac config. They also admit they hadn't finished checking their own machines for Postman's additions when they published. The self-check list covers ~/.codex/config.toml, hooks, ~/.local/bin and Claude Code and Cursor settings. Their summary is good: an agent's config file deserves the attention you'd give a password.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2118311658867097991
Pawel Huryn got tired of everyone discussing what a harness is and published Mini Harness, about 1,200 lines of Python with /goal mode, permissions plus auto mode, skills and MCP servers, down to server-side compaction. The goal was teaching, since other minimal harnesses skip those features. The surprise, by the author's account, is that it became the top harness on Claude Haiku. A harness you can read in one sitting is the best way to stop treating this layer as magic. If 1,200 lines can compete, a lot of the complexity elsewhere deserves questions.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2118350138347544618
Papermorph is a skill for Claude Code that takes a textbook PDF and produces an interactive narrated web course, with formulas and figures changing step by step and quizzes in each chapter. The author's linear algebra example runs 40 chapters, with the planes of a system of equations draggable in 3D. It also does picture books: one classic Chinese children's story stops the animation so a kid can tap what to look for, and a wrong tap isn't called wrong. Output is static pages with no server, and each chapter can export a 1080p video. Parents and teachers are becoming a serious audience for coding agents.
@NixFred [Claude Code]
https://x.com/NixFred/status/2118142508207099976
One power user runs about 60 Grok bots with exactly one job each, and Claude Code sits underneath them as the agent doing machine work. Chief of Staff keeps machines and connectors healthy, Larry Manager hands work to a Claude Code agent on each machine, Homelab Health only alerts, and The Phredd Times prints a newspaper overnight. They all check in at one page, and anything that spends money or can't be undone waits for a yes. The stated trick wasn't more bots, it was one job per bot and a clear owner for everything. That's org design, and it transfers to any agent stack.
@mushengziyue [OpenClaw]
https://x.com/mushengziyue/status/2118158262914940965
A secondhand but well-told OpenClaw story: Robby, an operations employee who isn't a programmer, gave an agent named Ron $100 and a $20,000 goal. Selling SWOT analyses on Fiverr went nowhere on a new account. The turn came from TikTok comments, where it read about 200 and saw people asking for an agent of their own, so it pivoted to a hosted AI companion community, presold at $10 to about 600 people, with about 270 staying at $29 a month. That's $8,374 in monthly recurring revenue after 13 days. It's 13 days and it's reported by someone else, so hold it loosely, but reading what people ask for beat picking a business from a list.
@superdoccimo [OpenClaw]
https://x.com/superdoccimo/status/2118268233140646378
A Japanese builder made a tool called Sakura Bridge and drove a home PC from the free Claude plan on Sonnet 5.5. It listed connected devices, saw a shared Windows screen, and went through Windows to SSH into another Linux box, with no paid plan or API billing. Claude got as far as preparing to update a remote OpenClaw, then ran out of free quota because its explanations ran long, so the update itself isn't confirmed. The author's next idea is to do the update by hand and have Claude do the checks, log reading and verification. The failure is the funniest part: chatty explanation is what used up the budget.
@lucsucces [OpenClaw]
https://x.com/lucsucces/status/2118470917722620304
A short note on why personal agents stick. This user started on OpenClaw, switched to Grok Bot a couple of months ago, and now says Muse and Dots look interesting but setting everything up again isn't worth it, since Grok runs a whole life at this point. The moat here is accumulated setup, not model quality. It also explains why the first switch happened at all: early on, there wasn't much to lose. For anyone building in this category, the window to be someone's first serious agent is the thing to fight for.
@BrandonJourney2 [OpenClaw]
https://x.com/BrandonJourney2/status/2118299232624836993
A complaint from an everyday OpenClaw user: half the time goes to telling an AI to fix OpenClaw, because it keeps losing connection to models. Another user who's spent a month on Grok Bot, Muse and Dots says frontier harnesses break a lot too, and the only efficient repair is asking them to fix themselves. So the maintenance burden isn't special to open source. 'Ask the agent to repair the agent' has quietly become the standard support channel, which works until the thing that's broken is the agent's own connection.
@dusangran [OpenClaw]
https://x.com/dusangran/status/2118426120529773002
A one-year tool diary that says a lot: IntelliJ to Zed a year ago, then four months ago Zed to Multica, then Hermes, and now OpenClaw. The author asks what's next and notes the monthly Anthropic bill is the only constant. Switching costs in agent tools are near zero right now when the model underneath stays the same. That's great for users and rough for anyone selling a harness. Harness vendors are competing for a seat while the model vendor collects the rent.
@lucianlamp [Claude Code]
https://x.com/lucianlamp/status/2118251858997514672
A Japanese user turned Muse's cloud PC into a small studio, installing Claude Code, Codex, Grok Build and Antigravity on it and telling Muse what each one is good at. Blender is on there too, with an external GPU connected, so Muse can hand 3D work around. It's a personal agent acting as a dispatcher over coding agents, which is a pattern showing up more each week. Nobody here is picking a winner among the four. They're being treated like contractors with specialties.
