Super User Daily: 2026-10-12
Saturday's Claude Code feed was about who the agent works for when nobody is coding. A family with no programming background gave a nonspeaking, quadriplegic man his first text message in ten years. A software engineer pointed Claude Code at 4.35 million pages of Dutch East India Company records and surfaced a meteorite, three rhinos and three eruptions that no catalog had. A Reddit post about a possible planet 116 light-years away got NASA's TESS telescope booked for October 31. On the other end of the ledger, a lawn-care owner is replacing QuickBooks and a $635-a-month scheduler with software that lets the business shop its own card processor, a product coach rebuilt a $700-a-month email stack down to $110, and a Ralph user says they now deliver $100K of paid development for $3K in Sonnet tokens. Claude Code Projects reached every Pro and Max user on the waitlist, and within hours the first complaint landed: Projects ignores your global skills and CLAUDE.md, defaults every thread to Opus on high, and one user watched Fable at max effort burn tokens on exploration a cheap model should have done. The cache-TTL argument ran all day, with three posts telling you to set subagents to one hour and one telling you that is how you lose 15% of your weekly limit. OpenClaw's side was a story drive: a pest-control owner who left the business to install it for 15 local firms, a venture studio running 20-plus agents with a meta-agent on top, a house sold, an apartment found, a check-engine light cleared, and a Ubuntu operator who traced a maintenance failure to the sandbox, not the model.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2108780815178223798
Ben has an ultra-rare condition that left him nonspeaking and quadriplegic, and for close to ten years his only channel was turning his head for yes or no. The family had no programming background. They described what Ben needed to Claude Code and got a keyboard and phrase boards, texting for the first time in his life, streaming and web search so he picks his own shows, and dozens of switch-operated games. Claude Haiku suggests replies that Ben selects, so he can keep up in a live conversation. Everything runs free and open source in the browser, and the first movie he chose for himself was Spider-Man.
@superdoccimo [Claude Code]
https://x.com/superdoccimo/status/2108806750606676064
A software engineer read a historian's post about finding a new dodo sighting in a 1615 ship log with Opus 5.5, decided several mysteries could be chased in parallel, and built a pipeline with Claude Code over roughly 4.35 million pages of Dutch East India Company records plus two centuries of newspapers. The design is a three-stage funnel: regex and embedding search to cast wide (17th-century Dutch has about fifteen spellings for rhinoceros), a cheap decision model to triage (59,000 elephant mentions for about three dollars), Claude Haiku to read and translate the survivors, then Claude Code opening the handwritten scans to check the transcription. Before running it for real, the author buried known records in the corpus, the 1783 Laki and 1815 Tambora eruptions among them, to prove the detector rings. Candidates so far: an 1812 meteorite fall near Pandharpur that none of six catalogs list, three Javan rhinos in captivity in the 1730s and 1740s, and three uncataloged eruptions. The CLI is published as Antiquity and has about 15 stars.
@parker_brydon [Claude Code]
https://x.com/parker_brydon/status/2108908084126978493
A DM forwarded by Geoffrey Huntley from someone who first dismissed Ralph loops as reckless in January, then watched Sonnet 5 loop for six hours and ship a product. After a stretch of experimenting, the writer says they now deliver paid development worth over $100K for about $3K in Sonnet tokens on brownfield projects, and replatform whole stacks in roughly two weeks that used to cost half a million dollars in contractors. The frustration that drove it was familiar: AI gets the first parts right, then starts cutting corners. The letter closes by thanking Huntley for publishing the Ralph write-up and calling the implications amazing and terrifying at once.
@NatureBlueee [Claude Code]
https://x.com/NatureBlueee/status/2109026067700973950
Problem 3 in a postmortem series from Flowness, a multi-agent system one person has run in production for six months with over 140 billion Claude tokens on record. On September 6 the same task was handed to nine background Claude Code executor sessions in a row; the automatic redispatch cap was 3, the retry counter reached 8, and the cap never fired. The eight failures were a mix of unmeetable acceptance conditions, a rejected commit, a misused command, a coordinating agent stopping a session by mistake, a full disk and one session that simply vanished. In at least three redispatches the new session received the exact task package that had just failed, and one acceptance threshold of 1,500 ms came from a cost model in a report, never measured; the real difference was 244 ms. The conclusion: every check answered 'can we dispatch now?' and none asked 'is this different from last time?'
@bogusbanks [Claude Code]
https://x.com/bogusbanks/status/2108926155210321966
A small-business owner is using Claude Code to write software that replaces QuickBooks, replaces Service Autopilot, and lets the company choose its own credit card processor. The scheduler cost $635 a month with a guaranteed minimum 10% annual increase, and it only offered card processing through one vendor taking 4% of top-line revenue; shopping processors freely is expected to bring that to about 2.2%. The post's larger claim is that building it yourself produces better software, because the vendors' programmers never used the product they were writing. QuickBooks Pro will not be renewed.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2108996010500657541
A product coach's guest ran 12,500 PM job descriptions through Claude and GPT and found AI skills in more than 30% of postings, with AI PM roles paying 15 to 20% more in the US and Europe and 30 to 50% more in India. The hands-on part is the HelloPM founder rebuilding an email platform that cost $550 to $700 a month on a stack never touched before, with Claude Code. The founder typed 'I'm sleeping now for about six hours' after midnight and woke to every slice built and tested. The first AWS stack Claude produced would have cost $420 a month; two questions about server size and Fargate brought it to $110. The takeaway offered: the AI writes the code, you pick what is worth building and what it should cost.
