Super User Daily: 2026-06-06
The center of gravity this week is not 'AI writes code', it is 'a non-coder runs a business on top of an agent'. The cases that stand out are a solo accountant keeping the books for sixty companies, a lawyer freezing his contract-review judgment into a skill, an investor cloning his own taste on pitch decks, and a marketer pointing parallel agents plus a refuter at a Meta Ads slump. Underneath all of it runs one loud anxiety, the token meter, with a buy-a-Mac-Mini-and-cancel-Anthropic argument echoing in thread after thread, even as somebody else casually burns two-point-seven billion tokens to build a multiplayer game in a day. The split is getting sharp: half the room is trying to spend less, the other half has figured out that spending more is the whole point.
@kandmybike [Claude Code]
https://x.com/kandmybike/status/2062340067583295747
A solo CPA quietly runs the books for sixty client companies by himself, and his whole back office is Claude Code. He treats it as an execution unit: it handles bookkeeping, drafts client correspondence, and gathers the scattered information a tax accountant normally chases down by hand. This is the kind of one-person-firm-that-acts-like-twenty story people keep predicting, except it already shipped. The accountant isn't writing code, he's running a practice.
@lawyer_j_kawano [Claude Code]
https://x.com/lawyer_j_kawano/status/2062349374592889259
A practicing lawyer took the exact procedure he uses to review contracts and turned it into a Claude Code skill, then open-sourced it. It covers NDAs, outsourcing agreements, software development and sales contracts, with sixty-plus review angles and over a hundred and eighty standard clause examples baked in. The output isn't a vague summary, it's a list of flagged issues plus a redlined Word version with tracked changes ready to send. This is the real shape of legal automation: not a chatbot that answers law questions, but a senior lawyer's checklist frozen into a file.
@takechan_lawyer [Claude Code]
https://x.com/takechan_lawyer/status/2062681277648814284
Another lawyer hit a wall with the Google Docs API not handling numbered lists and structure well, so he routed around it: let Claude Code generate a proper docx with the right layout and numbering, then convert that to Google Docs. He wrapped the whole thing into a reusable skill. It's a small story but a telling one, the interesting work in legal tech right now is people patching the boring gaps between tools, not building grand platforms.
@nikunj [Claude Code]
https://x.com/nikunj/status/2062659649732825549
This one is closer to cloning a person than automating a task. He pulled two hundred-plus of his own Granola meeting notes from founder pitches, distilled them down to about fifty-three, and built a Claude Code skill called Nock that mimics how he himself judges a pitch deck. Then he tuned it against real decks until its take matched his. An investor's taste is supposed to be the unteachable part, and he's grounding a skill in his actual recorded conversations to teach it anyway.
@0xMoysei [Claude Code]
https://x.com/0xMoysei/status/2062603548479852796
He scraped roughly fifty thousand Google Maps leads, dentists in Austin, in about thirty minutes for under a dollar, all on his existing Claude subscription. One skill file did the orchestration, subagents fanned out to hunt and verify the emails. The headline isn't the scrape, scrapers are old. It's that the marginal cost of fifty thousand verified leads is now lunch money and half an hour, which quietly breaks the pricing of every lead-gen tool that charges per contact.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2062519355267846449
He runs go-to-market at ColdIQ, a seven-million-ARR shop, entirely inside Claude Code, and he showed the wiring: a repo with thirty-two hooks, five agents, eighteen prompts and six skills connecting Claude into Apollo, Instantly, Gmail, Slack and the CRM. It enriches leads and writes sequences on live data. This is what a real GTM stack looks like when the seams disappear, the CRM stops being a place you visit and becomes a thing your agent operates.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2062330304829010200
He built a Meta Ads diagnostic in Claude Code that doesn't just answer 'why did performance drop' with one guess. It spins up parallel agents, each chasing a different hypothesis, then a refuter agent goes adversarial and kills the theories that don't hold, leaving a ranked diagnosis. That's the same generate-and-refute structure the research labs use, pointed at a paid-media problem. Most marketers are still asking the model one question and trusting the first answer.
