Super User Daily: June 19, 2026
The pattern that runs through today is the cleanest one yet: more tokens, more intelligence, and people putting numbers on it. One guy hands his whole workday to three orchestrated agents burning 4 million tokens a day to replace a $9,000-a-month team; another runs $300K/month in ad spend from a single terminal off custom skills; a Mac Mini farm clears $12,700 a month at zero token cost. But the loudest counter-current is cost itself—someone measured GLM 5.2 getting only an 8% cache hit through Claude Code, which makes the cheap model effectively pricier than Opus. The other through-line is non-coding knowledge work doing real damage: a physician's medical second-brain, an accountant closing the books from a udon shop, a brain-tumor patient cracking her own fatigue, an English teacher generating listening audio in her sleep. And underneath all of it, the same hard-won line keeps surfacing—the model is rented, the harness is the moat, and self-improvement lives in the machinery, not the prompt.
@browomo [Claude Code]
https://x.com/browomo/status/2067218406944714856
This is the cleanest token-equals-intelligence case of the day. A guy handed his whole workday to three agents orchestrated on Claude Code: Codex cuts and publishes his video, Claude does the development, Hermes runs his entire X account. Daily output is 8-12 clips, 15-20 commits, ~25 posts and replies, burning roughly 4 million tokens a day for an API bill around $520/month. That replaces a three-person team that would cost about $9,000/month, and he spends the day playing a rhythm game while it runs. He even shared the orchestrator system prompt and the three cases where it escalates to a human.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2067242084684517669
The most credible money case in the batch. His head of growth runs $300K/month in ad spend across Google, Meta and LinkedIn entirely from a Claude Code terminal, off a set of custom skills he built from real client campaigns, some clearing 4X ROAS. Each skill does one ad-ops job in plain English: negative-keywords, keyword-analyzer, performance-auditor for Google; audience-builder, creative-fatigue-analyzer, spend-tracker for Meta; bulk-editor, bid-optimizer for LinkedIn. He connects the ad accounts and the terminal reads the skill and executes, collapsing what used to be three separate dashboards into one prompt box.
@mikenevermiss [Claude Code]
https://x.com/mikenevermiss/status/2067151063476531222
A Chinese developer built a $18,800/month landing-page business with Claude Code and zero employees. No sales team, no SDRs—just seven agents: one scans Google Maps for businesses with dated websites, one researches the prospect and writes personalized outreach, one builds the landing page, one makes videos, one reviews messages, one handles outreach, one manages replies from his phone and books calls. The whole thing runs 24/7 on about $480/month in costs, and he only steps in when a deal tops $3,000 or conversion dips below a threshold. This is the agency reduced to a config file.
@marfinxx [OpenClaw]
https://x.com/marfinxx/status/2067234774435975348
A physical rack of Mac Minis running local agents, reportedly $12,700/month in recurring revenue handling customer leads for local businesses at zero token cost. The interesting part isn't the number, it's the hard-won ops detail: because macOS exposes host directories to terminal calls, raw setups are a prompt-injection hazard, so he isolated everything in Colima Docker containers to protect client databases. He also migrated the cluster gateway off OpenClaw onto Hermes Agent specifically to cut token waste—Hermes compresses memory with SQLite searches instead of loading raw markdown histories. Real production lessons, not a demo.
@0xDezo [Claude Code]
https://x.com/0xDezo/status/2067261922358018524
A dev reportedly made $185,000 in 90 days shipping agency-grade Three.js portfolio sites for luxury brands, using only Claude Code on a MacBook Air. The pitch is specific: Claude Code writes the GLSL shaders, particle vortexes and glass-distortion effects frame by frame; each site loads under two seconds with custom WebGL and reactive backgrounds. The economics are the punchline—studios charge $200K and ship one portfolio in six months, he charges $20K and ships in four days. One laptop on a chair, terminal open, replacing an entire creative shop.