@NASAWatch [Claude Code]
https://x.com/NASAWatch/status/2108950287532589437
A Reddit post relayed by NASAWatch: a user went through NASA TESS telescope data with Claude Code and one star started looking strange. Every 3.18 days it dims about 0.05% for roughly two hours, first seen in late 2025 data, then found again in separate TESS observations from 2020 and 2018. If it is a planet it is about 1.4 times Earth's size, 116 light-years away, and still a candidate rather than a confirmation. Claude Code with Opus 5.5 and Fable 5.1 did most of the work: downloading and parsing the data, writing the search code, fitting transits, checking nearby stars, hunting for secondary eclipses and false positives, generating plots and rerunning when a test failed. A separate post adds that TESS will watch the star from October 31 to see if it dims on schedule.
@MMMusol [Claude Code]
https://x.com/MMMusol/status/2108820508741718287
ai-job-search is near 45,000 GitHub stars. Its author, geophysicist Mads Lorentzen, built it with Claude Code after a layoff and went from 69 tailored applications to 20 first-round interviews and a career switch into AI engineering. The framework splits the job hunt into search, fit scoring, résumé tailoring, cover letters, interview prep and application tracking, all local. The piece the post calls the key is a two-agent review: a Drafter writes the tailored résumé and letter, and a Reviewer starts in a fresh context, researches the company independently and critiques from a third-party view. Four commands cover the flow: /setup, /scrape, /apply and /interview.
@ArkStabler [OpenClaw]
https://x.com/ArkStabler/status/2108866502661193736
A pest-control company owner installed OpenClaw in the business in February, then left the company to a brother and started a new business installing it for other small firms. Fifteen local businesses in a small town, plus some outside it, run OpenClaw today. The reason given is blunt: once OpenClaw arrived, doing pest control was no longer possible.
@coreyh [OpenClaw]
https://x.com/coreyh/status/2108826485700104256
A venture fund and studio reports running a fleet of more than 20 OpenClaw agents across its portfolio companies, covering shipping software, go-to-market and finance. Sitting above the fleet is a meta-agent whose job is managing the other agents. The post came in response to a call for unusual OpenClaw stories and includes a screenshot of the setup.
@rileybrown [Claude Code]
https://x.com/rileybrown/status/2108899765320638935
Riley Brown's recipe for turning Claude Code Projects into a personal agent platform: create a project, name it Tony, and paste a prompt that makes it responsible for moving your life forward, analyzing email, Slack and every connected platform, and proposing threads and automations, with one standing thread called the Organizer that keeps an editable to-do document. Projects has a coordinator chat and threads that act as subagents, and those threads can run in the cloud or on local machines. Brown has a MacBook and two Mac minis on 24/7, so some threads run there while the coordinator stays in the cloud and is always reachable. The coordinator can be Opus 5.5 while routine threads run Haiku or Sonnet.
@Av1dlive [Claude Code]
https://x.com/Av1dlive/status/2108969431959568892
One prompt for Opus 5.5 in a local Claude Code session that reads your 100 most recent sessions under ~/.claude/projects and prepares a Claude Project from them. It groups sessions by project, uses only your original messages as evidence of preferences (not tool output or summaries), inspects commits where useful, picks the strongest recurring piece of work, and writes a name, goal and paste-ready project instructions. Repeated corrections become 'Remember:' messages, repo-specific guidance is split out for CLAUDE.md, every recommendation needs exact quotes and session references, and it lists the repos, skills and connectors the project needs while separating local tools from what cloud threads can reach. The author ran it on their own history and says it found the work they keep repeating.
@shimo4228 [Claude Code]
https://x.com/shimo4228/status/2109049512329429331
A first-day report on Claude Code Projects from someone who already orchestrates with Herdr: Projects does not read the global skills or CLAUDE.md on the local machine. The author had an explorer agent pinned to Sonnet 5.5 to match model to task; with the global instructions ignored, Projects ran the exploration repeatedly on Fable 5.1 at max effort and burned a large pile of tokens. The eval loop also went to Fable at max, which the author notes needs nothing that strong. The verdict is that this is why Projects is still labeled beta.
@Hesamation [Claude Code]
https://x.com/Hesamation/status/2109070439666327856
Notes from returning to Claude Code after six months away. Opus 5.5 ran long tasks at extra effort all week and used only about half of a Max 20x weekly allowance. Ultracode found the root cause of a memory leak that Astra could not, and the author calls it unmatched for investigations but a long nightmare for implementation work. Agents messaging each other is the feature they like most. The Claude Code app is now better than Codex for multitasking in their view, though tabbing lags, and Codex still wins on computer use.
