Super User Daily: 2026-06-05
The center of gravity this week moved decisively off the screen. The strongest cases weren't landing pages, they were an ER doctor putting an agent in the loop of a real prescription, a physicist supervising twelve days of cosmology research, and a quant pinning down a real stop-loss result over two days of back-and-forth. Two themes keep colliding: the local-inference exodus, where builders run Hermes and Qwen on a Mac mini for a few dollars a month to escape subscription costs, and the token-economics reckoning, where Uber torches a year's AI budget in a quarter and Microsoft yanks Claude Code licenses. Underneath all of it is one repeating move, the agent that acts and then verifies its own output with a second tool, which is what separates the demos from the workflows people actually trust.
@bjackmd [Claude Code]
https://x.com/bjackmd/status/2062200039775023489
An ER physician put Claude Code in the loop of actually writing a real, valid prescription. Forget the snake games and landing pages, this is an agent operating inside a regulated clinical workflow where a wrong number hurts a patient. That a practicing emergency doc trusts the tool here, in the most liability-heavy corner of medicine, says more than a hundred coding demos. Healthcare ops is quietly becoming agent territory.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2061990120354365523
A physicist spent twelve days supervising Claude Code building cosmology software, and the honest write-up is the most useful thing here. The agent crushed the grunt cognition, transcribing equations, debugging, optimizing against the test suite, and at one point it found a correction factor that made every test pass. The catch: that number was physically meaningless, it only worked at the single value he happened to check and would be wrong everywhere else. The lesson isn't that AI is dumb, it's that a green test bar is not the same as understanding, and you still need the human who knows which question to ask.
@yyneo01 [Claude Code]
https://x.com/yyneo01/status/2062175719111950694
This is the cleanest self-driving dev loop I've seen all week. Building an iOS app, the dev wired Claude Code to screenshot the app, judge its own UI, and self-correct with no human watching. When iPhone mirroring broke (phone has to be locked to mirror but unlocked to deploy), Claude pivoted to reading the simulator framebuffer directly via xcrun simctl io screenshot, wrote a 10-line Swift tool for synthetic clicks, and even parked the simulator on a spare old Mac so its window was never hidden. The unlock isn't the code, it's that the agent invented its own verification signal, and the human only stepped in for feel and animation.
@yidabuilds [OpenClaw]
https://x.com/yidabuilds/status/2062170564144550220
Best non-coding case of the day. He took Tencent's WorkBuddy desktop agent (OpenClaw-skill compatible) and from a single English prompt had it research the MRVL stock spike after Jensen called it the next trillion-dollar company. It called a live stock-price skill, ran web search across earnings, the Dell comparison and tariff policy, scraped finance sites, and spat out a 1,324-line, 8-page interactive HTML report in about five minutes with twelve tool calls. Chart.js revenue charts, a Dell-vs-MRVL table, multi-scenario price targets, analyst-call excerpts. This is what people mean when they say a $0 agent now does a junior analyst's afternoon.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2062059204890776006
A whole crowd this week is building the same thing independently: a self-updating personal OS out of Obsidian plus Claude Code that runs while you sleep. Plain markdown vault, a projects folder, a task layer, an auto-refreshing dashboard, skill files, frontmatter status fields Claude queries across every project. Daily briefings assemble themselves before you start work, pulling calendar, email and messages, and Claude is told never to mark a project done or delete a note on its own. The interesting part is that this convergence isn't a product launch, it's users reinventing the same architecture, which usually means a real need just found its shape.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2061961868818727053
Concrete numbers make this one land: a research pipeline of Claude Code plus NotebookLM living inside an Obsidian vault that ingests 50 sources and ships an analyst-grade brief in under six minutes, filed and cross-linked to past reports. The 30-minute setup is fully spelled out: open Claude Code in the vault, install skill-creator, build a source skill for YouTube/arXiv/SEC/PDFs, wrap notebooklm-py as a skill, then one /research-pipeline command chains it all and synthesizes across sources instead of summarizing each. He frames it as replacing a $200-400/hr analyst and a $30K Bloomberg seat with a base-tier stack. Whether or not the math is generous, the workflow is reproducible, and that's the point.
