Super User Daily: 2026-06-10
The center of gravity moved off the keyboard today. The strongest Claude Code and OpenClaw stories weren't about shipping code faster, they were about people pointing these agents at their own lives and businesses: a years-long sleep mystery cracked from a ring's data, a household chief-of-staff that runs the family calendar, a monitor-less Mac Mini quietly running an Amazon book empire. And underneath almost every story is the same nervous question, the bill. Token cost stopped being a footnote and became the design constraint itself, people are running three accounts, capping spend at $1,500 a head, and racing to install tools that shrink what the agent reads. Here is what real users actually did.
@hosseeb [Claude Code]
https://x.com/hosseeb/status/2064042451850121632
This is the non-coding case of the day. Hosseeb had a sleep problem that doctors couldn't pin down for years, so he handed Claude Code his Oura ring data through the API and let it run Python statistics across heart rate, SpO2, wake events and sleep stages. Claude built a dashboard and then did the thing a tired human wouldn't, it formed a hypothesis: upper-airway resistance syndrome. A $200 at-home WatchPAT study later confirmed mild sleep apnea. A terminal agent ran a differential diagnosis from wearable data and was right.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2064045567483990210
Futia built a Claude Code plugin that does market research the way an agency wishes it could. It scrapes Reddit, YouTube, TikTok and Instagram Reels for a niche via Apify and Firecrawl, ranks every post by real engagement, then pulls out the winning hooks and the exact words customers use, dropping it all into a content-brief dashboard. He ran it live on 'magnesium for sleep' and in three minutes it surfaced a 5.1M-view hook and a recurring customer complaint. That's a week of an intern's work compressed into a coffee break.
@thinking_slow [Claude Code]
https://x.com/thinking_slow/status/2063989600729604310
This is what an AI-native marketing org actually looks like, not a thread about one. At Ahrefs they wired Claude Code and Codex into a real content machine: a static-site blog with MCP connectors, vector embeddings over the whole sitemap so topical authority and internal linking happen automatically, scheduled content audits that pull rankings and backlinks through the Ahrefs MCP, and daily cron jobs that refresh the top articles. It all lives in one Content OS dashboard. The point isn't that AI writes posts, it's that the whole pipeline runs itself.
@LucasQin77 [Claude Code]
https://x.com/LucasQin77/status/2063954133778894944
Qin built Serenity Watch, an open-source Agent Skill that turns a trader's public X feed into a structured signal. Claude reads the posts, classifies each stock mention as bullish, bearish or neutral, and rolls it up per ticker. The output is concrete enough to trade off of: 95 tickers and 379 mentions in seven days, with $SIVE pulling 74 mentions, 66 of them bullish, all on an hourly-refreshing dashboard. This is sentiment analysis that used to need a data team, shipped as a skill.
@igus_ai [Claude Code]
https://x.com/igus_ai/status/2064135240898040055
The headline number here is the cost. Igus installed Zero with Claude Code, pulled live data on 17,000-plus stocks plus crypto and financials with no APIs and no OAuth, and built a Bloomberg-terminal-style dashboard for Apple, Tesla, Nvidia, Bitcoin and Ethereum off Finnhub and CoinGecko, auto-hosted on a public URL. Total spend: three cents. A product that vendors charge thousands a year for, rebuilt for the price of a stick of gum.
@aniketapanjwani [Claude Code]
https://x.com/aniketapanjwani/status/2063789202668425664
Academia is quietly going agentic. Panjwani interviewed economics professor Scott Cunningham about using Claude Code for actual applied microeconomics: a 51-post Substack series, probing models' tendency to over-search specifications, and a multi-language replication of the Callaway and Sant'Anna 2021 difference-in-differences paper. He's now teaching agentic coding inside his causal-inference workshops. The interesting part isn't speed, it's a researcher using the agent to stress-test methods, not just write code.