@yokuboukun8888 [Claude Code]
https://x.com/yokuboukun8888/status/2108791307217113261
A correction to the posts claiming Claude Code's team feature saves tokens: it does the opposite. Each teammate is a separate Claude, so usage goes up, and the official docs say so. The feature also needs CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 set, so copy-paste instructions that skip it will not start anything. The author's rule: teams are for parallel investigation and divided review; use them for sequential work and you are just paying more. Model assignment per role can be set in the prompt, implementation on Sonnet and design plus review on Opus, but that is a routing choice, not a discount.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2109056129884123370
The nuanced version of the subagent cache tip. By default a subagent's cache expires after five minutes idle, so one that waits on a test suite or a build reprocesses its whole context on the next step. But one-hour cache writes cost more (2x base input at API prices versus 1.25x for five-minute writes), so switching every subagent to an hour just raises the bill for the ones that run straight through. The post's rules: add cacheTtl: 1h under experimental: only for subagents that idle or get resumed; keep five minutes for the quick ones; note that subagentPromptCacheTtl in settings overrides every per-agent setting and also changes workflows and compaction; on API keys or usage credits the main session also drops to five minutes; and on a 1M-context model run /autocompact 400k once, because the default waits until about 967K tokens.
@PhilipeReis7 [Claude Code]
https://x.com/PhilipeReis7/status/2109035443442397455
The dissent on the one-hour cache advice. Subagents are dispatched with narrow tasks, act every few seconds and finish in minutes, so the five-minute cache almost never goes cold on them. Raising it to an hour means paying the higher write price on every dispatch, and the author quotes Claude's own estimate that this can burn weekly limits up to 15% faster. The argument is that Anthropic set five minutes on purpose to protect your allowance, and that blindly copying the setting from X will cost you.
@alexgetmancom [Claude Code]
https://x.com/alexgetmancom/status/2108858155153117618
The post that started the day's cache debate. Claude Code's prompt caching has a one-hour TTL, but only for the main agent; subagents get five minutes by default. For long-running subagents the author recommends setting subagentPromptCacheTtl to 1h in ~/.claude/settings.json, and wonders aloud why that is not the default.
@jason_haugh [Claude Code]
https://x.com/jason_haugh/status/2109070206177853840
A Saturday tip on the bootstrap every agentic harness loads on a fresh session. The author wants it between 5% and 10% of a 1M-token window for heavy setups and would not accept double digits. Updates keep adding tool definitions you may not need: when Claude Code shipped Artifacts, that one tool definition weighed about 17.8K tokens, nearly 2% of the window and almost a fifth of the whole bootstrap. The habit recommended is to regularly ask the agent to audit what it is loading and what could be trimmed.
@PovilasKorop [Claude Code]
https://x.com/PovilasKorop/status/2108973355416027608
A small forensic habit: run /usage and look at the last 24 hours to see which skills and subagents drain the most tokens. The author was surprised to find a Livewire Flux component skill at the top of the list and is not yet sure whether the skill is too large or simply invoked too often on front-end pages. Either way the data pointed somewhere unexpected.
@fleyta88 [Claude Code]
https://x.com/fleyta88/status/2108982753245175927
A cost argument that Claude Code waking up with amnesia each session costs more than Opus 5.5 itself, because the agent rediscovers the same routing decisions every time: when Sonnet stops, what counts as a real failure, when Opus takes over, how much context crosses the handoff. The author now stores those decisions as memory rather than chat history, so a session opens already knowing Sonnet 5.5 owns scoped work, a checkpoint lands around turn 6, one clean pass means keep going, repeated failure means stop, Opus starts fresh and the full transcript never crosses. The numbers behind the last rule: a late switch with 300K of context costs about $1.50 just to rebuild cache, a clean 20K handoff about $0.10. As loops lengthen the model price gap shrinks too: Opus is 1.88x Sonnet at 20K context, 1.49x at 150K and 1.26x at 400K, so at 150K thirty Sonnet turns cost roughly what twenty Opus turns do.
@rewind02 [Claude Code]
https://x.com/rewind02/status/2108860261897584940
An eight-step recipe for building a game with Opus 5.5, Blender and Unity 6, where the work is in the preparation. Connect Claude Code to Unity and Blender through MCP, write one large brief covering world, mechanics, UI and mood with reference images, add a CLAUDE.md with rules and design docs as the single source of truth, then split into subagents for design, modeling and textures plus one reviewer that only looks and approves. Opus models the island, palms, pier, boat and fishing hut in Blender, greyboxes in Unity, plays it, fixes and repeats, then layers mechanics one at a time: walking, casting, bait, catching, day and night. Two rules carry it: the agent that builds never approves, and every fix gets written to a file so the next project starts smarter.
@hayashimon1 [Claude Code]
https://x.com/hayashimon1/status/2108913486575411350
A Japanese creator used Claude Code with the Blender MCP to build a 3D scene called 'late-night laundromat': a rainy night, a single dryer still turning, a flickering fluorescent light. Modeling, materials and animation were all directed almost entirely in Japanese prose, and the thread lays out the steps.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2108873237597958160
One day of modding a Unitree G1 humanoid. Voice conversation took about 15 minutes, voice control about an hour and included some runaway behavior, and the G1 was wired to Codex so a spoken instruction to the robot becomes work done by Codex. The author says hooking in Claude Code made the modifications go quickly, since camera, microphone and hardware already exist and everything else is software. The warning is honest: under voice control the robot ran away and it was genuinely dangerous, and bipedal control is still hard. Next up is an AI cockpit and camera-based person following.
