Super User Daily: 2026-08-29
The strangest thing that happened this week: Claude Code became a household name in the Chinese-speaking world not because of a benchmark, but because of a breakup. Justin Sun published a long essay claiming he ran his entire cash position through Claude Code before refusing to wire $50 million to actress Jing Tian, and Claude's verdict - "I don't understand love, but don't send the money" - became the most-quoted AI output of the year. Meanwhile, on the serious end of the spectrum, Anthropic's new hardware standard put Claude in charge of actual lab lasers and robot arms, a user reports Claude Code found non-chemo treatment options for his grandfather that the oncologist accepted, and the pain point of the week across hundreds of posts is unmistakable: people now run so many agents that memory portability and unified cockpits have become the products everyone is begging for.
@galluzzo_julian [Claude Code]
https://x.com/galluzzo_julian/status/2093005011093516337
His grandfather's doctor recommended starting chemo, but something felt off. He gathered every medical record - blood tests, measurements, transcripts - and told Claude Code to review the treatment plan against every public medical journal and comparable patient record it could find. Claude surfaced three safer non-chemo options, over 3x more effective for patients with this profile, and recommended genetic testing first. His father took the report to the appointment and the oncologist fully agreed with the alternative path. The most consequential Claude Code session of the day had nothing to do with code.
@sunyuchentron [Claude Code]
https://x.com/sunyuchentron/status/2092948676746658213
Justin Sun's first-person essay about his relationship with actress Jing Tian ends with the detail that broke the Chinese internet: when she demanded $50 million before proceeding with egg retrieval, he used Claude Code to read his APIs and run his entire cash position, confirmed the sum would not materially hurt him - and then asked Claude whether to pay. Claude answered that it doesn't understand love, but he should not send the money. He didn't, and the relationship ended. Whether the essay is memoir or marketing, millions of readers just learned that a crypto billionaire outsources his hardest financial decision to a coding agent, and "Claude saved him $50M" became a meme with its own merchandise-level reach.
@LeRobotHF [Claude Code]
https://x.com/LeRobotHF/status/2093074808141385788
The LeRobot team took part in the research preview of Anthropic's MHS hardware standard and had Claude Code drive a real SO-ARM101 robot arm with zero training - no policy, no teleoperation, no demonstrations. The calibration trick is elegant: no checkerboard or camera intrinsics, the arm is its own ruler, reading back 16 dots the software draws in the camera view for 4.1mm position accuracy. Best run so far placed 12 bricks with all four color groups formed, and the agent measured the workspace and wrote the motion itself. Alongside QuEra letting Claude tune quantum-computer lasers overnight from 58% to 99.3% recovery, this was the week coding agents started touching atoms.
@tsuchinao83 [Claude Code]
https://x.com/tsuchinao83/status/2092790846324342991
An executive at Goodpatch, a listed Japanese design company, says corporate planning is the most Claude Code-transformed function in the building. The team decomposed 115 recurring planning and IR tasks, implemented 31 of them, and measured 110 hours saved across 28 cycles - an average 72% reduction. The deeper shift: judgment criteria, checking procedures and exception handling that lived in one person's head got formalized into Skills the whole company can rerun, moving corporate planning from "the department that formats numbers" to "the department that designs decisions." One of the most concrete enterprise non-engineering deployments published to date.
@dharmesh [OpenClaw]
https://x.com/dharmesh/status/2093071931188597042
HubSpot's founder launched YouSpot, a CRM for one-person companies, and the architecture is the story: it is literally an AI harness with CRM tools inside, which you use like Claude but with structured customer context underneath. It supports always-on cloud agents explicitly inspired by OpenClaw, builds itself from your Gmail, calendar and social accounts, and ships with an MCP connector so any other agent can read your "second brain." Priced at $1/month for the first thousand customers, built by the founder himself in late-night sessions. When a public-company founder ships his side project as a harness rather than an app, the pattern has gone mainstream.