@DamiDefi [OpenClaw]
https://x.com/DamiDefi/status/2063954443427807578
This is the OpenClaw pitch in one screenshot. Dami turned a spare Mac Mini into an agent with ChatGPT as the brain, Telegram as the interface and full computer access through OpenClaw. From a single voice note it analyzed his business for revenue opportunities, researched YouTube content ideas complete with competitors, hooks and thumbnails, and laid out a 12-week newsletter strategy, all in minutes. The machine you weren't using is now a junior strategist that never sleeps.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2063879957848236302
Zixuan, a Chinese CS teacher, built a complete RPG in one weekend with Claude Code from a single idea, three maps, weapon, armor and potion systems, bosses, dozens of enemy types, and reportedly clears $10,000 a month off it, solo, no team, no investors. The number deserves skepticism, but the shape is now common: one person, one weekend, a shippable game. The barrier that used to be 'a studio' is now 'a prompt and a free Saturday.'
@NotionHQ [Claude Code]
https://x.com/NotionHQ/status/2064055129381421391
Notion put its own name on a non-engineer's automation, which tells you something. Their Workplace Experience Coordinator wired Claude Code and custom Notion agents into OfficeSpace and Slack to handle office seating end to end: one agent watches the new-hire database and pings managers on Slack, another assigns the actual seats. He's not a developer, and it saves him hours every onboarding cycle. The most repeatable lesson here is who built it.
@virattt [Claude Code]
https://x.com/virattt/status/2064086211199701346
Virattt did something subtle and smart: he built a side dataset specifically so Claude Code never has to read the long document. It's operational KPIs, load factor, book-to-bill, count of $100K customers, extracted and standardized from earnings releases, transcripts and IR sites, the stuff that isn't in XBRL or the 10-K. It updates in about 15 seconds and serves in milliseconds. The insight is that the cheapest token is the one the agent never has to read.
@ventry089 [Claude Code]
https://x.com/ventry089/status/2063939313767723089
This is a whole company running on a €200 box. The creator runs Claude Code on a monitor-less Mac Mini as a multi-agent hierarchy named Sophia that operates an Amazon book business: one agent scans hundreds of niches across four countries, another writes, a third reads the draft as the target customer persona and rates it, a fourth strips out the AI tells. The claim is $1.5M from Amazon plus a company sold for $300K. Treat the figures as marketing, but the org chart, one human, four agent roles, is the real artifact.
@VaibhavSisinty [Claude Code]
https://x.com/VaibhavSisinty/status/2063810805217505679
Testing is becoming something you just ask for. With MobAI connecting Claude Code to an iOS simulator, one plain-English prompt and zero test scripts lets Claude drive the whole app, navigate every screen, tap buttons, check the map, read the debug logs, and surface real bugs the developer had missed, then write up a structured report. QA used to be a role. Here it's a sentence.
@browomo [Claude Code]
https://x.com/browomo/status/2064008893550764248
A solo D2C founder runs Claude Code as seven specialized agents, email, morning briefing, approvals, feature dev, graph updates, sitting on a knowledge graph with persistent episodic memory and a human-approval layer. The claim is ~$19k a month in profit at $200-650 in agent cost, versus the $5,000-9,700 a human mini-team would run, with the morning's work prepared overnight. The economics are the whole point: the agent stack costs an order of magnitude less than the headcount it replaces.
@Luci_Grace_C [Claude Code]
https://x.com/Luci_Grace_C/status/2063927600796602530
Luci open-sourced a self-hosted AI companion built to run forever, persistent and self-healing on a personal Mac, reachable from phone or computer. The clever bit is the billing arbitrage: it exploits Claude Code's new split where non-interactive calls are metered separately, and uses Agent SDK credits to bridge into WeChat and Telegram. The repo even ships three flavors, human-readable, machine-readable, and 'feed it all to CC.' Someone built a relationship on top of a metering loophole.