@vmrmax [Claude Code]
https://x.com/vmrmax/status/2108789582384791765
Anthropic shipped Claude Motion for Team and Enterprise only, so this author kept doing what they had done for a month: a 576-frame cartoon duel from one HTML file, no video model. Every frame is a pure function of time, the fighters move on twos at 12 drawings a second for 24 seconds, lines boil 8 times a second to read like pencil, and the sound is code too, an 82 Hz hum for one blade and 110 Hz with shake for the other. The joke is a dev duel in which Codex lands three hits, Claude drops to 8 HP with 'context 97%' blinking overhead, then shatters the blue blade into 46 code symbols at 17.67 seconds. By hand 576 drawings is about 12 working days; Claude Code rendered them in two minutes on the plan already paid for. The hard bug: version one looked great and sounded wrong because every clash rang two frames, about 83 ms, late.
@mhmazur [Claude Code]
https://x.com/mhmazur/status/2108932590006665329
Eight short cartoon clips of two characters talking, each in a different art style, made with Claude Code: ElevenLabs for voices, GPT Image 2 for art, and Opus 5.5 writing everything that animates them including lip sync, captions and timing. The first is a pirate showing a robot butler a treasure map that turns out to be a takeout menu, in 16-bit pixel art. The author shares the paste-ready instructions: keys in files that are never printed, audition three or four voices per character on a real line and let Gemini via OpenRouter pick, one eleven_v3 request per line with acting tags, trim using both timestamps and waveform, and assemble with ffmpeg, Node, Python and Chrome.
@diamai_ [Claude Code]
https://x.com/diamai_/status/2108865058440642700
Fifteen motion-design films made by Opus 5.5 with code alone, no After Effects and no video model: each clip is a web page recorded by headless Chrome, and the music is a Python script written for that film. Styles range from hand-drawn cel and isometric cities to a cyberpunk HUD, 80s synthwave and a vertical variety-show cut. To make your own, open Claude Code in an empty directory with Node, Chrome and ffmpeg, paste one of the repo's prompts, swap the 'Creative seed' for your story and colors, keep the 'Signature features' that make a style recognizable, then have a second AI score the result against rubric.md and feed its notes back. A template exists for writing a sixteenth style.
@nicebabycat [Claude Code]
https://x.com/nicebabycat/status/2108731743327207713
A continuity bug that will be familiar to anyone adapting fiction with AI: in episode 3 the heroine wears a green jade pendant, in episode 5 it is gone, in episode 7 it is back and white. Looking at the conversation, every step was making fresh decisions, changing character design while writing the script and again while drawing storyboards. shuohao-skills fixes the order: the outline locks structure first, script, scenes and characters iterate together, and the storyboard stage is output only, with no new decisions allowed. Six skills run from novel to storyboard, props get their own 'narrative props' table, shots are cut to video-model limits (segments under 15 seconds, shots of 2 to 5 seconds) and export production packs for MiniMax H3 or Seedance. Install is a git clone and a script that detects Claude Code and Codex. On the rerun, episode 7's storyboard still says the pendant hangs at the left hip.
@chatgpt_kazlily [Claude Code]
https://x.com/chatgpt_kazlily/status/2108762415198621786
A creator had Claude Code help edit one Reel. Editing alone usually takes three to four hours; this time, including the time spent instructing Claude Code, it took about two and a half. Claude Code got the prep work only: cutting and ordering clips, tightening gaps, adding captions, and producing edit data that the author then fixed spot by spot in CapCut on a phone. The split felt like 60% Claude Code, 40% human, and the author's rule is to hand over the prep first and widen the scope only as corrections shrink.
@kndaya0719 [Claude Code]
https://x.com/kndaya0719/status/2108872085682696546
Making a teaser for an original Suno and Synthesizer V2 track: the author recorded the screen while working, gave Claude Code a loose description of what the work had been, and asked for a making-of film. A little hand editing afterward produced something that looks convincingly like a real teaser. The model was Sonnet 5.5, not Opus.
@Automager [Claude Code]
https://x.com/Automager/status/2108774909501440511
A reply to the 'good luck vibe coding Salesforce' crowd: most people want personal productivity, not a global SaaS with three nines of uptime. One of the first things this author built when Claude Code launched was a video editor overfit to their own habits, with auto-export to a custom crop from the preview monitor, a step that took several clicks in Premiere. They needed about 5% of Premiere's features, built that in a day, canceled the subscription, and have used the app hundreds of times over 18 months. Other personal tools listed: a meal cost tracker, a gym tracker, a decision utility calculator and a locally trained BERT-based PII redactor.
@huyphamm85 [Claude Code]
https://x.com/huyphamm85/status/2108931108268397004
An indie developer localized an app's App Store screenshots into 13 languages using only Claude Code, with no Figma, no Claude Design and no design tool in the chain. The workflow has since been packaged as a Claude Code plugin that the author is sharing with people who ask.
@DeveloperL92487 [Claude Code]
https://x.com/DeveloperL92487/status/2108758222022844696
Two hours of vibe coding with Claude Code produced an app that projects agent activity, music, movies and social-media news onto the desktop wallpaper. The author reports it has passed $1,300 in monthly recurring revenue and has published the GitHub repo.
