Super User Daily: 2026-05-07
The story today is not about Claude Code or OpenClaw being good. It is about how absurd the leverage gap has gotten. One Chinese guy fanning thirteen Claude Code windows across two monitors and quietly running a 200-store-a-month dropship factory for less than 80 dollars a day. A 12-year-old soldering a 12-dollar status light that blinks when his agents finish another OnlyFans payday. Anthropic engineers no longer writing a single line of code. The numbers stopped being a flex and started being structural. Below is what real users actually shipped on May 5, sorted by how much it bends your sense of what one person can do.
@helloitsaustin [Claude Code]
https://x.com/helloitsaustin/status/2051792721871004002
He got married last weekend. So he had Claude Code analyze twelve years of iMessages with his wife, then handed the dataset to Claude Design, and shipped a custom wedding website for the guests in minutes. Trick: on a Mac signed into iCloud, Claude Code can read the chat.db file directly. This is the canonical 2026 weekend project — five minutes of "what if" turning into a personalized data product that would have cost a developer two weeks last year.
@browomo [Claude Code]
https://x.com/browomo/status/2051606666454426091
A Chinese operator runs a thirteen-Claude-Code-window dropshipping factory off a single desk. Six windows on a wall-mounted LG, six on a vertical second monitor, one on the MacBook. Each agent gets a slice of a Shopify store build — catalog, homepage, cart, email sequences, banners, analytics. He charges $800 per finished store, serves around 200 dropshippers a month, burns under $80 a day in API tokens. The system prompt that flips it from assistant to operator: "you are my new founder-engineer." This is the cleanest agent-as-employee economics doc I have seen this year.
@iokawauchi [Claude Code]
https://x.com/iokawauchi/status/2051813812010103293
The actual Meta engineer who built Claude Code is now living in rural Nara, Japan, brewing miso. Ex-Apple iPod engineer in Itoshima making coffee grinders. Ex-Netflix team renovating a closed kindergarten in Yamanashi into a movie theater. The post is a vibe shift document: when the leverage tool you built lets you live anywhere, the geography of high-impact work decouples from the Bay Area. Worth reading less for the gossip and more for the signal of where the next wave of post-FAANG founders is heading.
@sai_borg [Claude Code]
https://x.com/sai_borg/status/2051678530291761201
Sai is dogfooding Claude Code from the phone. A teammate spotted a UI bug — black text on a black button — while away from his laptop. Reproduced it on his phone in iMessage, took screenshots, handed it to Claude Code. PR was up before anyone read the original Slack message. This kills the ticket → repro doc → tomorrow's standup loop. The mobile-to-PR pipeline is the bigger story than the bug itself.
@gagarot200 [OpenClaw]
https://x.com/gagarot200/status/2051582473096831024
Two Japanese students claim ¥40M/month in revenue running an OpenClaw + Hermes agent stack with almost zero human time. Cannot independently verify the number, but the workflow they describe — multiple AI agents handling content, ops, customer interactions on autopilot — is now technically reachable. The reason it ranks here is that even half of that number on this stack should not be possible by 2024 logic. The economics broke a year ago and most people still haven't updated their priors.
@helloitsaustin [Claude Code]
https://x.com/helloitsaustin/status/2051800551080501683
Same wedding website thread, but here is the actual technical seed: if you sign into iCloud on a Mac and use iMessage, there is a chat.db file you can point Claude Code at. Once you know this, every personal-data-as-product project unlocks. Twelve years of relationship data, decade of family group chats, your therapy text journal — all become readable corpus for a one-shot site, retrospective, gift, or memory book. This trick alone will spawn a thousand niche products in the next quarter.
@hiragram [OpenClaw]
https://x.com/hiragram/status/2051483866175877492
He created a fresh user account on a spare MacBook, installed OpenClaw + Codex Native Harness + Computer Use, gave it a Gmail address, and said "make a Tabelog account, make an X account." It did. No credentials of his were exposed. The agent uses the computer the way a junior assistant would. He calls it a fully self-contained PC-using secretary. This is the cleanest demonstration of why Computer Use plus an OS-level sandbox is the actual unlock — not better prompts.
@steipete [OpenClaw]
https://x.com/steipete/status/2051697991266795793
Peter, the OpenClaw creator, added Google Meet support and now his agent Molty is eager to join every call. The micro-update matters because the highest-friction missing surface for personal agents was meeting attendance. Once an agent can hop into Meet/Zoom, take notes, follow up afterwards — the gap between "I have an assistant" and "AI is my assistant" closes. Watch this turn into a feature war between OpenClaw, Google's rumored Remy, and Meta's Hatch.