@Sprytixl [Claude Code]
https://x.com/Sprytixl/status/2064003321086775768
Another solo dev, another Steam launch. A Chinese developer shipped a space-survival game built with Godot, one GPU and Claude Code, and reports about $14,000 a month. He described what he wanted and Claude Code built the player, planet and enemy scenes, the upgrade system, the space combat and the full game loop. It's live with a trailer and in-game purchases. The pattern is hardening: describe the game, get the game.
@techxsarfraj [OpenClaw]
https://x.com/techxsarfraj/status/2063819070764683583
Sarfraj packaged a whole business into a single skill file. His OpenClaw bot scrapes local businesses that have no website, auto-builds a custom site for each, emails them a preview link, and runs the whole loop daily on autopilot. Lead-gen, production and outreach, the three jobs a small agency hires for, collapsed into one shareable file that runs while he sleeps.
@earthtojake [Claude Code]
https://x.com/earthtojake/status/2064037239743721814
Hardware is falling to the same trick. Jake designed a 7-DOF robot arm with working kinematics entirely by prompting Claude Code and Codex, with no CAD software open at all. Custom skills had the agents generate STEP files and robot description files and run FEA checks straight from text and image prompts. CAD was a specialist gate; here it's a conversation.
@zaimiri [Claude Code]
https://x.com/zaimiri/status/2064049262225486170
A Ford legal-ops exec built a household chief-of-staff named Claudette out of Claude Pro, Claude Code and Claude Cowork. It scans email and calendar and sends a daily family-ops briefing, drop-offs, au pair handoffs, birthdays, recycling day, texts she needs to send. The honest detail is the tuning: she had to train it down from overly detailed to a genuinely useful operations manager. This is the agent quietly moving into the home.
@mfishbein [Claude Code]
https://x.com/mfishbein/status/2063977367010263211
Fishbein built a meeting-prep agent on the Claude Agent SDK and hasn't done manual prep in weeks. Every morning at 7 it pulls his calendar and, for each external attendee, runs a research playbook, Fullenrich, LinkedIn via Deepline, Exa and Perplexity, Fireflies transcripts, HubSpot classification, then emails a dossier via Resend, all deployed on Railway with a cron. The dossier even gets written back to HubSpot. This is the boring back-office grind, fully automated.
@cevenif [Claude Code]
https://x.com/cevenif/status/2063940602148889058
VibeOS is the stress test everyone wanted to see: across 64 logged Claude Code sessions, someone built a complete ARM64 operating system that boots on a Raspberry Pi, bootloader, kernel, filesystem, a GUI with draggable windows and a dock, a ported DOOM, a built-in C compiler, and its own VibeCode IDE. Logging each session turns it into real data about how far AI-assisted large engineering actually goes. The answer, apparently, is 'a whole OS.'
@0xMorlex [Claude Code]
https://x.com/0xMorlex/status/2063950646684901630
Morlex documents Noah Brier's setup that turns Claude Code plus Obsidian into a genuine second brain: point Claude Code at the root of a 1,500-note vault, run it in thinking mode so it asks questions instead of guessing, use the agents as sparring partners, and let it surface notes you'd forgotten you wrote. You resume work by saying 'catch me up.' This is the non-coding direction a lot of power users are converging on.
@Asteri_eth [Claude Code]
https://x.com/Asteri_eth/status/2063968844545228813
Asteri built the pipeline that catches your 3am ideas before you lose them. A raw note dropped into Obsidian gets auto-scanned, classified as project, task, content or noise, auto-researched, and turned into a structured plan. Then a human opens Claude Code to review and approve, after which one command spawns a project-manager agent plus dev, research and content sub-agents to actually execute. The human is in the loop for two minutes; the rest ships itself.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2063971657761595510
Lieben rebuilt a lead-gen agency inside Claude Code. Folder-based agents score his closed-won customers with a Python script, pull 150-plus enriched Apollo contacts in under two minutes, write copy off proven frameworks, and push campaigns into Instantly and Smartlead. It runs 24/7 on Railway and let him cancel $300-plus a month of scraping tools. The replaced thing here isn't a task, it's the agency itself.