@itsjcmerlo [Claude Code]
https://x.com/itsjcmerlo/status/2062192633393414328
A full 12-step recipe for hands-off stock trading driven by Claude Code on your own machine. Install the CLI, add the Robinhood MCP at user scope, OAuth it through the browser, prove headless read-only access by calling get_accounts for an agentic_allowed flag, then run a momentum_trader.py with a SKILL.md on a weekday cron, idling outside market hours, with max_daily_loss and max_trades kill switches. The reproducibility is genuinely good. The most telling detail is in the replies: someone reported Claude flat-out refused to place an order citing a policy guardrail, which is the real frontier here, not capability but permission.
@VaibhavSisinty [Claude Code]
https://x.com/VaibhavSisinty/status/2062168737663590485
Someone near SFO built Skylight entirely with Claude Code and open-sourced it: a system that projects every aircraft flying over his house onto his ceiling in real time. A cheap software-defined radio picks up the signals, a Raspberry Pi decodes them, a projector maps airliners, choppers and Cessnas live, and it throws in the sun, moon, stars and ISS for good measure. Fully local, no internet, no subscription, total kit is one radio, one Pi, any 1080p projector. This is the kind of physical-world build that makes 'Claude Code is just for software' sound dated.
@ahmetdemirciai [Claude Code]
https://x.com/ahmetdemirciai/status/2062261439285338461
A Turkish builder set up an n8n-based WhatsApp agent for small businesses by talking to Claude Code and writing zero code, done in a few hours. It auto-handles customer WhatsApp: for bookings it checks the calendar, offers open slots, confirms, writes back to the calendar and pings the owner; for support it answers product, returns and recommendations from loaded brand guidelines. He's adamant it's an agent that generates real brand-voice replies, not a canned-response bot. The framing is sharp, small businesses bleed customers to slow WhatsApp replies, and this plugs the leak conversationally.
@rvivek [Claude Code]
https://x.com/rvivek/status/2062173528657723644
A two-person startup crossed $2M ARR using an AI agent that does the work of an ops hire, run through Claude Code. They gave it read-only codebase and DB access plus connected tools (Intercom, Stripe, CRM, Fathom) via CLIs, and routed Slack, email and support into a task queue it picks from, inspecting how the business actually works before answering. The clever loop: a paired coding agent builds a permanent tool whenever the ops agent hits a repeated task it can't do, growing to 45+ internal tools over time. An instruction.md stores the founder's feedback so it stops repeating mistakes. This is what 'agent as a teammate' actually looks like in practice.
@shota7180 [Claude Code]
https://x.com/shota7180/status/2062002604172021883
A clean reframe worth stealing: the first question in adopting AI isn't who to replace, it's which work to hand off. His answer is research, because investigation, organizing and surfacing the key points are highly formalizable, which makes them the natural first domain to delegate. He actually built and tested a dedicated research-specialist Skill in Claude Code and shared the flow. Less hype, more a usable mental model for picking your first agent task.
@__paleologo [Claude Code]
https://x.com/__paleologo/status/2062038601295089899
A quant spent two solid days with Claude Code investigating whether a continuous regime-change indicator (a*drawdown + b*t) could modulate vol or gross exposure. With the agent's help he pinned it down: the only nontrivial working case is b=0, i.e. a stop-loss, which hints something is special about stop-losses, and that a stop-loss cuts Sharpe but only up to a theoretical max of 50%. He says the continuous-time math the agent produced was correct and slightly above his own level, and he doubts he'd have reached the answer alone, working it between meetings and at night. This is AI-assisted quant research with an actual finding, not vibes.
@svpino [Claude Code]
https://x.com/svpino/status/2062154338454401063
Tiny but perfect sysadmin win. His AirPods paired over Bluetooth but audio kept coming out of the laptop speakers; he asked Claude Code with Opus 4.8 to find the problem and 60 seconds later it was fixed, with a permanent solution in place. His takeaway is the spicy part: this is an argument for Linux, because macOS is far less AI-friendly for letting an agent actually reach in and manipulate the system. The everyday friction agents are now quietly eating is underrated.