@QingQ77 [Claude Code]
https://x.com/QingQ77/status/2051630487282565428
book-to-skill: drop a PDF or EPUB technical book into a folder, and a Claude Code skill auto-extracts the structure into a callable skill file — chapter summaries, terminology glossary, design pattern reference. Uses Docling for code-heavy books to preserve tables and code blocks, pdftotext for plain prose. After indexing, you can call /skill-name and Claude pulls from the book directly, no PDF re-reading and no hallucination risk. Single best example today of treating books as agent fuel rather than human reading material.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2051565075614589385
obsidian-second-brain turns an Obsidian vault into a self-rewriting AI second brain that runs as a Claude Code skill. Inspired by Karpathy's LLM Wiki idea but goes further: new content rewrites existing notes instead of appending, contradictions get auto-reconciled, hidden cross-note patterns get extracted into new pages. Includes 31 slash commands, 4 scheduled overnight agents that "close today, reconcile contradictions, synthesize patterns, repair orphaned notes, rebuild index." The vault stops being a dead archive and becomes a compounding asset.
@nickleginza [Claude Code]
https://x.com/nickleginza/status/2051807510336536653
Replaced a $2,000+ commercial shop floor PA system with a $200 build: Raspberry Pi, $35 amp, horn speaker, Claude Code-written control software, browser UI for remote management. Plays whatever they want, when they want. The line "engineers don't buy what they can build" was always aspirational — Claude Code makes it actually true for embedded systems hobbyists.
@ozwxy [Claude Code]
https://x.com/ozwxy/status/2051621555093078412
WEIN / BACKSTAGE Group rolled out Claude / Claude Code to all 240+ members. Monthly company-wide AI training covers Claude Code basics through building actual workflow tools. Members are shipping their own automation tools right after each session. This is the part of enterprise adoption that gets undercounted — not the headline contracts, but mid-sized companies turning their full headcount into builders. Once a 240-person team learns to ship internal tools, the productivity delta against their non-AI peers becomes uncrossable in a year.
@geekbb [Claude Code]
https://x.com/geekbb/status/2051652549938389311
Tiny tool that displays Claude Code and Codex 5-hour and weekly usage limits inside the MacBook notch. Hover for percentages, click to expand, slide between Usage and Cost views, Cost screen estimates spend and token throughput from local session logs. Boring on the surface, but watching limits in real time has the same cognitive effect as watching latency on a trading screen — you start optimizing different decisions when the meter is visible at all times.
@hachiya [Claude Code]
https://x.com/hachiya/status/2051604020813594972
Confirms the Nara miso story — the Claude Code creator literally relocated to a Japanese countryside village and is making fermented soybean paste. Treat it as the same vibe data point as iokawauchi's post above, but the joint coverage from two well-followed Japanese accounts on the same day signals this story crossed into local meme status. Real impact: Nara real estate just got more interesting.
@amenya_nelson [Claude Code]
https://x.com/amenya_nelson/status/2051583766481571904
Built his first AI executive assistant entirely in Claude Code. It reads, sends, and triages email, manages calendar events, creates and shares Google Docs, runs morning briefings, executes arbitrary bash automation. He now books meetings by typing "find a suitable time on my calendar for this event and send them invites." This is a workflow most people pay $100/month for through Superhuman + Reclaim + Otter — replicated in a weekend with one agent harness.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2051666220332794238
Gabor runs 21 specialized agents inside one Claude Code workspace structured exactly like a software company: system analyst, product spec architect, CTO agent, UX flow architect, designers, brand agent, implementation agents, code maintainability agent, performance agent, test architect, product council. Used the setup to ship a hockey rules iOS app to TestFlight solo. The system analyst is the bottleneck — its spec quality determines every downstream agent. The reframe: stop thinking of Claude Code as a chatbox, start thinking of it as an org you are staffing.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2051589237880009026
Six-month build log on running Obsidian as the persistent memory layer for Claude Code via MCP — vault becomes the second brain that thinks back. CLAUDE.md sits in the vault and loads automatically; project context, voice, standards persist between sessions. Morning briefing is a single command that reads the daily note, active projects, inbox, returns three priorities and one decision before noon. After ninety days, the vault stops being a passive archive and becomes a compounding asset every additional note increases the value of.