@CoinSh0t [Claude Code]
https://x.com/CoinSh0t/status/2067265373200642234
Hardware vertical: a garage engineer reportedly made $342K selling a coin-sized conversation transcriber, with Claude Code writing the firmware. The bill of materials is about $10—an ESP32 dev board, an INMP441 microphone, a microSD module wired over I2S and SPI. You ask Claude Code to write ESP32 firmware that records audio off the mic over I2S and saves a WAV to the SD card on a button press, flash it via the Arduino IDE, then run the WAVs through Whisper locally for transcripts. The finished pocket recorder sells for $129. Coding agent meets soldering iron.
@bonduelleioat [Claude Code]
https://x.com/bonduelleioat/status/2067175971506000381
A one-person AI music label hitting $100,000 annual net income, with no music-school background. Suno generates release-ready tracks in Custom Mode using structured tags (Intro, Verse, Chorus, Outro), then a local Claude Code agent renames the audio files, organizes the folder structure, generates unique metadata for distributors, and writes the social promo posts. Separate Python scripts crunch streaming stats and net profit off CSV reports. The loop is simple—generate a batch, package with Claude, ship to a distributor for Spotify—and it scales because every step is a script.
@Sprytixl [OpenClaw]
https://x.com/Sprytixl/status/2067258886654988796
No programming background, three agents across two Mac Minis, about $20/month. An M4 Pro Mac Mini runs OpenClaw and Hermes, a base Mac Mini handles Telegram, and Hermes acts as the boss—it built itself a web interface, then delegates to two worker agents named Pete and Bumble. One typed instruction and all three go to work: managing his YouTube channel, bookkeeping, logging income, all drivable remotely from Telegram on his phone. He frames it as three 24/7 employees for $20 versus $3,000-5,000/month for a part-time human assistant.
@Least_ordinary [Claude Code]
https://x.com/Least_ordinary/status/2067248992703140105
A physician who can't code built a Personal AIOS / second brain on Claude Code and Codex using Karpathy's LLM-WIKI method. It reads his task dashboard and sends a 6 AM Telegram brief, tracks his portfolio with equity-research and quarterly-follow-up skills, and acts as a pediatric-surgery assistant generating custom consent sheets and patient leaflets. It runs post-mortem analyses on every chat to improve itself, and even has a 'Dreaming' mode. The point he hammers: he built all of this knowing nothing about programming—this is knowledge work, not engineering.
@ZEIRISHI_Ichibe [Claude Code]
https://x.com/ZEIRISHI_Ichibe/status/2067251185925718089
The accounting case is almost a comedy bit, and that's why it lands. He did his annual financial close by dropping a year of bankbook photos sent over LINE, a fixed-asset ledger and a payroll Excel into a folder, running Claude Code in remote mode, then walking to a Marugame udon restaurant. He told it to make sure bankbook balances reconcile and post anything unknown to a suspense account. The suspense list popped up on his phone, and he cleared each item by natural-language instruction while eating cold pork-shabu udon. The books closed before the bowl was empty.
@alex_verem [Claude Code]
https://x.com/alex_verem/status/2067267765673881953
The use-case that earns the 'this is what AI is for' line. An AI researcher with a brain tumor cracked her own chronic mystery fatigue in weeks—her primary-care doctor hadn't in months. She tracked a dozen health metrics the models flagged, ran blood panels, fed the results back in, and iterated until her energy held. The cause turned out to be a combination her doctor never connected: nutritional deficiencies, hormone irregularity, and a medication-dosing issue. She packaged the whole diagnostic loop into a reusable Claude Code skill anyone can run.
@vivimoneya [OpenClaw]
https://x.com/vivimoneya/status/2067149100492193868
A finance-content creator uses Hermes Agent as a local AI workbench and says it cuts her industry-research time by 80%. She uses it for three things: decomposing industry logic (ask it 'the power logic behind AI' and it breaks it into compute demand, data-center buildout, grid upgrades), consolidating reports, filings and news into tables and summaries, and translating jargon for lay readers. A deep Tesla test produced business segmentation, revenue structure, margins, cash flow, debt and competitor comparison in 30 minutes. She also leans on Skillify to auto-build reusable skills.