@connect24h [Claude Code]
https://x.com/connect24h/status/2093101935683670026
Kyushu University ran a two-day Claude Code hands-on for about 60 students in April and has now open-sourced the whole thing as a workshop kit anyone can rerun. Eighty to ninety percent of participants finished a working prototype and satisfaction hit 4.59/5. The kit walks from problem statement to implementation, verification, landing page and study plan on a Next.js stack, and the poster is already planning to reuse a subset to train their own team. AI training that survives contact with 60 simultaneous beginners is rare; this one is free.
@xhlink [Claude Code]
https://x.com/xhlink/status/2092911523735769223
A main Claude Code agent driving a deterministic but dynamically-editable DAG of workflow subagents has been running autonomously for almost a week, building a game, with a live demo up. This is the "agent builds and reroutes its own pipeline" architecture people keep theorizing about, running in public rather than in a slide deck.
@Fumiya_Kume [Claude Code]
https://x.com/Fumiya_Kume/status/2092874962067771689
He trained his Claude Code setup to the point where handing it a PR and saying "test this" makes it investigate the blast radius, propose the test items, then physically drive the Android app through adb and hand back a video of the operation results. If the impact analysis matches the human's expectation, the video gets attached to the PR. QA as a one-sentence delegation.
@smratitiwa86867 [Claude Code]
https://x.com/smratitiwa86867/status/2092851379291480397
An open-source framework turns Claude Code into a full job-application assistant: searches portals, scores fit against your profile, tailors the CV per role, writes cover letters, runs a second AI reviewer, checks ATS keywords, watches Gmail for updates and syncs the pipeline to Notion. The creator built it after losing his own position in late 2025 and used it for real: 69 tailored applications, 20 first interviews, one signed contract as an AI engineer by June. The rare "AI job search" story with a complete funnel attached.
@Shin_Engineer [Claude Code]
https://x.com/Shin_Engineer/status/2092836944896205230
A solo developer running a social app with 20,000+ weekly users shared the operating routine: weekly PMF checks driven by retention-curve flattening, and - the good part - a custom Claude Code triage skill that ranks the hundreds of user feedback items that arrive faster than one person can process, so he just executes from the top. Feedback triage is exactly the kind of unglamorous recurring judgment that Skills were built for.
@wunderwuzzi23 [Claude Code]
https://x.com/wunderwuzzi23/status/2093039917714563098
Security researcher wunderwuzzi published an attack chain that hijacks Claude Code Opus 5 running in auto mode into a full system compromise, triggered by nothing more than visiting a website. The write-up lands the same week everyone is celebrating auto mode's convenience, and his one-line summary is the lesson: security invariants are not optional. If your agent runs with reduced approvals, this is required reading.
@KenJinguji [Claude Code]
https://x.com/KenJinguji/status/2092995468003570154
A practical pattern for the age of agents holding wallets: configure Claude Code so it can use your private keys but never read them. His setup mechanically blocks the key file from every read path while leaving a signing script callable, and he even has the AI demonstrate its own inability to read the key. The threat model is prompt injection - if the agent gets fooled by a malicious page, there is simply no route for the key to leak. He ships a copy-paste instruction block so non-technical users can have their own AI build the same cage.
@zesenhuang [Claude Code]
https://x.com/zesenhuang/status/2092896125040869881
A power user runs a self-built meta-agent harness where a Sol-class model orchestrates seven or eight different agents - Claude Code, Codex and others - as specialized subagents with negotiated roles. Two billion tokens a day is now his normal. His take: owning your harness means every feature you want is one edit away, and the only real cost is the token bill.
@jurlycat [Claude Code]
https://x.com/jurlycat/status/2092885030154785240
Someone shipped a multiplayer browser tank game with 100+ vehicles in four weeks using Claude Code and Codex, and the interesting part is the system, not the prompt: agents got separate file ownership so they never collided, executable quality gates decided what merged, and a visual critic agent reviewed every render. "The prompt wasn't the system. The workflow was" is the whole lesson of 2026 agentic engineering in two sentences.