@deanwball [Claude Code]
https://x.com/deanwball/status/2063973900896371039
Wall used Claude Code's dynamic workflows to stand up a virtual publishing house, fact-checkers, editors, researchers and competing writers arranged like a corporate org chart, generating 50-75k-word histories of any topic he feeds it. He ran hundreds of agents and burned roughly 10 million tokens on a single run, and the output crushed plain decentralized LLM calls. This is the clearest 'more tokens, more intelligence' artifact of the day, and he knows exactly what it cost.
@sadek [Claude Code]
https://x.com/sadek/status/2063930847074336886
Sadek built a product-auditor skill that screenshots an app, diagnoses the UX and UI fixes, classifies issues red, orange or yellow, and scores the product 0-10, then hooked it inside a /goal loop targeting an 8 with token and loop limits. He walked away for three hours, monitoring through Claude Dispatch and TeamViewer from 1,800km away, and came back to a substantially built-out app. This is the loop pattern as overnight contractor, set a bar, leave, return to finished work.
@0xcherry [Claude Code]
https://x.com/0xcherry/status/2063886587684069628
Cherry, author of OpenAlice (5,000 GitHub stars), rebuilt the trading-assistant product in v0.30.0 by throwing away his own agent framework. The home-grown one burned millions of tokens on pre-trade analysis and couldn't support skill injection, so after studying how Claude Code, Codex and OpenClaw do it, he dogfooded a rewrite that fixed usability and re-accelerated growth. The lesson buried here: rolling your own agent loop is often the expensive mistake.
@Chenzeze777 [Claude Code]
https://x.com/Chenzeze777/status/2063962196477190417
This is the token-cost anxiety turned into a tool. Chen tested Headroom, an open-source compressor that hit 14k GitHub stars in a week, which squeezed a code search from 17k tokens down to 1,400, a 92% cut, reversibly and locally, with no change to the answer. It wraps Claude Code, Codex or Cursor in one line. He's running it and waiting to see what next month's bill looks like. When the bill is the constraint, the compressor is the product.
@iam_elias1 [Claude Code]
https://x.com/iam_elias1/status/2064061537791402235
Elias breaks down Graphify, a free MIT-licensed tool that pulled 55,100 GitHub stars in seven weeks by cutting Claude Code's token use 49x to 71x. It turns a codebase into a queryable graph via tree-sitter AST across 40-plus languages, then installs a hook so Claude navigates by structure instead of grepping every file. Same story as Headroom from a different angle: the breakout open-source projects this week are all about spending fewer tokens, not bigger models.
@KingBootoshi [Claude Code]
https://x.com/KingBootoshi/status/2063999432077795579
Bootoshi's workflow writeup is a real shift in how people orchestrate. He moved from Claude Code bossing around Codex sub-agents to a single Codex agent running a task A-to-Z in /goal mode on gpt-5.5 and opus-4.8. The load-bearing idea is ADRs, Architectural Decision Records as plain .md files, used as the durable context the agent reads, plus a discussion phase before any code and computer-use E2E dogfooding. The moat isn't the agent, it's the written-down decisions.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2063899498259128670
Huryn shares a cross-model review trick with the full skill spec to copy: have Claude build, then a Codex skill review the result, because the model that wrote the bug shares the blind spot that hides it. He posts the verbatim codex-ideation skill, a SKILL.md plus a codex.py wrapper around 'codex exec', so you can paste in a second pair of eyes. Two models checking each other beats one model trusting itself.
@sanchoyai [Claude Code]
https://x.com/sanchoyai/status/2063876453859811523
Sancho's team adopted TestSprite in one sprint, and unlike diff-reading tools, it opens the app and clicks through staging like a real user before it writes any test plan. In week one it caught a billing-flow bug that three engineers had waved through in PR review. It rolls into Claude Code and Codex as an MCP server. The recurring theme this week: the agents that earn trust are the ones that actually use the product, not just read the code.