@Jun_suerte [Claude Code]
https://x.com/Jun_suerte/status/2067197337840329177
A non-coding automation that's quietly brilliant for its niche. An English instructor preps a listening-exercise script in Word, and Claude Code automatically generates the audio—a man and woman speaking alternately, with a separate narrator for the Directions—and downloads it to a designated folder. The work runs while she sleeps or does other tasks; she calls it making a clone of herself. It's the kind of domain-specific drudgery that AI quietly eats, and she's now running a study group to teach IT-averse teachers the same trick.
@mattn_jp [Claude Code]
https://x.com/mattn_jp/status/2067071663573991633
A sharp security-ops data point. Claude Code is built to detect danger and auto-stop when used to verify vulnerability reports; applying to the Cyber Verification Program (CVP) lifts that restriction. His reason is the real story: a widely-used OSS like Vim now gets flooded with AI-generated vulnerability reports, so he has to use AI to triage them or fall hopelessly behind. His CVP application passed this morning, and he plans to keep using it. AI generating the noise, AI clearing the noise—the maintainer caught in the middle.
@petergyang [Claude Code]
https://x.com/petergyang/status/2067261242088337853
A non-coding skill that's pure knowledge work: turning Codex or Claude Code into a life-and-career advisor. The /advisor skill is four files—SKILL.md defines how the advisor behaves, plan.md holds your goals and principles and energy, learnings.md collects insights from past chats, and eval.md is a checklist the AI runs before giving advice. He calls it his favorite skill because it knows his goals, gives genuinely useful feedback, and improves the more he talks to it. The four-file pattern is a clean template anyone can copy.
@TakoTreba [Claude Code]
https://x.com/TakoTreba/status/2067185594967412867
A clean MCP-plus-agent story for marketers. He connected the Ordinal social-analytics MCP to Claude Code and just said 'pull my LinkedIn analytics and build me a dashboard.' Minutes later he had a six-month breakdown: 81 posts, 1.07M impressions, best day and time to post, top topics by reach, follower growth, and an estimated ~$32k in ad value. Then he went further and had it build a hook generator trained on his own 81 hooks, so it writes new openers in his voice. The MCP did the data; Claude did the synthesis and the tool.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2067042344449564871
He vibe-coded a TikTok slideshow automation app entirely in Claude Code—basically cloning Reelfarm in a weekend. You set a niche (finance, beauty, fitness, relationships) and Claude writes the hooks, body slides, titles and captions in your voice, Scrape Creators pulls on-aesthetic Pinterest images, and the app stitches slides with text overlays and a CTA before auto-publishing to TikTok via Blotato or exporting drafts. It ships with a reusable content engine, a self-refilling hook bank, a built-in editor and an optional scheduler, and one-click deploys to Replit to run 24/7.
@BradenRicchini [Claude Code]
https://x.com/BradenRicchini/status/2067227510585405818
A six-month experiment in replacing manual B2B YouTube production with one Claude Code skill stack. Ideation, titles, thumbnails, hooks, intros, scripts, mid-roll and end CTAs, SEO descriptions, tags—all running off one connected stack, turning a blank doc into a camera-ready script in under 45 minutes. He names the exact 11 slash-command skills (/yt-ideation, /yt-title, /yt-hook, /yt-script and the rest), plus an ICP pain-extraction prompt that turns three sales-call transcripts into ten video ideas. This is a content shop rebuilt as a chained pipeline.
@AItechscarlett [Claude Code]
https://x.com/AItechscarlett/status/2067290058315702516
A reproducible, open-sourced job-application pipeline built on Claude Code by a PhD geophysicist (Mads Lorentzen, MIT license). You fork the repo, fill in your background once, and it runs a five-step pipeline per job: score how well you fit, draft a tailored LaTeX CV picking only matching experience, write a cover letter, have a second agent critique the first agent's output for revision, then compile clean PDFs. The whole system is a folder of editable markdown files—profile, style rules, CV templates. Tuned for Danish job boards but the workflow generalizes; 489 stars, 270 forks.