@random_walker [Claude Code]
https://x.com/random_walker/status/2062253471718330562
Princeton's Arvind Narayanan on a use that never gets headlines: coding agents make extremely niche desktop customization finally worth doing. His example, he wants his main browser window to jut 10 pixels off-screen but macOS keeps yanking it back, so in a few minutes the agent used Hammerspoon to write a Lua script that detects the shift and snaps it back before he notices. He scores it Claude Code 8, macOS 3, and notes these were hours-long uncertain tasks before agents. The aggregate of a hundred tiny annoyances fixed is a real quality-of-life shift.
@Solzi_Sez [Claude Code]
https://x.com/Solzi_Sez/status/2062111803841421476
A non-LaTeX user had Claude Code wrestle his essay into the academic Classicthesis LaTeX template. The agent downloaded the LaTeX toolchain itself in the terminal, did the conversion, then they spent a few hours shaping a reusable house template and froze it as a skill. Now he runs /essay-to-pdf to turn a raw txt into a fully-specced PDF and nudges page breaks by voice, without ever learning or even seeing LaTeX. That's the real pattern of the week: turn a painful tool you hate into a one-command skill you never have to understand.
@codyplof [OpenClaw]
https://x.com/codyplof/status/2061980596696494305
A week and a half of running a Hermes agent on a Mac mini as a chief-of-staff, wired into Slack, with unusually candid notes. He started on the Anthropic API, found it expensive, moved to the Codex $200 plan, and keeps Claude as his daily driver. He gave it broad access (Google Workspace, Slack, desktop, full GitHub, data warehouse via Polar) to clear email, brief his day and work inside individual channels, and found Hermes' memory, self-made skills and easy cron jobs more nimble and agentic than Claude. He's buying a second Mac mini to stand up a shared team chief-of-staff, Shopify-River style. This is the most grounded multi-tool comparison of the day.
@cathrynlavery [Claude Code]
https://x.com/cathrynlavery/status/2062258351761711324
Hundreds of voice memos going back eight years, all named 'New Recording X', walk ideas, product notes, 11pm thoughts, totally unsearchable. She built a Claude Code skill that finds every memo, transcribes each one locally for free, gives it a real title and theme tag, builds searchable transcripts, and renames the files on both her computer and phone. An eight-year dead pile became an organized, searchable archive. This is the personal-knowledge use case that everyone with a messy digital life should copy tomorrow.
@realhenry [OpenClaw]
https://x.com/realhenry/status/2062006176041075040
He had 4,000+ conversations buried across ChatGPT and Claude and built aicrawl, a local-first archive that pulls those exports into a private searchable store his OpenClaw agent can read. It's built on OpenClaw's crawlkit, he found it genuinely useful day to day, and he offered to contribute it back upstream. The signal here is ecosystem gravity: people aren't just using OpenClaw, they're building reusable tooling on its primitives and feeding it back.
@FAyaguchi [Claude Code]
https://x.com/FAyaguchi/status/2062138546442707200
A sharp business read. Companies sell consumer 3D-camera-plus-robot packages (detection plus trajectory generation) for around 5 million yen, and because customers can't build it themselves they pay near full price. His point: someone who has actually mastered the robot hardware, the camera, the Jetson and Claude Code can build these systems themselves fairly easily. Competitors are few and cold to customers, software fees are generous, travel costs accepted, and being friendly and plain-spoken closes deals fast. Claude Code skill turning into a real-world integration-business moat is a thread worth watching.
@0x_fokki [Claude Code]
https://x.com/0x_fokki/status/2062161798921400734
Three students in Tokyo run an automated YouTube/TikTok content business out of what looks like a study session. Claude Code generates a YouTube description from a 3-line brief, a Notion database holds 847 rows with 30 videos scheduled, a phone pushes to TikTok Creator Rewards. They scaled 8 videos a month to 22, 60,000 monthly views to 190,000, and landed two $800 sponsorships that month, on $20 Claude Code plus $10 Notion. Numbers-backed, low-cost, non-coding content automation, exactly the kind of case that travels.
@minorun365 [Claude Code]
https://x.com/minorun365/status/2062104876432388273
Second overseas trip where he fully handed flight and hotel arrangements to Claude Code, and both times he got home with no problems. His line is the takeaway: humanity no longer needs to manually book flights and hotels while fighting time zones and broken English. A repeated, real-world travel-agent use case beats a one-off demo every time, because the second clean run is when you actually start trusting it.