@GaoShanghua [Claude Code]
https://x.com/GaoShanghua/status/2062601452674277533
ToolUniverse is a plugin that drops two thousand-plus life-science tools and a hundred and twenty-plus research skills into Claude Code. The commands are the interesting part: cross-validate a claim across three or more sources, sweep fifteen-plus literature indexes, and ground every answer in a real citation. This is autoresearch wearing a lab coat, the bottleneck in science was never typing, it was the hours of cross-referencing, and that's exactly the part being collapsed.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2062395168272060514
A PhD geophysicist in Denmark built a Claude Code pipeline that reads a job posting, writes a tailored CV in LaTeX, drafts a cover letter, and has a second AI agent review the whole package before it goes out. He built it for his own search, open-sourced it, and it picked up four hundred and eighty-nine stars and two hundred and seventy forks because everyone job-hunting wants exactly this. The reviewer-agent step is the tell, people have internalized that one agent producing and another checking beats one agent doing both.
@mamiya_afi [Claude Code]
https://x.com/mamiya_afi/status/2062410281914110000
A self-described complete beginner says he's generated over sixty-five million yen in sales across roughly five months by pairing Claude Code with consistent posting on X. Take the exact number with the usual grain of salt, but the pattern is real and recurring this week: people with no coding background using these agents as the production engine behind a one-person internet business. The skill that's paying off isn't programming, it's knowing what to build and ship.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2062499588838228166
A guy stopped paying web designers eighteen hundred dollars a landing page and built the same caliber of site with Claude Code in one afternoon for under seventy dollars. The craft was in the brief, not the code: he fed it five reference screenshots, forced it to ask seven questions first, then pushed it toward one clear visual direction instead of accepting the generic AI-looking default. First pass looked solid, the second made it feel expensive. The edge isn't 'AI builds websites', it's that one person can brief, critique, polish and launch in an afternoon.
@obsidianstudio9 [Claude Code]
https://x.com/obsidianstudio9/status/2062353647946899952
He documented a four-tool research stack worth copying: Claude Code as the orchestrator, a Skill Creator, NotebookLM doing the heavy analysis server-side to save his own tokens, and Obsidian as the persistent memory. Feed it ten YouTube videos and roughly six minutes later you get a NotebookLM note, the analysis, an infographic and a markdown vault entry. Offloading the expensive analysis to NotebookLM to spare local tokens is a clever cost move most people miss.
@kaz_photon [Claude Code]
https://x.com/kaz_photon/status/2062501182539698492
He built a Claude Code command, slash ai-video plus a theme, that auto-researches popular posts on X, writes a script as structured JSON, and renders a sixty-second Apple-Event-style video with background music through Remotion. One line in, a finished promo video out. This is the content-factory pattern maturing, the human picks the topic and the whole research-to-render pipeline runs itself.
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2062500643818873253
He leaves VoiceClaw Realtime running on his desk all day and just talks to it, using his voice to drive OpenClaw and Hermes in the background: brainstorming, querying his CRM, updating dashboards, checking his agenda. When he steps out, he keeps going from his phone and Apple Watch. This is the ambient-agent future people sketch on slides, except it's somebody's actual Tuesday, the computer became a thing you talk at while doing other work.
@iamlukethedev [OpenClaw]
https://x.com/iamlukethedev/status/2062605267129262116
After six months as one of OpenClaw's loudest supporters, he finally tried Hermes and rebuilt his Jira-to-agent-to-GitHub workflow in it. The OpenClaw version took about seven hours to get right, the Hermes version took one, and on day one it closed twelve Jira tickets in under three hours, work he estimates would have eaten two-plus months of engineering. What moved him wasn't a feature OpenClaw lacked, it was that Hermes suggested edge cases and protections while he built, and worked on the first pass.
@gagarot200 [Claude Code]
https://x.com/gagarot200/status/2062423289642012901
In a single day he built a playable multiplayer League-of-Legends clone, christened LMAO, with Claude Code and Opus 4.8: TypeScript, React, Canvas, PartyKit, a separate subagent per champion, all wired through the ultracode workflow. The number that matters is the meter, two-point-seven billion tokens, fifteen and a half million of them output, the rest mostly cache. This is what 100X intelligence physically looks like, you don't get a playable multiplayer game in a day without spending tokens like water.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2062466899922051484
A developer handed Claude Code a thousand-dollar budget and went to bed. He woke up to a production platform live on Cloudflare: four teams of agents that had split the scope and run in parallel, four Workers deployed, three hundred and eighty-seven tests all passing, seven thousand five hundred lines written and committed. While his colleagues were still arguing architecture in Notion. The one decision he changed was to stop writing code and become the person who assigns work to agents.