@yamachan_ai_log [Claude Code]
https://x.com/yamachan_ai_log/status/2118292005398921334
A small but transferable finding from someone testing Claude Code and Codex on business model diagrams: asked to generate complex diagrams from a text description into HTML slides or PowerPoint, the best result so far comes from having the agent draw icons and arrows as SVG. Anyone who has fought AI-made slides will recognize the issue. Icons and connectors are usually where it falls apart, and vector output keeps them crisp and editable. The post is a modest test note, and that's why it's believable.
@ai_am_furufuru [Claude Code]
https://x.com/ai_am_furufuru/status/2118308166207459767
A Japanese creator made a video titled 'How to get rid of packed trains' using only Claude Code on Opus 5.5, with the Pro plan at 3,500 yen as the entire cost. No video generation AI was used, which is what surprised the author most. Motion built from code keeps replacing generated footage for explainer-style content. It's cheaper and it can be edited precisely. This is a useful reminder that a coding agent is also a video tool when the output is a program that renders frames.
@kgsi [Claude Code]
https://x.com/kgsi/status/2118147692950159496
Canary, a Japanese company, runs PRD through to design-system UI generation with Claude Code, and the write-up names a specific failure. With only DESIGN.md, every generation decided how to treat each component on the spot, so precision wobbled. The fix was adding a list of component types and an index of when to use each. That's a useful correction to the idea that one design file solves it. Agents need the decision logic spelled out, not just the look.
@gorisan_it [Claude Code]
https://x.com/gorisan_it/status/2118342590521389170
An AI diary posted under Sumida Ward activity-report tags offers a neat habit for a CLAUDE.md that's getting fat. After big changes the instruction is 'optimize the whole thing, show me the improvements first, and fix after I confirm', so nothing gets rewritten blindly. The post is tiny, but the order is the point. A self-rewriting instruction file is exactly where you don't want a silent change. It also shows how far this has spread, when a local activity report mentions context file maintenance.
@RennieHaylock [Claude Code]
https://x.com/RennieHaylock/status/2118398399087300742
A builder pushed back on 'SaaS is dead' with a real result: a World Cup app built in a month with Claude Code, reaching 28,000 users and $12,000 in revenue, with 10,000 concurrent users during the final. Event-driven apps are a great fit for agent building, because the deadline is fixed and the window for demand is short. Before, that window often closed while you were still hiring. The author reads this as a reason to be more bullish on bootstrapped software. A single event isn't recurring revenue, but the speed to catch one is new.
@guoda2066 [Claude Code]
https://x.com/guoda2066/status/2118227583523307882
A Chinese developer is pre-releasing Remmo, a phone app for following and continuing desktop agent conversations across Claude Code, Codex, Cursor CLI and Antigravity CLI. There were two reasons to build it: the official apps organize projects and sessions poorly, and most alternatives ship with a heavy desktop app when the developer prefers the official one. It supports multiple Macs and is in daily personal use, with Apple's review still pending, so 30 TestFlight slots are open. Remote agent control keeps spawning small products because no official app feels finished yet. Being its own daily user gives this one a decent shot.
@Gorden_Sun [Claude Code]
https://x.com/Gorden_Sun/status/2118292548561371463
Tsinghua and Berkeley researchers built uni-robot-agent, using Codex and Claude Code as the runtime so open models can control robots through a standard toolset for seeing, moving and grasping. Binding camera frames to live arm state and letting the model think while moving took Qwen3.8-Flash-Next from 1.90 to 31.39 in RoboDojo, past GPT-6 Astra's 28.97 on the old setup. On 19 real dual-arm tasks it scored 53.50, ahead of a tuned robot model. Cost per task was $1.09 against $21.29. Same model, far better score with a better harness, now in robotics too.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2118425535667679410
Someone fed a full ChatGPT export, 3 years and 2,800 chats, into a Claude Code wiki built on Karpathy's LLM Wiki idea, and woke up to 412 pages. The page about the author was the hardest to read: all 3 side projects stopped the week money had to be asked for, sleep questions got asked 14 times, and a Spanish study plan got requested 9 times, always in January. The post is promotional in tone, so treat the details as the author's. The idea stands anyway. Your chat history is the most honest diary you've kept, and nobody has read it yet.
@lucifer_x007 [Claude Code]
https://x.com/lucifer_x007/status/2118383141832540328
A small handoff test: Claude Code explored a project, identified bugs and wrote a handoff, then Codex read it, fixed the bugs and ran the tests, with no manual context transfer. The tool in the middle is Fulcra Dynamics, and the post reads partly as a recommendation for it. The complaint behind it is what matters, since having to repeat context across coding agents is one of the most repeated gripes this month. A written handoff file gets you most of this on any stack. What people want is for that step to stop being their job.