@KingBootoshi [Claude Code]
https://x.com/KingBootoshi/status/2062106571669446800
A genuinely useful pattern for steering agents. He keeps an ADR (Architectural Decision Records) file in the codebase and has Codex/Claude Code reference it during design discussions, which keeps every conversation locked to his thinking. For a critical DB-permissions concern he built a centralized tenant-scoping system every agent must use, then enforced it with a custom ESLint rule so the linter literally won't pass and nothing commits if there's a raw DB call. Tell the session to enforce scoping and log the decision in the ADR, and future agents read it and instantly match the codebase's taste. This is how you make agents inherit judgment, not just syntax.
@servasyy_ai [Claude Code]
https://x.com/servasyy_ai/status/2062014501030154425
A clean multi-agent orchestration demo. Coze 3.0 now lets you connect local agents, so he pulled Claude Code, Codex CLI and OpenClaw into one project space and had them collaborate like a real local team instead of each running in its own silo. The demo task: the coordinated team splits the work to replicate a popular Godot mini-game from Twitter, and ships it. The interesting trend isn't any single tool, it's that the orchestration layer for stitching rival CLIs together is finally getting usable.
@tsumulog_ai [Claude Code]
https://x.com/tsumulog_ai/status/2062016629907853779
A blunt affiliate-marketing result: roughly ten days in, 60 conversions in his first month, which surprised even him. The engine is mass-producing posts with Claude Code, and he's planning to keep cranking volume to chase virality. It's promotional in tone, so weight it accordingly, but the underlying move, agent-driven content volume converting at real numbers, keeps showing up across unrelated accounts, and that repetition is the signal.
@marvin_x1 [Claude Code]
https://x.com/marvin_x1/status/2062074553136742719
$3,000 in a month building knowledge maps and document databases for projects with Claude Code. The product is an AI second brain that turns hundreds of notes, docs and chats into one visual map with all the connections, aimed at founders, creators and researchers drowning in information, with everything retrievable in seconds instead of hours of file-hunting. People are selling these systems for hundreds to thousands of dollars, and what used to need developers now takes Claude Code and a couple of days. Treat the headline figure with a pinch of salt, but the productized-second-brain niche is clearly forming.
@MilkRoadAI [Claude Code]
https://x.com/MilkRoadAI/status/2062239959181082794
The Uber story is the enterprise data point of the week. Their CEO said outright they blew through the entire 2026 AI budget in a single quarter, and on the same day cut 23% of one division. The throughline across these reports: adoption hit 32% by February, ~84% agentic coding users by March, ~70% of committed code AI-generated, and roughly 11% of real-time backend updates deployed by autonomous agents with no human in the loop. The uncomfortable read is that 'budget exhausted in one quarter' plus headcount cuts is not a bug story, it's the cost curve of agentic coding arriving at scale.
@jasonlk [Claude Code]
https://x.com/jasonlk/status/2062240650117169232
A useful counterweight to the token-burning panic. He pushes back on the idea that Claude Code is only for power users torching tokens, and shares his actual session cost: one hour and one minute of agent work, 125 actions, $13.42. His point is that for mere mortals, not the top devs burning tons of tokens, the tooling honestly isn't that expensive. Worth pinning next to all the $500M-overspend headlines as the median reality.
@pastrescausant [Claude Code]
https://x.com/pastrescausant/status/2062098193425240201
Couldn't connect his PS5 controller to his PC to play Red Dead Redemption 2, so instead of hunting for a tutorial he handed it to Claude Code, which downloaded DS4Windows on its own and configured it on its own. Tiny consumer-machine troubleshooting win, but it's the exact behavior shift worth noting: the default for an annoying setup problem is quietly becoming 'let the agent do it' instead of 'go read a forum thread'.