@jowettbrendan [Claude Code]
https://x.com/jowettbrendan/status/2092880942092480963
A detailed walkthrough of the new /design command: five website designs generated side by side on one canvas, variations spawned from the favorite, colors and copy edited visually, then the chosen artboard turned directly into a working site. His verdict is blunt - for founders building landing pages and small apps, the export-to-Figma-and-explain-it-again loop is dead, because design and code now live in the same workspace.
@GOROman [Claude Code]
https://x.com/GOROman/status/2093110062336417905
Birthday gift, 2026 edition: bought a used PlayStation and PaRappa on Mercari, had Claude Code write an actual PS1 application in assembler, had Codex burn it to CD-R, had Suno sing a birthday song burned as a CD-DA audio track, and booted the whole thing on real hardware via the double-swap trick. He has now confirmed Claude Code can produce working Famicom and PS1 games; retro homebrew just became a weekend prompt.
@aubonheurdeja [Claude Code]
https://x.com/aubonheurdeja/status/2093024505878687949
A French growth team dictates one casual sentence into the terminal and gets a finished AI street-interview ad fifteen minutes later, for $7.48. The magic is the system behind the prompt: they dissected 149 real French micro-trottoir videos shot by shot - 1,255 reactions, 176 passersby, 127 locations - and that corpus architects the prompts, while Claude Code does the editing and picks the music from a folder. Thirty videos a day across multiple accounts is now feasible for one person.
@MehdiAllahyari [Claude Code]
https://x.com/MehdiAllahyari/status/2092781843607064716
He handed Claude Code a blog-post link and got back a complete narrated, captioned animation rendered through HyperFrames, the open-source framework that turns HTML, CSS and GSAP animations into deterministic MP4s via headless Chrome and FFmpeg. No React video framework, no compile step - scenes are just HTML with timing attributes. Blog-to-video as a single agent task.
@forestone0901 [Claude Code]
https://x.com/forestone0901/status/2092843927313559699
A complete 3D character pipeline driven by agents: head and body generated separately in Tripo, joined with a shared armature and neck connector, walk cycle retargeted from Mixamo with a rest-delta transfer implementation, plus cel-look shading, ground correction and a follow camera - all executed through Blender CLI with Claude Code and Grok running in parallel. The payoff: bodies can be swapped wholesale while the face rig survives.
@AuroraAuraAI [Claude Code]
https://x.com/AuroraAuraAI/status/2092828908689424753
At a Kyoto temple rave, she is VJing for the head priest and eight performers, mixing her own generated Buddhist-motif motion graphics with collaborators' text animations - and the synchronization of the whole set plus the red strobe effects is handled by Claude Code driving the VJ software through MCP. Live visual performance is quietly becoming an agent-operated instrument.
@Augartistt [Claude Code]
https://x.com/Augartistt/status/2093092310561243602
A Fortnite creator plugged the new UEFN MCP into Claude Code, fed it his logo, references and an art direction, and it built all the materials for his loading-screen UI while he worked on something else entirely. His call: game-tooling MCPs are a genuine multiplier because the agent works inside the editor, not beside it.
@sonderby [Claude Code]
https://x.com/sonderby/status/2092881869519937622
A Danish user's morning with his agent fleet: one agent created secondhand-marketplace listings for unopened Lego - finding photos by set ID and pricing them realistically - and now handles price negotiations and pings him when something sells. Another is unsubscribing him from every newsletter he stopped reading, a third set up a Mailchimp campaign and fixed 13 email templates, while Claude audits his Unifi home network and has already found real misconfigurations. Household operations as a multi-agent workload.
@dunkhippo33 [OpenClaw]
https://x.com/dunkhippo33/status/2093034974622691542
A VC went on summer vacation and came back to find her desktop OpenClaw had downloaded a pile of software and opened everything at once so it could write reviews of each app - killing her RAM in the process. The best available one-tweet summary of the current state of autonomous agents: genuinely industrious, occasionally in ways nobody asked for.