@fkysly [Claude Code]
https://x.com/fkysly/status/2062381984161374350
Short but striking: he reports running a hundred and fifty-one agents simultaneously with Claude Code's new dynamic workflow, calling it a personal record. No grand workflow attached, just a raw data point on where the ceiling is moving. A year ago the conversation was about one good agent, now power users are casually clearing a hundred and fifty in parallel.
@shannholmberg [Claude Code]
https://x.com/shannholmberg/status/2062652746508173796
He laid out how he's building an agent company inside his marketing agency, and the structure is the lesson: a company brain ingested with transcripts, past campaigns and strategy docs, an orchestrator Hermes agent on top, then department verticals, specialist agents, and narrowly-scoped sub-agents under those. Each client gets an isolated pod so context never bleeds across accounts. His repeated point is that narrow scope, not more tools, is what makes an agent actually good, a vague marketing agent just produces vague output faster.
@fujibee [Claude Code]
https://x.com/fujibee/status/2062336709904490601
He shipped a tiny bash-plus-SQLite messaging layer so Claude Code and Codex instances can talk to each other directly, instead of using the human as a copy-paste relay. To prove it, he had two Claude Code instances autonomously play tic-tac-toe against each other, and the repo went from five stars to three hundred-plus in a week. It's a small hack pointing at a big gap, we have great single agents and almost no plumbing for them to coordinate without a person in the middle.
@Charles77xixi [Claude Code]
https://x.com/Charles77xixi/status/2062396668662993091
He got tired of Claude Code reading CLAUDE.md and Codex reading AGENTS.md and the two drifting apart, so he made a single source of truth at tilde-slash-dot-agent-harness and symlinked both files to it. Now skills, memory and rules stay in sync across both agents automatically. It's a five-minute fix, but it's the practical answer to a complaint half the community is voicing this week.
@sitinme [Claude Code]
https://x.com/sitinme/status/2062350373139120487
CodeGraph pre-builds a tree-sitter map of your codebase into local SQLite so Claude Code, Cursor and Codex stop burning tokens re-reading files every time they explore. The numbers he cites are blunt: sixty-three percent fewer tokens, eighty-two percent fewer tool calls, response time on some tasks cut from over a minute to seconds. As the token meter becomes the thing everyone watches, the products that win are the ones that make the agent read less, not think more.
@r3nzsec [Claude Code]
https://x.com/r3nzsec/status/2062541708907475111
On the security side, IRFlow Timeline collects and normalizes your local AI-assistant usage history, Claude Code, ChatGPT Desktop, Cursor, Copilot, Codex, into one unified forensic timeline, plus an AI Secret Hunt that finds secrets people pasted into their assistants. This is a new category quietly forming: incident response for the AI tools themselves. Every credential someone dropped into a chat box is now a finding waiting to happen.
@bepituLaz [Claude Code]
https://x.com/bepituLaz/status/2062422072832200949
His actual workday: the boss files a request card in Basecamp, he adds context in the comments, then Claude Code on slash-loop picks up all the tasks and opens a PR, while he leaves for a yoga class. That's the loop people keep theorizing about, running on a normal job at a normal company. The work happens inside an existing project-management tool, no new platform, no human babysitting the run.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2062512271646171235
He plugged MiniMax M3 into Claude Code, pasted in a Dribbble design, and got production React back in one session. At his agency he's now replaced Opus 4.8 with M3 for about eighty percent of coding tasks, at a fraction of the cost. This is the cost rebellion getting concrete, people aren't abandoning the harness, they're keeping Claude Code and swapping the expensive model out from under it for the routine work.
@MLWhiz [Claude Code]
https://x.com/MLWhiz/status/2062324749360111679
He posted the full working config to run Claude Code against a local Ollama model through LiteLLM, the exact config.yaml and settings.json, and then, refreshingly, the honest result: two tokens in six minutes on a forty-eight-gigabyte MacBook Pro. The setup works, the hardware doesn't. It's a useful counterweight to all the buy-a-Mac-Mini-and-cancel-Anthropic threads, local is real but the local-runs-everything dream is still mostly aspiration on consumer machines.
@dunik_7 [Claude Code]
https://x.com/dunik_7/status/2062458191586181204
His thread crystallizes the cost anxiety running through this whole week. A developer posts a hundred-and-seventy-dollar Claude Code bill from ten days, someone replies that they bought a Mac Mini M4 and haven't paid Anthropic since, and the dev community runs the math: a stacked cloud setup at around four hundred and fifty-nine dollars a month versus a one-time six-hundred-dollar box and three dollars of electricity, one env var to redirect Claude Code at local. The point isn't that everyone should switch, it's that enough people are doing the arithmetic out loud that it's reshaping behavior.