@yanndine [Claude Code]
https://x.com/yanndine/status/2051713556651913241
Built an entire GTM agentic OS that runs from a single Claude Code terminal session every week. CLAUDE.md encodes ICP, voice, tool stack, hard rules. Skill library handles lead scoring, outreach writing, content generation, client reporting. Sub-agents run heavy parallel tasks and return clean summaries. MCP connects to CRM, inbox, calendar, outreach tools. Hooks ping Slack on completion, gate sensitive files, run quality checks. Modal deployment turns any skill into a live URL the team can trigger. Solo founder GTM stack at parity with what most agencies sell for $5K/month.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2051805717003161602
A Moonside Halo desk lamp wired to Claude Code state. Color shifts when Claude is thinking, writing, done, or errored — peripheral status indicator instead of constant context-switching to check the screen. People are already adapting it to Philips Hue and Govee setups. Could also map color to remaining usage limit. The deeper move is that ambient computing for AI is starting to look like the smart-home build of 2018 — DIYers patching it together before any vendor ships a polished version.
@MushtaqBilalPhD [Claude Code]
https://x.com/MushtaqBilalPhD/status/2051639321032343949
Concrete academic workflow. Feed Claude Code your raw notes plus your style preference and ask for a presentation — it one-shots the slide deck. Drop in lecture notes, get a draft article. After a few iterations, ask Claude Code to turn the workflow into a Skill it can invoke later. From that point forward your style and requirements compound. The pattern beats hiring a research assistant — and it scales to any other professional whose output is mostly artifacts on top of a personal voice.
@kkk_cun [Claude Code]
https://x.com/kkk_cun/status/2051497079705571721
Built a company that uses Claude Code to generate every part of an X (Twitter) growth operation: niche selection, post writing, engagement, product creation, sales funnel, LINE flow, sales copy. Claims it nets ¥1M/month. The granular task list is what matters — every line item used to require a specialist contractor. Eight specialist contractors compressed into one agent stack. The overhead of running a side business just dropped through the floor.
@danieloflagos [Claude Code]
https://x.com/danieloflagos/status/2051615742227878278
Saw a design from another creator he liked, couldn't sleep until he built it in Claude Code. Live link shipped same night. The flow: see → vibe → build → ship in one sitting. Four years ago this was a two-week sprint with a designer and a developer. Now it is a Saturday night and one prompt iteration loop. The barrier between desire and shipping just compressed by orders of magnitude.
@Atenov_D [Claude Code]
https://x.com/Atenov_D/status/2051623265685180587
Plan Mode (Shift+Tab in Claude Code) prevents 80% of errors and almost nobody uses it. Forces Claude to write out every step and ask questions before touching a file. Plus a list of fifteen other commands and settings most users never find: /init builds the architecture cheatsheet, /st shows live token spend, /compact compresses history at 60% memory, /re rolls back without breaking the chat, /loop schedules checks, /hooks pings on long generation finish. Best low-effort optimization checklist of the day for current Claude Code users.
@MattBruenig [Claude Code]
https://x.com/MattBruenig/status/2051687936949653852
Used Claude Code to download the 2022 Survey of Consumer Finances, filter for age over 64, and compute average net worth by elderly net worth decile. Whole thing as a single prompt. This is the kind of data work that took policy researchers a graduate-student-week of pandas wrangling. He shared the prompt verbatim. Anyone reading can verify the number against any government dataset in roughly five minutes now.
@FOLKS_FUMITO [Claude Code]
https://x.com/FOLKS_FUMITO/status/2051468629494104089
Iwai from the band Galileo Galilei is using Claude Code to build Lexicon and Valhalla DSP-style FDN reverbs and Prophet-style synths. He uses them in actual song production and live performance. He plans to perform on his next tour exclusively with synths and reverb plugins he and Claude Code built together. Guitarist coding professional-grade audio plugins because the natural-language-to-DSP layer finally works. The audio production world hasn't fully caught up yet, but it will.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2051673615989670192
Burned through three Binance API keys before figuring out the leak. It wasn't him. Claude Code's Write/Edit tools auto-inject the new file content back into the conversation context — designed to avoid the model reading stale versions, but a disaster for .env files. His three-layer fix: don't let the AI use Write/Edit on .env, use Bash touch + open instead; PreToolUse hook that blocks any sensitive path; the real moat is exchange-side IP whitelist + read-only + withdrawals disabled. Best operational security writeup of the day.