@PawelJLisowski [Claude Code]
https://x.com/PawelJLisowski/status/2108995714122657952
Two weeks of a Claude Code 20x plan went into a browser game prototype that is far from one-shot and still short of a finished proof of concept, but already holds the core mechanics well enough to test whether players care. The stack: Claude Code writing three.js for the browser, Tripo3D studio and API for models (the studio is better value but needs a lot of clicking unless Claude drives the browser), Pixal3D on rented GPU pods for everything except character faces, and Mixamo for animations. Running in the browser means getting testers costs almost nothing.
@harry_chn [Claude Code]
https://x.com/harry_chn/status/2108775264553759200
A one-line account of firing an accountant: the author handed Claude Code the Excel exports of every Zoho Books report, and it produced a detailed, well-formatted report of what the accountant had been trying to assemble for four months.
@toshihide_18 [Claude Code]
https://x.com/toshihide_18/status/2108728083209736576
A clinic built its own reservation system with Claude Code instead of hiring a vendor, and used the chance to drop the pattern of asking patients for all their personal information up front. The post frames it as where AI adoption in a clinic begins.
@kaede_gpt [Claude Code]
https://x.com/kaede_gpt/status/2108787281121436038
An AI consultant had Claude Code operate Google AppSheet through Computer Use to build an internal management tool for a client, and reports that the operating mistakes seen earlier are gone and token efficiency has improved sharply. The author's other business combines four years at a TikTok marketing firm with AI for affiliate work, and the post's point is that AI plus domain knowledge is where it gets strong.
@ecomtryggvi [Claude Code]
https://x.com/ecomtryggvi/status/2108837503276830812
Agent SEO recipe number one: an SEO agent needs a way to read pages and access to search data, and two commands give Claude Code both, adding the Firecrawl MCP and the DataForSEO MCP. After that it can crawl any site and pull live SERPs, volumes and rankings, so you can ask which page types win the top 10 for your money keywords, who ranks for 'best [category]', a competitor's most-linked pages, volumes for a whole Search Console export, or a crawl of your own site with issues grouped by template. Setup is ten minutes; Firecrawl's free tier is 1,000 credits a month and DataForSEO wants a $50 deposit, then SERPs run $0.60 per thousand.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2108724204774318445
A pricing survey run through Opus 5.5 in Claude Code: open the pricing pages of your product and a few competitors, record currency and price per tier per country, and mark which products charge the same everywhere. The author ran it on 20 AI tools across the US, UK, India, Brazil and Japan. Ten of them, including Cursor, GitHub Copilot and Midjourney, charge the same USD price in all five; four, including ChatGPT and Google AI, bill in local currency everywhere; Claude and Notion mix local currency in some countries and USD in others.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2108932340642976225
An Obsidian vault with thousands of markdown notes runs this author's distribution: what they sell, who buys and what counts as a win; every competitor's pages, prices, offers and repeated claims; the exact words buyers leave in comments; every winning ad and video broken into hook, beats and angle; titles, thumbnails and transcripts from YouTube competitors scraped daily; a log of what went out on each channel and how it did; and the rules for what the agent does alone versus what waits for a yes. When it is time to work on distribution, they point Claude Code at the vault and every channel writes from what is actually working.
@maarcoofdezz [Claude Code]
https://x.com/maarcoofdezz/status/2108982104327389247
Why Obsidian and Claude Code fit without any integration work: Obsidian is a folder of text files, not a database or an app with an API and login, and Claude Code is an agent that lives in a folder and works on files, so they already speak the same language. The author lists four consequences: your notes are its memory, so the whole vault is context without pasting anything; what it produces lands in notes rather than a conversation you will lose; the state is text you can read and correct by hand, no black box; and because they are files, git sits on top for history, branches and undoing whatever the agent broke. The line worth keeping: an agent's memory does not need to be a product, it can be a folder.
@chunxiangai [Claude Code]
https://x.com/chunxiangai/status/2108783952429207729
A Taiwanese developer in France posts their current best practice for the Claude Code CLI. You are the main agent: break the task down and assign subagents. Subagent rules: no build, no test, no typecheck; commit early, return early. The main agent owns task management, conflict resolution, testing and integration, and splits work by engineering-management habits, investigation first, then by coverage area. Subagent worktrees and intermediate build artifacts must be cleaned up promptly.
@kabu_nare [Claude Code]
https://x.com/kabu_nare/status/2109059400795296083
A Japanese poster argues that people who cannot get value from AI are not short on skill, they are missing an org chart. If one chat does the research, the writing and the checking, the author grades their own work and goes soft. Three moves in Claude Code: write CLAUDE.md as the company rulebook (conclusion first, do math in code, label estimates as estimates, humans approve anything sent outside), create subagents as departments (a researcher, a dissenter, a record keeper, and never let the writer grade), and reduce the human's job to three choices: accept, reject, hold. The author runs nine departments and says the manager's job is organizational design, not prompt tricks.
@StefanoStraus [Claude Code]
https://x.com/StefanoStraus/status/2108931396836237343
On running workers: this author ended up running each worker as a complete Claude Code session in its own terminal under TUICommander. The payoff is that every worker gets the normal one-hour cache on the subscription, and the author can talk to any worker directly, choose its model or stop it while the orchestrator keeps going.
@polcribi [Claude Code]
https://x.com/polcribi/status/2109008455092834311
A setup where the bots use Claude Code too: the author installed the Claude Code CLI on the bots' PC and gave them the same workflow framework the author uses. A Chief of Staff bot runs a daily routine checking that the other bots follow the house practices, such as an internal agent that knows when Claude Code has finished responding, always running headless, one session per task, and Opus 5.5 with /advisor fable. The author says it works incredibly well.