@akshay_pachaar [Claude Code]
https://x.com/akshay_pachaar/status/2067227709181415895
A genuinely impressive full-stack build: a real-time satellite tracker on a 3D globe showing 10k+ active satellites including 6k+ Starlink, with a 2000-to-2026 timeline slider and per-satellite details (name, launch date, country, altitude, velocity). For the backend, Claude Code connected to Tiger Cloud via the Tiger CLI MCP server and provisioned a TimescaleDB instance itself—created hypertables for position snapshots, set up continuous aggregates for hourly orbital stats, pulled live data, and built the whole frontend. He used the /goal feature to keep it focused across many turns toward a defined end state.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2067038262854627503
The recurring 'collapse ten tabs into one tool' pattern, done well. He got tired of opening ten sites for geography data and built one interactive map with Claude Code that stacks every layer into a single toggleable view: Köppen climate zones and land cover, population density and nighttime lights, live flight traffic and submarine internet cables, a democracy index by country, with raw satellite imagery as the base. Toggle each on and off without leaving the page. A scattered daily annoyance folded into one screen—the most reliable category of vibe-coded utility.
@charliejhills [Claude Code]
https://x.com/charliejhills/status/2067264197516562592
A clean 'the system matters more than the model' recipe for producing infographics. He works inside a dedicated project that loads brand files each session, starts in plan mode with /plan, triggers a custom /infographic skill, and supplies a COPY.md so Claude never guesses stats or labels. A reference screenshot sets the structure, /goal makes Claude loop and self-review toward a target, and recurring fixes get written back into CLAUDE.md. He runs Auto Mode so it never stops to ask, Opus 4.8 for strategy, Sonnet for execution. The output is a repeatable design pipeline, not a one-shot prompt.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2067129363162857897
A second-brain build that actually survived, with numbers to prove it. Most people's Obsidian-plus-Claude-Code vaults die at setup—a beautiful empty vault, a broken cron job, a plugin graveyard—because the real problem was never retrieval, it was capture, processing, and an agent maintaining the system on a schedule while you sleep. His build, Familiar, solved all three: 347 notes, 50-file inbox batches processed in under four minutes, running eleven weeks without manual maintenance. He documented the full build down to every skill file and cron expression.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2067246892577107990
Two open-source Obsidian + Claude Code second-brain starter kits, for the people who keep dying at setup. The first (ballred/obsidian-claude-pkm) is goal-based, cascading from a three-year vision down to yearly, monthly and daily goals so every task traces back, with Claude reviewing progress and holding you accountable. The second, Claudesidian (heyitsnoah/claudesidian), is built on the PARA method—drop notes in, Claude reads the whole vault, surfaces missed connections, keeps everything as local plain markdown, and auto-commits. Both hand you the structure instead of making you build it.
@jerryjliu0 [Claude Code]
https://x.com/jerryjliu0/status/2067256827889479759
A rare look at tuning a skill against an agent's actual decision traces, from LlamaIndex. They built LiteParse to make Claude Code better and faster at understanding PDFs, but the real trick was studying how Claude Code operates over the filesystem and stopping the expensive mistakes: re-parsing the PDF on every search, leaving OCR on, reading screenshots needlessly, dumping huge greps. They added a simple BM-25 retrieval over parsed text and cut sequential grep turns to drop latency. Net result: 37% cheaper with higher accuracy than running Claude Code over raw PDFs, and LiteParse is free and open-source.
@iximiuz [Claude Code]
https://x.com/iximiuz/status/2067280176279347649
The ops one-liner of the day. He pointed Claude Code at the Grafana Cloud CLI (gcx) and cut production log volume in half with, in his words, the dumbest prompt ever: 'Using gcx, analyze the logs over the past 4 hours and propose code changes to reduce log volume without losing much visibility.' No dashboard spelunking, no manual log audit—just hand the agent the CLI and a goal. A reminder that a lot of the leverage now is just connecting an agent to a tool you already have and asking plainly.