@petergyang [Claude Code]
https://x.com/petergyang/status/2051675536217571703
Mercury VP built a Claude Code system that ingests his meeting transcripts from Granola and cross-references them against his performance review feedback. The system tells him in real time when he is doing the exact thing his last review flagged. He says it keeps him honest at a frequency no human coach could match. Personal ops loop — the kind of self-improvement workflow that used to require an executive coach charging $500/hour and a 40-hour-delay feedback cycle.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2051671692313854454
Plans to run OpenClaw agents in production: auto-create landing page variants, deploy them, run Facebook ads against each, iterate on results to find new proven angles and customer profiles. Marketing experimentation loop fully closed. The piece that makes it real now versus six months ago: Nano Banana-class image models can finally produce reliably high-quality website graphics, so the agent can complete the full pipeline without a human design step.
@AmitLeViAI [Claude Code]
https://x.com/AmitLeViAI/status/2051688430728282210
Three months grinding hundreds of POCs across multiple GPU clusters with five Claude Code sessions running in parallel. Result: an interpretability research finding he believes can actually matter. Solo researcher running a parallel-experiment infrastructure that two years ago required a three-person ML lab. Five sessions in parallel is no longer exotic — it is the default loadout for anyone serious about agentic ML research right now.
@MrKevinRothrock [Claude Code]
https://x.com/MrKevinRothrock/status/2051477568734584866
A bot that monitors and aggregates Russian-language Telegram news channel buzz, fires every thirty minutes, gives him a handle on a chaotic news beat. Built with Railway plus Claude Code. The use case — open-source intelligence aggregation in a domain where machine translation and source curation used to require a junior analyst — is exactly the sort of work that quietly gets automated and never makes it into a benchmark.
🗣 User Voice
User Voice
The room is loud about the same five things. Tokens, tokens, tokens. @adxtyahq, @Parul_Gautam7, @akshay_pachaar, @09pauai and dozens more are pointing at how vanilla Claude Code burns context badly across long sessions; WozCode and Insforge Skills keep coming up as the patches that actually move the cost meter. Then the harness-vs-model debate: @MrAhmadAwais, @0xcherry and @realsigridjin all argue separately that "model X is bad at coding" almost always decomposes into "this harness was tuned for a different model" — closed labs ship server-side fixes silently, open-model harnesses are stuck with the visible tax. Stability churn is the third drumbeat — @badlogicgames left Claude Code after summer 2025 because invisible system-prompt edits kept breaking his workflow, @testingcatalog confirms OpenClaw is shipping a long-term-support release this month for the same reason. Fourth: nobody loves the Codex vs Claude Code coin flip — @kimmonismus reports Codex passed Claude Code on downloads on April 30, but @cyrilXBT, @yanhua1010 and @Astronaut_1216 insist the right answer is run both, with Claude Code as the orchestrator and Codex as the executor. Fifth: control. @ds_nakajima warns against companies giving employees personal Max plans because logging, masking, and prompt blocking are absent — @MariusHobbhahn announces Watcher specifically as MDM/EDR for coding agents, treating them as untrusted by default. The signal: people graduated from "is this useful" to "how do I run this safely at scale."
The room is loud about the same five things. Tokens, tokens, tokens. @adxtyahq, @Parul_Gautam7, @akshay_pachaar, @09pauai and dozens more are pointing at how vanilla Claude Code burns context badly across long sessions; WozCode and Insforge Skills keep coming up as the patches that actually move the cost meter. Then the harness-vs-model debate: @MrAhmadAwais, @0xcherry and @realsigridjin all argue separately that "model X is bad at coding" almost always decomposes into "this harness was tuned for a different model" — closed labs ship server-side fixes silently, open-model harnesses are stuck with the visible tax. Stability churn is the third drumbeat — @badlogicgames left Claude Code after summer 2025 because invisible system-prompt edits kept breaking his workflow, @testingcatalog confirms OpenClaw is shipping a long-term-support release this month for the same reason. Fourth: nobody loves the Codex vs Claude Code coin flip — @kimmonismus reports Codex passed Claude Code on downloads on April 30, but @cyrilXBT, @yanhua1010 and @Astronaut_1216 insist the right answer is run both, with Claude Code as the orchestrator and Codex as the executor. Fifth: control. @ds_nakajima warns against companies giving employees personal Max plans because logging, masking, and prompt blocking are absent — @MariusHobbhahn announces Watcher specifically as MDM/EDR for coding agents, treating them as untrusted by default. The signal: people graduated from "is this useful" to "how do I run this safely at scale."