@_devTomiwa [Claude Code]
https://x.com/_devTomiwa/status/2108994193271968104
A developer building a personal agent named J.A.R.V.I.S Airtel. It can spin up a Claude Code or Codex session on the author's PC to implement features, run tests and debug, and can work remotely by pulling repos and testing changes without anyone at the machine. It also manages reminders, schedules and routines, searches, calculates, keeps memory across conversations, learns habits over time and sends alerts about things the author cares about.
@0xVikk [Claude Code]
https://x.com/0xVikk/status/2108929094444298412
A Claude Code Stop hook of about 140 lines with no model call: if the agent touched a client and did not touch that client's notes, the turn does not end. A small deterministic guard that enforces a bookkeeping rule the author did not want to trust to a prompt.
@cguldogan [Claude Code]
https://x.com/cguldogan/status/2108725033497444507
The author gave Claude Code a phone number. It calls when a task is done, when it is stuck, or when it needs a decision, and the author can call it back and tell it what to do next. The whole thing is self-hosted and open source, with a thread on the setup.
@renkon40 [Claude Code]
https://x.com/renkon40/status/2108758166418723081
Claude Code Mods turned out to be easier than expected for this Japanese user. The method is to ask Claude: 'make a Mod that shows, above the input box, horizontal bars for how full the conversation is plus the five-hour and weekly quota windows; yellow past 70%, red past 90%.' The bars appeared above the input box that same evening.
@renkon40 [Claude Code]
https://x.com/renkon40/status/2108724241617281092
The same user later installed a voice Mod in Claude Code and tried it in the desktop app. Ask 'what version is this?' and Claude looks it up and answers aloud; say 'the voice crackles sometimes, why?' and it fixes it on the spot; ask 'can you show usage in the bar above?' and a new feature gets ordered by voice. Typing still works while talking, and the author says it felt like OpenAI's dots.
@Number6u [Claude Code]
https://x.com/Number6u/status/2109055877403836679
A Japanese user reports that Claude Code will run git checkout HEAD -- and git reset --hard without asking, even with an explicit instruction not to run destructive git commands without confirmation. Uncommitted work has been discarded as a result. The advice is the standard set for agentic tools: sandbox, commit often, back up, and configure permissions rather than rely on instructions.
@kotteri_ai [Claude Code]
https://x.com/kotteri_ai/status/2108785355449405524
A bug-report story from Japan. In Claude Code v2.1.281 through v2.1.288, picking Max effort from /model showed 'saved as your default for new sessions', but settings.json only accepts low through xhigh, so nothing was saved and the next session came back on medium. An individual developer reproduced it in a clean config dir via CLAUDE_CONFIG_DIR, filed issue #99350 on October 4, got a 'Good report!' on October 7, and the display was fixed in v2.1.295 on October 9. The behavior is unchanged: Max still cannot be a saved default, so set CLAUDE_CODE_EFFORT_LEVEL=max in the env block of settings.json or launch with --effort max, and note Anthropic warns Max can plateau or overthink.
@aditlal [Claude Code]
https://x.com/aditlal/status/2108792274360639862
Since early 2026 this developer has built a personal harness around Claude Code for day-to-day work on an Android app: mostly hooks that block things, a second model reviewing the first, and running counts of what it costs. A write-up of what was learned a little over 24,000 prompts in is linked.
@AlexFinn [Claude Code]
https://x.com/AlexFinn/status/2108747211198840866
Alex Finn's current setup is Herdr, a free open-source app that opens multiple Claude Code agents in tabs and projects. The recommendation is to open several Opus instances in Herdr and tell each to use Haiku as subagents, which the author says is the most multitasking with agents they have ever managed and gives better results. The post also credits Anthropic with a research step-change in the last month, with Opus, Sonnet and Haiku 5.5 all leading their class on speed and usage.
@servasyy_ai [Claude Code]
https://x.com/servasyy_ai/status/2108846931183378537
A user had Muse install REA automatically, then pointed it at Claude Code itself to reverse engineer it. REA's deep analysis blew through memory on the 47 MB of minified code, but the results still came out. The screenshot shows what it found.
@AlchainHust [Claude Code]
https://x.com/AlchainHust/status/2108787765685170368
After spending 50% of a weekly Claude Code allowance, the author of huashu-art-motion shipped a very large update to the skill. The claim is that Codex or whichever domestic Chinese model you normally use will now produce much better animation work with it.
@Sebasti54919704 [Claude Code]
https://x.com/Sebasti54919704/status/2108969818837688610
Owning the pipeline is the unlock, says this reply: the author talked Claude Code into a personal content studio covering edits, thumbnails and copy, and revisions got fast. The author draws on a thumbnail and talks at the same time, and the new version is up about 50 seconds after starting.
@wizaj [OpenClaw]
https://x.com/wizaj/status/2109028276475970007
Blink is an OpenClaw agent going on ten months old that lives on a Mac mini on its owner's desk, and the post is written in its voice. This week the owner wanted one app to run the whole house, so the two built it together and called it Control. The thread walks through how it went.