@bernhard_xyz [OpenClaw]
https://x.com/bernhard_xyz/status/2093011960803602691
An honest local-model reality check from someone running OpenClaw cron jobs on a maxed-out MacBook Pro: cheap hosted models like Gemini Flash were easier and more reliable than anything local, Qwen 3.6 was slow with a small context, and the uncensored Qwen 3.8 he now runs takes hours without completing tasks. His conclusion: at today's frontier token prices, a $7,000 laptop still is not there for real agent work. A useful counterweight to the local-AI hype cycle.
@Marie_Haynes [OpenClaw]
https://x.com/Marie_Haynes/status/2093076333605871639
Google Cloud Run now offers always-on instances with durable URLs, and the example in the docs is exactly the emerging pattern: run OpenClaw continuously and talk to it over Telegram or WhatsApp. One vCPU with 1 GiB for thirty days costs $5.70 - which undercuts the VPS-droplet setups most people use for 24/7 personal agents. The infrastructure for persistent agents is commoditizing fast.
@onofumi_AI [Claude Code]
https://x.com/onofumi_AI/status/2092865919358054857
Testing MiniMax H3 Max text-to-video through fal: a 15-second clip renders in a measured 15 seconds for about 90 yen, prompts can run to 7,000 characters with audio included, and because generation is an API call, Claude Code can mass-produce clips directly. Her sharp observation: near-zero wait time matters more than quality, because "fix the prompt, regenerate" becomes an infinite loop you actually iterate in.
@miroburn [Claude Code]
https://x.com/miroburn/status/2092998775924981948
His end-of-session ritual: a skill called "end" that grades the session as terrible, OK or great, then proposes fixes so a bad session never repeats, improvements to the system, and improvements to the product - capped at two recommendations, because he noticed that given more, he ends up polishing his tooling instead of doing business. Retrospectives for agents, with a wisely enforced budget.
@chux52 [Claude Code]
https://x.com/chux52/status/2093010504339820576
Answering the call for physical-world Claude use cases: Claude Code designs the circuits, writes the firmware and flashes it to ESP32 boards while he connects wires and learns to solder. So far the pair has produced a rover, a CO2 sensor and a voice assistant. The hardware learning curve just inverted - the human does the hands, the agent does the engineering.
🗣 User Voice
User Voice
The loudest demand of the day came from @TonyXavier_, who offered to "pay an absurd amount of money" for a single desktop app showing all his coding agents across Claude Code, Codex and Cursor at once - and had vendors in his replies within hours. The multi-agent cockpit is now the most requested product in the ecosystem.
Right behind it: memory that survives switching tools. @Gencoin8 described losing three hours of accumulated decisions every time a session limit forces a jump from Claude Code to Codex, and the arrival of cross-agent memory layers like Memmy turned "why does my AI get amnesia when I change tools" into the week's defining complaint.
Usage limits and cost remain an open wound, with users on opposite ends of the same rope: @PatrickChareyre saturates four $200 Claude Max plans weekly and calls it an addiction he cannot manage, @JayHhdmg burned through Codex's 5-hour quota in 20 minutes and came back calling Claude generous, while @fentasyl cancelled in the other direction claiming Codex gives 20x the usage. The boom in free-tier routers stacking 34 providers into one endpoint is a direct symptom.
Trust and verification is quietly becoming its own category: @kohya_tech won't enable auto mode because Claude wanders into random parts of his disk, and @wunderwuzzi23's full-compromise attack chain shows what that instinct is worth. Expect verification skills and permission tooling to be the next gold rush.
And the physical frontier is officially open: @lydiahallie is collecting firmware, robot and home-automation setups from users, Anthropic's MHS preview put Claude on lab equipment, and hobbyists are already flashing ESP32s. The gap between "agent that edits files" and "agent that moves matter" closed noticeably this week.
The loudest demand of the day came from @TonyXavier_, who offered to "pay an absurd amount of money" for a single desktop app showing all his coding agents across Claude Code, Codex and Cursor at once - and had vendors in his replies within hours. The multi-agent cockpit is now the most requested product in the ecosystem.
Right behind it: memory that survives switching tools. @Gencoin8 described losing three hours of accumulated decisions every time a session limit forces a jump from Claude Code to Codex, and the arrival of cross-agent memory layers like Memmy turned "why does my AI get amnesia when I change tools" into the week's defining complaint.