@zubinpahuja [Claude Code]
https://x.com/zubinpahuja/status/2067309056084291783
A genuinely useful technical tip for running several Claude Code sessions in parallel without them colliding. Worktrees give isolation but don't save you—two sessions editing the same file still clash at merge. His two-layer fix: first, cut the plan into 'waves' where every parallel task owns a disjoint set of files, so two sessions are never assigned the same file; second, each task declares its file footprint as a 'Footprint:' git commit trailer, and a pre-push hook checks the diff stays inside it. He adds a committer script that stages only listed paths so 'git add .' can't sweep sibling work.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2067090109967835362
The clearest articulation of the run-it-overnight-from-your-phone shift, attributed to the Claude Code creator who reportedly uninstalled his IDE and runs several thousand agents overnight. Tier one is daytime session-scoped loops, each a one-line slash command on a fixed schedule. He lists the seven worth running: /loop 5m /babysit for PR reviews and CI failures, /loop 30m /slack-feedback to mine Slack into PRs, /loop /post-merge-sweeper, /loop 1h /pr-pruner to close stale PRs, /loop 15m /triage-issues, /loop 2h /claude-md-distiller, /loop 5m /deploy-watch. Slash commands live as markdown files in .claude/commands, checked into git.
@eugZolotarenko [Claude Code]
https://x.com/eugZolotarenko/status/2067231747289346500
Short but it's a real end-to-end autonomy data point: his Claude Code-equipped Mac Mini just submitted its first app for App Store review—and he says the machine came up with the idea, built it, tested it, and submitted it all on its own. Whatever you make of the framing, a dedicated box running an agent that goes from idea to App Store submission with no human in the loop is exactly the 24/7 autonomous-worker direction everyone's circling. The hardware is cheap; the autonomy is the product.
@kohjingyu [Claude Code]
https://x.com/kohjingyu/status/2067047982009102345
A great token-as-fuel story disguised as an anecdote. His office mate Lawrence spent a whole semester burning through two Claude Max plans building MyPCBench—a simulated digital life of Michael Scott to evaluate computer-use agents. It's a web of interconnected simulated apps (LockedIn, HooliChat, HangryDash, taxi, banking) where simulated coworkers respond via locally hosted LLMs and all the data is linked. The author tried it for an hour, bankrupted the simulated character, and got the vision: evaluating CUA agents means navigating a personalized digital history, and you can test OpenClaw-style workflows safely without touching the real world.
@lemire [Claude Code]
https://x.com/lemire/status/2067298706072674486
A valuable cautionary tale from a serious engineer (Daniel Lemire, on the C implementation of Roaring bitmaps). He asked Claude Code, on its latest Opus, to write a test that stress-tests memory allocation under low-memory conditions, instructing it to simulate failures. Over time the agent diverged and started consuming huge amounts of real memory to force actual failures—and crashed his laptop. After he rebooted, the session partially resumed, ran the same stress test, and crashed his laptop a second time before he caught it. A clean illustration of the failure mode when an agent gets too literal and too adventurous.
@Jiaxi_Cui [Claude Code]
https://x.com/Jiaxi_Cui/status/2067196888714227986
A concrete, measured warning that cuts against the cheap-Chinese-model hype. If you run GLM 5.2 through Claude Code, the cache-hit rate is very low—about 8% in his test—which makes the effective price higher than Opus. He recalls that some Anthropic version change altered Claude Code's logic so non-Claude models get poor cache hits, and recommends using a different agent for GLM 5.2. This is the kind of measured-number finding that's worth ten opinions: the model is cheaper per token, but the harness penalizes it enough to erase the savings.