@davideciffa [Claude Code]
https://x.com/davideciffa/status/2063984025258082813
Ciffa launched Luce Spark, which calibrates expert selection for 33-35B MoE models based on your own Claude Code and Codex sessions, then caches the cold-hit experts in a small GPU cache. The result is 40% less memory and a 1.5x speedup versus naive offloading, enough to run those models full-speed on a 16GB machine. It's a quiet but real unlock: bigger local models on the hardware you already own.
@Abdulr3hm33n [Claude Code]
https://x.com/Abdulr3hm33n/status/2064010162080236025
Affaan Mustafa won the Anthropic and Forum Ventures hackathon by building a complete startup in eight hours with Claude Code, then open-sourced the engine behind it, 'ecc', a system he'd refined over 10-plus months of daily use that turns Claude into a full engineering team with a skills library and specialized subagents like OWASP security review and memory optimization. The eight-hour win is the demo; the ten-month system is the actual moat.
@ardakutsal [Claude Code]
https://x.com/ardakutsal/status/2064081048821027231
Kutsal built GARL, a Global Agent Reputation Ledger, an open-source Apache-2.0 trust layer for the agent economy. It signs every agent action with a cryptographic receipt, tracks reputation across five dimensions, uses shrink-only permission tokens, and anchors receipts to Base mainnet, with a 5-line GitHub Action that signs AI commits from Claude Code, Codex, Copilot or Cursor. A few thousand verifications are already running in production. As agents start acting, someone has to make the actions accountable.
@gagarot200 [Claude Code]
https://x.com/gagarot200/status/2063853598040289366
This is the weirdest arbitrage of the day. Gagarot buys unused or secondhand Kindles for around 1,500 yen, uses Codex and Claude Code to build the custom features a client asks for onto the device, and resells each modified Kindle for 150,000 yen, selling out in five minutes. A hundredfold markup on a dead-cheap device, powered entirely by an agent writing the firmware nobody else will. Hardware hacking just became a margin business.
@S1TA10 [Claude Code]
https://x.com/S1TA10/status/2063926336121393541
Small discovery, real consolidation. S1TA10 realized GPT Image 2 was already inside their Codex subscription and callable from the same terminal, so now they generate logos, mockups and hero images directly inside Claude Code without leaving the prompt, and cancelled a separate Higgsfield subscription. The interesting trend is centripetal: capabilities keep getting pulled into the terminal until the standalone tools fall away.
@daken_in_market [Claude Code]
https://x.com/daken_in_market/status/2064130040355516899
A clean illustration of what cheap implementation unlocks: someone had Claude Code build 100 stock-trading bots, backtested all of them, kept the best performers and pushed those to live trading. When writing a strategy costs almost nothing, you stop picking one idea and start running a tournament. The model isn't 'build a bot,' it's 'build a hundred and let the data choose.'
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2064104431063617797
The token bill in one tweet. Masahiro uses Claude Code's /loop to keep running E2E tests until the build is actually done, and reports that token consumption is so heavy that even $200 a month isn't enough, so he runs a three-account setup. This is the most honest line of the day about the real cost of loops, the smarter the loop, the bigger the bill, and people are paying it three subscriptions at a time.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2063788774166020458
Leo frames agent-loop cost as the real expense of AI coding and brings the receipt: Uber capped Claude Code and Cursor at $1,500 per person per tool per month after burning the annual budget in four months. His own fix is process discipline, fixed Codex threads, active-task and pickup JSON, daily routing threads, all aimed at not re-paying for the same context every handoff. The frontier problem has quietly shifted from capability to cost control.