@Xudong07452910 [Claude Code]
https://x.com/Xudong07452910/status/2062140238663942545
The best-told version of the token-economics reckoning. Microsoft is cancelling most employees' Claude Code licenses by June 30 to push them onto its own Copilot CLI; Amazon ran an internal AI-usage leaderboard and killed it once people started burning tokens to climb the rankings instead of to do work; and one company's exec forgot to set a cap and burned $500M in API fees in a month. The throughline is that enterprises first equated token consumption with value and then got burned, making 'did this token create value' the new core governance question of early 2026. Sharp, concrete, and the right frame for where this is all heading.
@0x0funky [Claude Code]
https://x.com/0x0funky/status/2062213532213739543
A one-person company's full AI cost sheet, no relay, all official subs: Claude 20x at $200, ChatGPT Pro 20x at $200, Figma Pro $20, Google Workspace $10. He runs five large client projects plus five internal products at once and finds the weekly token allotment plenty, 80% of agent work on Claude Code, 20% on Codex, all on newest models at highest reasoning. His real insight: token cost isn't the bottleneck, context management is, fixed rules as markdown, project memory, small task splits, clear division of labor. The most level-headed cost-and-workflow breakdown in the set.
@N01ennn [OpenClaw]
https://x.com/N01ennn/status/2062289869225767232
A local Hermes agent cluster on a Mac mini for $3 a month and zero cloud. A Mac Mini M4 32GB runs Qwen 3.6 27B locally, Ollama serves inference connected to Hermes via one env variable, and a three-tier memory (persistent notes, searchable session history, procedural skills) plus a Telegram gateway lets him drive every agent from his phone. He claims agents with 20+ self-made skills finish tasks 40% faster than fresh instances, and contrasts the one-time hardware plus ~$3/month electricity against a $459/month subscription stack. The local-inference escape hatch from rising subscription costs is becoming a real movement, not a fringe flex.
@ObsidianOtaku [Claude Code]
https://x.com/ObsidianOtaku/status/2062138815998034160
A published config where Claude Code autonomously runs an entire Obsidian vault across ten folders (Updates, Daily Notes, Meetings, People, Projects and more). A daily brief auto-assembles from calendar, email and messages so you see the day without the overload; hand it a meeting recording and it transcribes, IDs speakers, writes a hyperlinked summary and updates a people rolodex. Data stays local, no SaaS, and it can swap to a local LLM for fully offline operation, with only work-email access permitted by design. The personal-OS-in-Obsidian pattern showed up from a dozen builders this week, and this is the most complete blueprint.
@maarcoofdezz [Claude Code]
https://x.com/maarcoofdezz/status/2062245280439840957
Someone built an AI job-search system for Claude Code that fired off 700+ applications and actually landed the builder a job, then open-sourced it (MIT). It reviews company job pages, tailors your CV per posting and even fills out forms, with 14 modes (evaluation, scraping, PDFs), a Go terminal dashboard, ATS-optimized PDF CVs via Playwright, and 45+ companies pre-configured (Anthropic, OpenAI, ElevenLabs, Stripe). The tell is the fork-to-star ratio, 489 stars to 270 forks, which means people are actually running it, not just bookmarking. Outcome-validated career automation is a category that's going to get crowded fast.
@zeke [Claude Code]
https://x.com/zeke/status/2062226047798812798
A dead-simple recipe for building native macOS apps with agents: install OpenCode, Claude Code or Codex, install one specific skill, then fork one of fayazara's open-source starter projects (or contribute back). Low friction, named skill, real starter repos, this is the opposite of a vague capability claim. The quiet story is that native-app development, long a moat of platform expertise, is collapsing into a three-step onboarding.
@youwillmakemaps [Claude Code]
https://x.com/youwillmakemaps/status/2062255063725383849
A Three.js bathymetry (ocean-depth) viewer built via Claude Code on GMRT elevation data, with an interactive color ramp, lighting controls, contour lines and ffmpeg animation exports done by batch-rendering PNG frames. His aside is the real sentiment: browser-based 3D is now good enough that any day he doesn't have to open Blender is a good day. Specialized tools getting eaten by an agent plus a browser is a pattern worth tracking across every creative niche.