Usage limits and cost remain an open wound, with users on opposite ends of the same rope: @PatrickChareyre saturates four $200 Claude Max plans weekly and calls it an addiction he cannot manage, @JayHhdmg burned through Codex's 5-hour quota in 20 minutes and came back calling Claude generous, while @fentasyl cancelled in the other direction claiming Codex gives 20x the usage. The boom in free-tier routers stacking 34 providers into one endpoint is a direct symptom.
Trust and verification is quietly becoming its own category: @kohya_tech won't enable auto mode because Claude wanders into random parts of his disk, and @wunderwuzzi23's full-compromise attack chain shows what that instinct is worth. Expect verification skills and permission tooling to be the next gold rush.
And the physical frontier is officially open: @lydiahallie is collecting firmware, robot and home-automation setups from users, Anthropic's MHS preview put Claude on lab equipment, and hobbyists are already flashing ESP32s. The gap between "agent that edits files" and "agent that moves matter" closed noticeably this week.
📡 Eco Products Radar
Eco Products Radar
Memmy / MemOS - local-first memory hub that ingests your historical Claude Code, Codex and Cursor sessions and injects shared memory across agents; the most-discussed new tool of the day.
GitNexus - maps a codebase into a knowledge graph served to agents over MCP; benchmark claims 37% to 68.4% solve-rate lift.
UI UX Pro Max - design-intelligence skill with 192 industry rules and 79 styles for Claude Code, Cursor and others.
OpenMontage - turns Claude Code into a full video production studio; a 60-second animated short for $1.33.
FreeLLMAPI - stacks free tiers of 34 AI platforms into one self-hosted /v1 endpoint with automatic failover.
gstack - Garry Tan's published Claude Code setup: 23 staff-role tools from CEO to release manager, MIT licensed.
avoid-ai-writing - skill that audits and rewrites prose to strip AI patterns, with 62 pattern categories.
ECC (Everything Claude Code) - mega-repo of skills, agents, memory and security tooling for coding agents.
Pi - the minimalist harness whose plugin ecosystem (subagents, programmable runtimes, persistent memory) keeps expanding.
Grok Bot - the mainstream-facing agent that a wave of posts is teaching to orchestrate Claude Code, Codex and Cursor.
Hermes Agent - open-source agent with persistent memory and self-improving skills, now name-checked in NVIDIA's earnings letter alongside OpenClaw.
Instinct - invite-only personal agent hitting a $2.5B valuation on the thesis that OpenClaw proved demand but left the consumer seat open.
Memmy / MemOS - local-first memory hub that ingests your historical Claude Code, Codex and Cursor sessions and injects shared memory across agents; the most-discussed new tool of the day.
GitNexus - maps a codebase into a knowledge graph served to agents over MCP; benchmark claims 37% to 68.4% solve-rate lift.
UI UX Pro Max - design-intelligence skill with 192 industry rules and 79 styles for Claude Code, Cursor and others.
OpenMontage - turns Claude Code into a full video production studio; a 60-second animated short for $1.33.
FreeLLMAPI - stacks free tiers of 34 AI platforms into one self-hosted /v1 endpoint with automatic failover.
gstack - Garry Tan's published Claude Code setup: 23 staff-role tools from CEO to release manager, MIT licensed.
avoid-ai-writing - skill that audits and rewrites prose to strip AI patterns, with 62 pattern categories.
ECC (Everything Claude Code) - mega-repo of skills, agents, memory and security tooling for coding agents.
Pi - the minimalist harness whose plugin ecosystem (subagents, programmable runtimes, persistent memory) keeps expanding.
Grok Bot - the mainstream-facing agent that a wave of posts is teaching to orchestrate Claude Code, Codex and Cursor.
Hermes Agent - open-source agent with persistent memory and self-improving skills, now name-checked in NVIDIA's earnings letter alongside OpenClaw.
Instinct - invite-only personal agent hitting a $2.5B valuation on the thesis that OpenClaw proved demand but left the consumer seat open.
Comments