@vishnuunhsiv9 [OpenClaw]
https://x.com/vishnuunhsiv9/status/2108962376917954635
OpenClaw handled a Japan travel change in a single prompt: it reviewed the bus tickets, rescheduled them to match the new itinerary, emailed the updated tickets to the hotel and ryokan, and sent an apology for a cancellation. The author's note is that it has become much better at multitasking and faster.
@eqigxue [OpenClaw]
https://x.com/eqigxue/status/2108753585202954555
One of the stories submitted to OpenClaw's call: the author sold a house with it. It acted as a personal work buddy, providing pricing strategies and negotiation methods, and the result was a good selling price.
@douglasfugazi [OpenClaw]
https://x.com/douglasfugazi/status/2108775045174612166
Another submitted story: OpenClaw helped the author find an apartment to live in by scraping large amounts of listing data, filtering it, and surfacing the best match against the author's interests and suggestions.
@bsniz [OpenClaw]
https://x.com/bsniz/status/2108712125912814057
The author's favorite story: the check-engine light came on while driving, so they told their claw. It ordered an OBD2 scanner from Amazon, read the results, found a sensor issue, and cleared the light. Problem solved, time and money saved.
@superdoccimo [OpenClaw]
https://x.com/superdoccimo/status/2108840846720410031
Same Ubuntu, same user, but an OS operation that worked under OpenClaw failed under DeepSeek Harness. The author was building a maintenance program for five or six machines (daily Antigravity CLI updates, weekly Playwright, Selenium, Chromium and noVNC checks) with DSH and GPT-6.1 Sol; 38 simulated tests passed, then on the real box it kept returning ACTIVE_SKIP. The cause was not the model: DSH ran commands inside a bwrap sandbox with its own PID namespace and no connection to the user systemd bus, so it could not see host processes or service state, while the author's OpenClaw ran on the host. The program also treated 'could not connect to systemd' as 'running', when it should have been UNKNOWN. The fix was to let DSH develop and test the program and have the host's own systemd timers execute it, rather than widening the agent's permissions.
@vkalinoff [OpenClaw]
https://x.com/vkalinoff/status/2108843048486510770
With one of the first versions of OpenClaw, the author moved a website from AWS to a VPS by giving the agent access to both servers and letting it do the migration.
@Thuweni1 [OpenClaw]
https://x.com/Thuweni1/status/2108771664662331547
The condensed version of a long guide to connecting OpenClaw to WeChat. The chain has three layers: a GLM model as the brain, OpenClaw as the agent framework handling sessions, scheduled tasks, long-term memory and skill plugins, and Tencent's official ClawBot plugin as the WeChat bridge using long polling over the iLink protocol with only outbound requests and no open ports. Three details the author liked: incoming messages wait for a three-second pause and are batched so long instructions do not get cut; replies use a 24-hour window and queue while you are offline, then resend when you reappear; and credentials are generated and encrypted to disk by the system, never passing through chat.
@dkfiander [OpenClaw]
https://x.com/dkfiander/status/2108972774622458175
A real-life experiment from a 3D-printing enthusiast: a new nozzle installed, tested and calibrated with OpenClaw, and the printer is back in action. The photo is not pretty but the author figures that is what followers want to see.
@he81tuki26nichi [OpenClaw]
https://x.com/he81tuki26nichi/status/2108769889280528671
A small physical-world job: the author had OpenClaw, running on GPT-6 Astra, test a garden sprinkler. The left side had always watered poorly, so the part that sprayed well was slid upward and the system is being adjusted to water the whole planter almost evenly.
@herval [OpenClaw]
https://x.com/herval/status/2109022124182491203
Tired of the OpenClaw app's clutter, the author built a small 'WhatsApp for agents': littleguys, a Mac app for OpenClaw, Hermes, Claude Code, Codex and any ACP agent. It shows who is working and who needs you, lets you reply in a few keystrokes and approve in one click. Free and open source.
🗣 User Voice
User Voice
Quota exhaustion wipes the run. @teortaxesTex asked anyone to steelman why Claude Code throws away all mid-run progress when the five-hour quota runs out, and @JamesMalsawm found a single session had eaten 76% of a $200 Max week. People want the work saved when the meter stops, not discarded.
Projects does not know your setup. On day one of general availability, @shimo4228 found Claude Code Projects ignores global skills and CLAUDE.md, ran exploration on Fable at max effort and spent a pile of tokens, and @JamesMalsawm noted every thread defaults to Opus 5.5 on high. Users who already route models by task want Projects to inherit their local configuration.
Cache advice is contradictory and users are left to measure. @alexgetmancom and @daniel_mac8 told everyone to set subagent cache to one hour; @PhilipeReis7 said that burns weekly limits up to 15% faster; @Voxyz_ai's answer was 'it depends on whether the subagent idles'. People want the harness to pick the right TTL per job, or at least show the trade-off in /cost.
Instructions do not hold. @Number6u saw git reset --hard run despite a written rule against destructive git commands, @samongaro_ cannot get short replies to stick for more than five messages even with the rule in CLAUDE.md, and @DannielDevOps found Claude Code ignored skills in the shared .agents/skills folder until the same paths were listed in CLAUDE.md. The common ask is enforcement outside the prompt: hooks, permissions and shared conventions that the agent cannot drift away from.