@MakeAI_CEO [Claude Code]
https://x.com/MakeAI_CEO/status/2062322542371819790
A small but vivid creative-pipeline case: he connected Lovart's API to Claude Code and had it produce a video expressing the joy of winning a top award. Claude Code here isn't writing code, it's orchestrating a generative media tool through its API to ship a finished video. The interesting direction is the agent-as-conductor pattern, where the CLI becomes the control plane for a stack of generative services rather than the thing doing the generation itself.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2062161629593113066
A neat closed-loop TTS-tooling trick. Engines like AivisSpeech expose an API for dictionary registration, so he delegates the whole thing to Claude Code or Codex: register entries via API, then verify pronunciation by having the engine read text aloud and checking it with a local Whisper CLI, then re-register anything missing, repeating until coverage is complete. Adding that self-verification step is what made unregistered words fill in accurately and automatically. The pattern, agent acts then checks its own output with a second tool, is the quiet backbone of every reliable automation in this roundup.
🗣 User Voice
User Voice
Usage and rate limits are the loudest pain point. People feel limits tightened, sub-agents and parallel workflows drain them fast, and a single rate-limit hit can nuke every concurrent session at once instead of queuing. @doodlestein
Subscription cost is the new anxiety, with per-engineer token spend rivaling salaries and $459/month stacks, which is exactly what's pushing builders to run models locally on a Mac mini via Ollama for near-zero marginal cost. @N01ennn
Setup and maintenance fatigue is real, people spend more time keeping their agent workspace alive (dependencies, broken upgrades, flaky workflows) than actually using it, and they want a finished system, not a science project. @codyplof
Most users badly underutilize Claude Code, treating it like a chatbot instead of configuring CLAUDE.md, memory, skills and sub-agents to run like an engineering team, and the gap between casual and power users is widening fast. @0x0funky
Safety guardrails are now blocking legitimate agentic work, the clearest example being Claude refusing to place a real trade it was authorized to make, which forces workarounds and model-switching and is becoming the real frontier. @itsjcmerlo
Usage and rate limits are the loudest pain point. People feel limits tightened, sub-agents and parallel workflows drain them fast, and a single rate-limit hit can nuke every concurrent session at once instead of queuing. @doodlestein
Subscription cost is the new anxiety, with per-engineer token spend rivaling salaries and $459/month stacks, which is exactly what's pushing builders to run models locally on a Mac mini via Ollama for near-zero marginal cost. @N01ennn
Setup and maintenance fatigue is real, people spend more time keeping their agent workspace alive (dependencies, broken upgrades, flaky workflows) than actually using it, and they want a finished system, not a science project. @codyplof
Most users badly underutilize Claude Code, treating it like a chatbot instead of configuring CLAUDE.md, memory, skills and sub-agents to run like an engineering team, and the gap between casual and power users is widening fast. @0x0funky
Safety guardrails are now blocking legitimate agentic work, the clearest example being Claude refusing to place a real trade it was authorized to make, which forces workarounds and model-switching and is becoming the real frontier. @itsjcmerlo
📡 Eco Products Radar
Eco Products Radar
Products and tools mentioned 3+ times across today's posts:
Codex — 80
OpenClaw — 60
Cursor — 33
Hermes Agent (Nous) — 28
MCP — 20
Mac mini — 18
Skills / SKILL.md — 15
Ollama — 12
Microsoft Scout — 12
OpenCode — 11
Gemini — 10
DeepSeek — 9
Obsidian — 9
GitHub Copilot — 8
Everything Claude Code (ECC) — 7
Notion — 6
Qwen — 5
NotebookLM — 4
Coze — 4
Playwright — 4
ffmpeg — 4
Figma — 3
n8n — 3
ElevenLabs — 3
Products and tools mentioned 3+ times across today's posts:
Codex — 80
OpenClaw — 60
Cursor — 33
Hermes Agent (Nous) — 28
MCP — 20
Mac mini — 18
Skills / SKILL.md — 15
Ollama — 12
Microsoft Scout — 12
OpenCode — 11
Gemini — 10
DeepSeek — 9
Obsidian — 9
GitHub Copilot — 8
Everything Claude Code (ECC) — 7
Notion — 6
Qwen — 5
NotebookLM — 4
Coze — 4
Playwright — 4
ffmpeg — 4
Figma — 3
n8n — 3
ElevenLabs — 3
Comments