📡 Eco Products Radar
Eco Products Radar
WozCode — the most-cited tool of the day for trimming Claude Code token waste, mentioned by @akshay_pachaar, @Parul_Gautam7, @adxtyahq, @JaynitMakwana, @heyDhavall.
Obsidian + Claude Code MCP — second-brain stack hit ten distinct mentions including @cyrilXBT, @GitHub_Daily, @sumika45379, @obsidianstudio9, @Stanleysobest, @bridgemindai, @Pluvio9yte. The default knowledge layer for serious users.
Codex CLI — passed Claude Code on download charts April 30 per multiple sources. Most users converge on the dual-stack pattern: Claude Code for design/architecture, Codex for execution.
DESIGN.md / Open Design — open-sourced design system files (@so_ainsight, @0xKingsKuan, @nutlope, @berryxia x2). Eliminates the AI-generated-UI smell.
HyperFrames + Remotion — motion design and video editing as Claude Code skills (@gauravsbuilding, @VaibhavSisinty, @shari1345). Last creative bottleneck dissolving.
Skills marketplaces — Daniel Avila's repo at ~30K stars (@DeepTechTR), claude-code-templates with 1000+ components (@sitinme), @SuguruKun_ai's "Skills 946 collection" with 15.8K stars.
Pi (badlogicgames) — minimalist self-modifying coding agent built as a Claude Code reaction (@Pragmatic_Eng, @GergelyOrosz, @matidotlol, @realsigridjin).
Multi-agent orchestrators — Scion (@geminicli, @JackWoth98), Octogent (@om_patel5 ecosystem), ruflo, multica, RightAgent — all converging on the "claude-code-instances-as-team-members" pattern.
SubQ — 12M context architecture at 5% the price of Opus (@DLKFZWilliam2, @theresanaiforit, @elliotchen100). Architectural unlock, not just scaling.
Watcher — coding-agent MDM/EDR (@MariusHobbhahn). Treats agents as untrusted, ships hard guardrails for security teams.
Ollama Cloud + Claude Desktop — one command swaps Anthropic models for Kimi K2, GLM-5, DeepSeek, Qwen inside Claude Cowork and Claude Code (@yanhua1010, @berryxia, @Lonely__MH, @maxifirtman, @erhanmeydan).
WozCode — the most-cited tool of the day for trimming Claude Code token waste, mentioned by @akshay_pachaar, @Parul_Gautam7, @adxtyahq, @JaynitMakwana, @heyDhavall.
Obsidian + Claude Code MCP — second-brain stack hit ten distinct mentions including @cyrilXBT, @GitHub_Daily, @sumika45379, @obsidianstudio9, @Stanleysobest, @bridgemindai, @Pluvio9yte. The default knowledge layer for serious users.
Codex CLI — passed Claude Code on download charts April 30 per multiple sources. Most users converge on the dual-stack pattern: Claude Code for design/architecture, Codex for execution.
DESIGN.md / Open Design — open-sourced design system files (@so_ainsight, @0xKingsKuan, @nutlope, @berryxia x2). Eliminates the AI-generated-UI smell.
HyperFrames + Remotion — motion design and video editing as Claude Code skills (@gauravsbuilding, @VaibhavSisinty, @shari1345). Last creative bottleneck dissolving.
Skills marketplaces — Daniel Avila's repo at ~30K stars (@DeepTechTR), claude-code-templates with 1000+ components (@sitinme), @SuguruKun_ai's "Skills 946 collection" with 15.8K stars.
Pi (badlogicgames) — minimalist self-modifying coding agent built as a Claude Code reaction (@Pragmatic_Eng, @GergelyOrosz, @matidotlol, @realsigridjin).
Multi-agent orchestrators — Scion (@geminicli, @JackWoth98), Octogent (@om_patel5 ecosystem), ruflo, multica, RightAgent — all converging on the "claude-code-instances-as-team-members" pattern.
SubQ — 12M context architecture at 5% the price of Opus (@DLKFZWilliam2, @theresanaiforit, @elliotchen100). Architectural unlock, not just scaling.
Watcher — coding-agent MDM/EDR (@MariusHobbhahn). Treats agents as untrusted, ships hard guardrails for security teams.
Ollama Cloud + Claude Desktop — one command swaps Anthropic models for Kimi K2, GLM-5, DeepSeek, Qwen inside Claude Cowork and Claude Code (@yanhua1010, @berryxia, @Lonely__MH, @maxifirtman, @erhanmeydan).
Comments