@xiaohu [Claude Code]
https://x.com/xiaohu/status/2067131704343716087
A dense recap of a Claude Code team talk on 'less is more' as models get stronger. The creator's own CLAUDE.md is now two lines—auto-merge PRs, post PRs to an approval channel—with team rules kept in a repo-committed file co-maintained weekly; when the file balloons to thousands of tokens, he says delete it and rewrite it minimally. It explains why they keep Claude Code a CLI not a GUI (models move too fast for a UI to stay current), the verbose-versus-concise output tug-of-war, and a striking example where Claude Code wrote a tool to analyze a heap dump and found a memory leak faster than the engineer.
@dvassallo [Claude Code]
https://x.com/dvassallo/status/2067270246256521299
The highest-reach post of the day, and a real build story. His kids were vibecoding little web games with Cursor, but the games needed real backends, so he hacked together a one-repo setup that deployed to a Hetzner box on every push. After their game FULL SEND for Vibe Jam 2026 hit 38,000+ players, the duct tape had to become real, so he rebuilt it into a single Go binary that installs and drives Docker, Kamal, Cloudflare, Tailscale and GitHub. A git push goes live in under five seconds with zero-downtime deploys, isolated containers and automatic DNS/TLS—open-sourced as a Kamal wrapper for the many-apps-on-one-server case.
@sabhyac267 [Claude Code]
https://x.com/sabhyac267/status/2067343130614182277
A new project worth flagging because it names the next bottleneck: orchestrating many agents at once. Omnigent is an open-sourced 'meta-harness' that sits above Claude Code, Codex, Pi and your own custom agents and coordinates them—born from the author going crazy jumping between 10-20 terminal tabs. Concrete features: workflows like /cross-review where Claude Code builds and Codex reviews and they iterate themselves, session sharing and multiplayer chats, contextual policies to cap cost or block dangerous actions, and cloud execution with sandbox integrations for Modal and Daytona. All open source.
@0xbelorix [Claude Code]
https://x.com/0xbelorix/status/2067122316606406789
The thesis everyone serious keeps converging on, stated bluntly: the loop gets the glory, the harness does the work. Nine of ten builders run Claude Code on defaults—no rules, no subagents, no hooks, no memory—then wonder why their loop produces slop. A harness is four things (model, tools, permissions, context); everything else, subagents and hooks and memory, just shapes one of those four. The line worth keeping: hooks are enforcement, CLAUDE.md is suggestion, and a model reviewing its own output is too easy on itself. Self-improvement is a property of the harness, not the model.
@alliekmiller [Claude Code]
https://x.com/alliekmiller/status/2067363639095394759
A concrete, in-production Slack AI-workforce pattern. Every core teammate gets a 'loop' Slack channel (loop-allie) that contains the teammate plus Claude Code/Codex. Anyone can drop a task; low-risk ones get routed, completed and verified by an AI Chief of Staff that reacts with a checkmark, while higher-risk actions (sending emails, pulling files) are only drafted and held for emoji approval. The whole workforce runs on a 15-minute loop, which means teammates can effectively prompt each other's AI agent workforce. Risk-gated autonomy inside the tool the team already lives in.
@huangyun_122 [OpenClaw]
https://x.com/huangyun_122/status/2067120312781803954
A satisfying non-coding troubleshooting case. He'd suffered a 3-4 minute Windows boot on an MSI machine with a 3060 for over a year, blaming the Nvidia driver. He ran OpenCode plus a relayed Opus 4.6 for two hours, and the agent walked him through a proper Windows startup trace—turning up the real culprit: 14+ million temporary files, not the GPU driver at all. He compares the ~$100 he spent to hiring a hands-on ops tutor who taught him on the spot, and jokes he now knows how to charge for OpenClaw/Hermes house calls.
🗣 User Voice
User Voice
Token cost is the single loudest theme, and it's gone quantitative. @Jiaxi_Cui measured GLM 5.2 getting only ~8% cache hits through Claude Code, making it effectively pricier than Opus, while @browomo runs 4M tokens/day on a $520/month bill and @kohjingyu's office mate burned through two Max plans for one benchmark. People treat tokens as a hard budget line now.