@dotey [Claude Code]
https://x.com/dotey/status/2064044201369567454
Dotey gives the sharpest explanation of why OpenClaw wins where polished tools don't: it has broad permissions and access to private data, and runs scripts at high privilege on a persistent machine, iterating and burning tokens until the task succeeds. He notes he avoids the more locked-down Claude Cowork for exactly this reason, and that restricted setups like WeChat AI can't compete because they lack both the permissions and the always-on, token-burning environment. Access plus persistence beats a nicer UI.
🗣 User Voice
User Voice
Five things real users kept saying this run.
First, token cost is now the binding constraint, not a footnote. People are running three accounts to dodge the $200 cap and routing every handoff through JSON to avoid re-paying for context (@masahirochaen, @runes_leo). A single dynamic-workflow run can burn 10 million tokens (@deanwball).
Second, the answer they want isn't a bigger context window, it's spending fewer tokens. Compression and graph tools that cut what the agent reads by 49x to 92x are getting installed the day they ship (@Chenzeze777, @iam_elias1), and people now build side datasets just so Claude Code never has to read the long document (@virattt).
Third, the best use cases keep drifting away from code, into health, family ops, research and content (@hosseeb, @zaimiri, @phosphenq).
Fourth, Claude Code nails function but not polish, so users bolt on dedicated image and design tools rather than wait for it to get pretty (@S1TA10, @igus_ai).
Fifth, permissions plus persistence are why OpenClaw-style setups win, broad access on an always-on machine that can burn tokens until the task is actually done (@dotey, @techxsarfraj).
Five things real users kept saying this run.
First, token cost is now the binding constraint, not a footnote. People are running three accounts to dodge the $200 cap and routing every handoff through JSON to avoid re-paying for context (@masahirochaen, @runes_leo). A single dynamic-workflow run can burn 10 million tokens (@deanwball).
Second, the answer they want isn't a bigger context window, it's spending fewer tokens. Compression and graph tools that cut what the agent reads by 49x to 92x are getting installed the day they ship (@Chenzeze777, @iam_elias1), and people now build side datasets just so Claude Code never has to read the long document (@virattt).
Third, the best use cases keep drifting away from code, into health, family ops, research and content (@hosseeb, @zaimiri, @phosphenq).
Fourth, Claude Code nails function but not polish, so users bolt on dedicated image and design tools rather than wait for it to get pretty (@S1TA10, @igus_ai).
Fifth, permissions plus persistence are why OpenClaw-style setups win, broad access on an always-on machine that can burn tokens until the task is actually done (@dotey, @techxsarfraj).
📡 Eco Products Radar
Eco Products Radar
Codex — the default second agent. People pair it with Claude Code for cross-model review, design passes and image generation (@PawelHuryn, @KingBootoshi, @S1TA10).
NotebookLM — the emerging reader in a read-versus-build division of labor that keeps Claude Code's context free for implementation (@shota7180, @0xMorlex).
Obsidian — the second-brain substrate people point Claude Code at for notes, ideas and memory (@0xMorlex, @Asteri_eth).
Cursor — still the everyday IDE companion sitting next to Claude Code (@efecim1sn, @Chenzeze777).
MCP — the connective tissue wiring tools, data and verification into every one of these workflows (@thinking_slow, @sanchoyai).
Higgsfield — the go-to image and video generator wired straight into terminal build flows (@S1TA10).
Codex — the default second agent. People pair it with Claude Code for cross-model review, design passes and image generation (@PawelHuryn, @KingBootoshi, @S1TA10).
NotebookLM — the emerging reader in a read-versus-build division of labor that keeps Claude Code's context free for implementation (@shota7180, @0xMorlex).
Obsidian — the second-brain substrate people point Claude Code at for notes, ideas and memory (@0xMorlex, @Asteri_eth).
Cursor — still the everyday IDE companion sitting next to Claude Code (@efecim1sn, @Chenzeze777).
MCP — the connective tissue wiring tools, data and verification into every one of these workflows (@thinking_slow, @sanchoyai).
Higgsfield — the go-to image and video generator wired straight into terminal build flows (@S1TA10).
Comments