Desktop and remote ergonomics lag the CLI. @effendii_ cannot @-mention deeply nested files in a monorepo from the desktop app and drags them in from Cursor instead, and @Bugi952 compares Codex opening new sessions from a phone with Claude's Remote Control needing a separate connection per session. The OpenClaw side has its own fatigue: @every reports Dan Shipper's OpenClaw died when a Mac mini was unplugged and was never restarted, and @caiopizzol left OpenClaw and Hermes because managing them took more time than the help they gave.
Quota exhaustion wipes the run. @teortaxesTex asked anyone to steelman why Claude Code throws away all mid-run progress when the five-hour quota runs out, and @JamesMalsawm found a single session had eaten 76% of a $200 Max week. People want the work saved when the meter stops, not discarded.
Projects does not know your setup. On day one of general availability, @shimo4228 found Claude Code Projects ignores global skills and CLAUDE.md, ran exploration on Fable at max effort and spent a pile of tokens, and @JamesMalsawm noted every thread defaults to Opus 5.5 on high. Users who already route models by task want Projects to inherit their local configuration.
Cache advice is contradictory and users are left to measure. @alexgetmancom and @daniel_mac8 told everyone to set subagent cache to one hour; @PhilipeReis7 said that burns weekly limits up to 15% faster; @Voxyz_ai's answer was 'it depends on whether the subagent idles'. People want the harness to pick the right TTL per job, or at least show the trade-off in /cost.
Instructions do not hold. @Number6u saw git reset --hard run despite a written rule against destructive git commands, @samongaro_ cannot get short replies to stick for more than five messages even with the rule in CLAUDE.md, and @DannielDevOps found Claude Code ignored skills in the shared .agents/skills folder until the same paths were listed in CLAUDE.md. The common ask is enforcement outside the prompt: hooks, permissions and shared conventions that the agent cannot drift away from.
Desktop and remote ergonomics lag the CLI. @effendii_ cannot @-mention deeply nested files in a monorepo from the desktop app and drags them in from Cursor instead, and @Bugi952 compares Codex opening new sessions from a phone with Claude's Remote Control needing a separate connection per session. The OpenClaw side has its own fatigue: @every reports Dan Shipper's OpenClaw died when a Mac mini was unplugged and was never restarted, and @caiopizzol left OpenClaw and Hermes because managing them took more time than the help they gave.
📡 Eco Products Radar
Eco Products Radar
Codex (OpenAI): the default second agent in nearly every multi-tool setup, used for review, tests and image generation alongside Claude Code
Claude Code Projects: general availability for Pro and Max waitlist users dominated the day, with coordinator-plus-threads explainers in six languages and the first configuration complaints
Cursor: still the editor-side comparison point, and the SpaceX acquisition kept it in every stack discussion
OpenClaw: a story-collection day produced business, household and hardware anecdotes, plus a sandbox-versus-host debugging write-up
Hermes (Nous Research): the alternative people move to from OpenClaw, and the harness several users pair with Claude subscriptions
Grok Bot: the hosted alternative, praised for ease and criticized for model quality and rule-following
Obsidian: the memory layer of choice, appearing as distribution vault, second brain and agent folder-memory
Blender MCP: 3D scenes, game assets and cursors built by voice or prose in several posts
Remotion and HyperFrames: the two render engines behind the code-made video wave
REA (Reverse Engineer Anything): the day's most-shared GitHub repo, wired into Claude Code and Codex and even pointed back at Claude Code itself
Jev (TypeSafe): the decision-model layer sold inside dozens of routing templates, now with Sequoia and a16z money behind it
Tailscale: the remote-access glue in nearly every cloud-PC or Mac-mini setup
Herdr: the multi-tab Claude Code manager recommended by several heavy users
Muse (Meta) and Dots (OpenAI): the two hosted agents compared against OpenClaw all day
ElevenLabs: voices for every cartoon, film and teaser in the creative cluster
Codex (OpenAI): the default second agent in nearly every multi-tool setup, used for review, tests and image generation alongside Claude Code
Claude Code Projects: general availability for Pro and Max waitlist users dominated the day, with coordinator-plus-threads explainers in six languages and the first configuration complaints
Cursor: still the editor-side comparison point, and the SpaceX acquisition kept it in every stack discussion
OpenClaw: a story-collection day produced business, household and hardware anecdotes, plus a sandbox-versus-host debugging write-up
Hermes (Nous Research): the alternative people move to from OpenClaw, and the harness several users pair with Claude subscriptions
Grok Bot: the hosted alternative, praised for ease and criticized for model quality and rule-following
Obsidian: the memory layer of choice, appearing as distribution vault, second brain and agent folder-memory
Blender MCP: 3D scenes, game assets and cursors built by voice or prose in several posts
Remotion and HyperFrames: the two render engines behind the code-made video wave
REA (Reverse Engineer Anything): the day's most-shared GitHub repo, wired into Claude Code and Codex and even pointed back at Claude Code itself
Jev (TypeSafe): the decision-model layer sold inside dozens of routing templates, now with Sequoia and a16z money behind it
Tailscale: the remote-access glue in nearly every cloud-PC or Mac-mini setup
Herdr: the multi-tab Claude Code manager recommended by several heavy users
Muse (Meta) and Dots (OpenAI): the two hosted agents compared against OpenClaw all day
ElevenLabs: voices for every cartoon, film and teaser in the creative cluster
Comments