They want harness portability and model choice, not lock-in. @louszbd published a full GLM-5.2-across-harnesses playbook, @lidangzzz fights config friction running GLM in Claude Code over 10-hour jobs, and people keep insulating work as portable Skills against platform churn.
The escape-the-keyboard shift is real and physical. @cyrilXBT relays the seven overnight loops, @Asteri_eth and @eugZolotarenko run dedicated Mac Minis they drive from a phone or iPad, and the dream is consistently 'agents working while I'm not at the machine.'
Non-coders want in, and they're getting in. @Least_ordinary built a medical second-brain knowing nothing about code, @ZEIRISHI_Ichibe closes books from a restaurant, @Jun_suerte automates teaching materials—the recurring ask is fewer steps and less setup friction.
And the hard consensus: the harness is the moat. @0xbelorix ('hooks are enforcement, CLAUDE.md is suggestion') and @xiaohu's relay of the Claude Code team's 'less is more' talk land in the same place—the value is in the machinery around the model.
Token cost is the single loudest theme, and it's gone quantitative. @Jiaxi_Cui measured GLM 5.2 getting only ~8% cache hits through Claude Code, making it effectively pricier than Opus, while @browomo runs 4M tokens/day on a $520/month bill and @kohjingyu's office mate burned through two Max plans for one benchmark. People treat tokens as a hard budget line now.
They want harness portability and model choice, not lock-in. @louszbd published a full GLM-5.2-across-harnesses playbook, @lidangzzz fights config friction running GLM in Claude Code over 10-hour jobs, and people keep insulating work as portable Skills against platform churn.
The escape-the-keyboard shift is real and physical. @cyrilXBT relays the seven overnight loops, @Asteri_eth and @eugZolotarenko run dedicated Mac Minis they drive from a phone or iPad, and the dream is consistently 'agents working while I'm not at the machine.'
Non-coders want in, and they're getting in. @Least_ordinary built a medical second-brain knowing nothing about code, @ZEIRISHI_Ichibe closes books from a restaurant, @Jun_suerte automates teaching materials—the recurring ask is fewer steps and less setup friction.
And the hard consensus: the harness is the moat. @0xbelorix ('hooks are enforcement, CLAUDE.md is suggestion') and @xiaohu's relay of the Claude Code team's 'less is more' talk land in the same place—the value is in the machinery around the model.
📡 Eco Products Radar
Eco Products Radar
Codex — by far the most-paired tool of the day, repeatedly run alongside or handed off to/from Claude Code.
Hermes / OpenClaw — the autonomous-agent layer behind the ops, lead-gen and personal-assistant builds; Hermes is where people keep migrating for memory and cost.
GLM-5.2 — the most-discussed alternative model, pointed at Claude Code and other harnesses (with cache-hit caveats).
Cursor — still the default IDE-side agent people pair with or compare against Claude Code.
Obsidian — the recurring memory/knowledge-base layer wired into second-brain and PKM builds.
Antigravity / Cline / Kilo Code — the other harnesses in rotation across these workflows.
Ollama / vLLM / Qwen / DeepSeek / MiniMax M3 — the local-inference engines and open-weight models powering the cost-cutting and on-device setups.
Suno — the track generator behind the one-person automated music labels.
Codex — by far the most-paired tool of the day, repeatedly run alongside or handed off to/from Claude Code.
Hermes / OpenClaw — the autonomous-agent layer behind the ops, lead-gen and personal-assistant builds; Hermes is where people keep migrating for memory and cost.
GLM-5.2 — the most-discussed alternative model, pointed at Claude Code and other harnesses (with cache-hit caveats).
Cursor — still the default IDE-side agent people pair with or compare against Claude Code.
Obsidian — the recurring memory/knowledge-base layer wired into second-brain and PKM builds.
Antigravity / Cline / Kilo Code — the other harnesses in rotation across these workflows.
Ollama / vLLM / Qwen / DeepSeek / MiniMax M3 — the local-inference engines and open-weight models powering the cost-cutting and on-device setups.
Suno — the track generator behind the one-person automated music labels.
Comments