🗣 User Voice
User Voice
Announced savings need arithmetic. Status lines now cost about 99 percent less, and @RadineerE10 and @ai_fengshou both point out that works out to about 5 percent on a session using background subagents. @nikolayxyz cites a 4-hour session where 339 of 1,496 API calls were this one line, and @BoboLinNotes estimates about $4.80 an hour at a 200k-token transcript. @rahulmanuwas spent a $250 cloud credit on 7 routines and 6 draft PRs, then found the credit never moved and the docs didn't explain which sessions it pays for.
Computer use is where people are comparing and complaining. @Yuchenj_UW calls it Codex's biggest advantage, having watched it finish an expense app in minutes, while @plusyip says Claude Code keeps stealing the cursor. @yasser_elsaid_ wants browser use in an SSH session and working remote control, and @FAatc168 finds raw mobile SSH close to useless for pasting context.
Long unattended runs still stall. @vinrambone goes to bed and wakes to find Claude Code stuck on the first iteration of a several-hour shell process, with no status check or timeout. @crackeddl asks why there's still no simple benchmark running the same tasks and model across harnesses.
Teams want to share a thread, not trade summaries. @ryanflorence wants every thread to be a chat room where two people and the agent can all steer, and asks if it exists. @unbosoms says small teams end up throwing individual Claude Code conclusions at each other, and arguing with it together in Slack would be faster.
People are starting to check the claims. @BruhWoofer and @thesimaximalist questioned the viral hidden-hardware stories in the replies, and @scanningwithsam wrote a check on the deleted-photos versions. @CognitionWire did the same for Postman: the auto-install and self-approved hook are in the GitHub issue, and data leaving the machine isn't. @ohayou792 reports the game-name rule was loosened in 2.1.317 after days of complaints.
Announced savings need arithmetic. Status lines now cost about 99 percent less, and @RadineerE10 and @ai_fengshou both point out that works out to about 5 percent on a session using background subagents. @nikolayxyz cites a 4-hour session where 339 of 1,496 API calls were this one line, and @BoboLinNotes estimates about $4.80 an hour at a 200k-token transcript. @rahulmanuwas spent a $250 cloud credit on 7 routines and 6 draft PRs, then found the credit never moved and the docs didn't explain which sessions it pays for.
Computer use is where people are comparing and complaining. @Yuchenj_UW calls it Codex's biggest advantage, having watched it finish an expense app in minutes, while @plusyip says Claude Code keeps stealing the cursor. @yasser_elsaid_ wants browser use in an SSH session and working remote control, and @FAatc168 finds raw mobile SSH close to useless for pasting context.
Long unattended runs still stall. @vinrambone goes to bed and wakes to find Claude Code stuck on the first iteration of a several-hour shell process, with no status check or timeout. @crackeddl asks why there's still no simple benchmark running the same tasks and model across harnesses.
Teams want to share a thread, not trade summaries. @ryanflorence wants every thread to be a chat room where two people and the agent can all steer, and asks if it exists. @unbosoms says small teams end up throwing individual Claude Code conclusions at each other, and arguing with it together in Slack would be faster.
People are starting to check the claims. @BruhWoofer and @thesimaximalist questioned the viral hidden-hardware stories in the replies, and @scanningwithsam wrote a check on the deleted-photos versions. @CognitionWire did the same for Postman: the auto-install and self-approved hook are in the GitHub issue, and data leaving the machine isn't. @ohayou792 reports the game-name rule was loosened in 2.1.317 after days of complaints.
📡 Eco Products Radar
Eco Products Radar
Codex: the constant comparison and the second agent in most stacks, 142 mentions in the posts read
Cursor: still named next to both, 62 mentions
Grok Bot: the personal agent people run coding agents under, and the one they complain about limits on
Hermes: named with OpenClaw as the self-hosted choice
Notion: agent memory and skills launch, used as a shared handoff page
Muse: personal agent with a cloud PC, used as a dispatcher
Conductor: bought by Meta, cloud shuts down November 20
Antigravity: often installed alongside Claude Code and Codex
Postman: desktop app added plugins to coding agents without asking
Obsidian: the default vault for agent notes
OpenCode: named as a harness option in tool stacks
claude-mem: session memory, passing 110,000 stars
T3 Code: control surface over several agents and subscriptions
Blender: 3D work driven by agents over MCP
HyperFrames: video from HTML inside coding agents
Codex: the constant comparison and the second agent in most stacks, 142 mentions in the posts read
Cursor: still named next to both, 62 mentions
Grok Bot: the personal agent people run coding agents under, and the one they complain about limits on
Hermes: named with OpenClaw as the self-hosted choice
Notion: agent memory and skills launch, used as a shared handoff page
Muse: personal agent with a cloud PC, used as a dispatcher
Conductor: bought by Meta, cloud shuts down November 20
Antigravity: often installed alongside Claude Code and Codex
Postman: desktop app added plugins to coding agents without asking
Obsidian: the default vault for agent notes
OpenCode: named as a harness option in tool stacks
claude-mem: session memory, passing 110,000 stars
T3 Code: control surface over several agents and subscriptions
Blender: 3D work driven by agents over MCP
HyperFrames: video from HTML inside coding agents
Comments