@mike_hj_w [Claude Code]
https://x.com/mike_hj_w/status/2062374857732702264
He couldn't find an X Pro / TweetDeck replacement he liked, so he built and open-sourced his own with Claude Code in about twenty minutes plus an hour of fixes, Windows packaging, persistent local config, a Grok-page back-button bug, mouse side-button navigation, and shipped Windows and Mac v1.0.0. This is the quiet shift under all the hype: when a missing app is a twenty-minute build, you stop waiting for someone to make it and just make it.
@pastel_orbit [Claude Code]
https://x.com/pastel_orbit/status/2062550607416943052
A game developer who hand-wrote his own Minecraft-style voxel renderer over years had long shelved one feature, smoothing the voxel edges, because reimplementing it meant rebuilding context he'd forgotten. With a well-formed harness, thirty minutes of planning and thirty minutes of edits-on got it working on the first Unity run. His note is the real insight: the joy was sharper precisely because it was his own engine, max resolution of understanding meant max payoff from the assist.
@omarsar0 [Claude Code]
https://x.com/omarsar0/status/2062553527730540611
He got hooked on generating harnesses on the fly, reverse-engineered Dynamic Workflows into his own agent orchestrator, then built an HTML dashboard to monitor tasks, metrics and reports. The use cases he's landing aren't coding: branching and parallel deep research with verification, session mining across all his past agent runs, bug hunting, triage, fact-checking, LLM councils. His read is that dynamic workflows are becoming a core primitive, and the exciting part is it extends well past code into business, science and research.
@AuroraMar1eL [Claude Code]
https://x.com/AuroraMar1eL/status/2062574446804869168
This one's wrapped in hype but the substance is real: an open-source Personal AI Operating System built on Claude Code, twenty-two thousand hours of dev work and six thousand logged sessions behind it, twelve thousand stars. No embeddings, no vector database, every memory and decision lives in plain markdown you can read with cat and version with git, split into work, knowledge, people and learning memory. A code-enforced hook called ContainmentGuard physically blocks sensitive data from leaving its zone. The design philosophy, readable plaintext memory over opaque vectors, is the part worth stealing.
@dionisiodev [Claude Code]
https://x.com/dionisiodev/status/2062630781881602319
To close on the human cost: he reports someone spent seventy-six thousand reais on Claude Code building a SaaS nobody will use, and the punchline is that the guy's girlfriend, reading the bank statement, assumed he'd bought an engagement ring. It's a joke and a warning in one. The same meter that makes 100X intelligence possible will happily let you incinerate a small fortune on a product with no users.
🗣 User Voice
User Voice
The loudest signal this week is cost. Anthropic moving subscription tokens to API billing on June 15 lit a fuse, and the timeline is full of model-routing tricks, local-Mac-Mini math, and token leaderboards. As @dunik_7 's thread shows, people are doing the arithmetic out loud and acting on it.
Config fragmentation is a real, fixable irritation. Claude Code reads CLAUDE.md, Codex reads AGENTS.md, and users like @Charles77xixi are symlinking the two into one source of truth while @petergyang openly asks for an industry standard.
Memory is still the unsolved layer. Agents forget project context every session, pushing people toward Obsidian, plain-markdown memory and a wave of tools like Memanto and Micro.
Honesty about limits matters. @MLWhiz publishing a local setup that produces two tokens in six minutes is more useful than ten hype threads, and the community rewards that candor.
The frontier of desire is parallelism and autonomy: dynamic workflows, slash-loop, @fkysly casually running 151 agents. People don't want a better chatbot, they want more agents running longer with less babysitting.
The loudest signal this week is cost. Anthropic moving subscription tokens to API billing on June 15 lit a fuse, and the timeline is full of model-routing tricks, local-Mac-Mini math, and token leaderboards. As @dunik_7 's thread shows, people are doing the arithmetic out loud and acting on it.
Config fragmentation is a real, fixable irritation. Claude Code reads CLAUDE.md, Codex reads AGENTS.md, and users like @Charles77xixi are symlinking the two into one source of truth while @petergyang openly asks for an industry standard.
Memory is still the unsolved layer. Agents forget project context every session, pushing people toward Obsidian, plain-markdown memory and a wave of tools like Memanto and Micro.
Honesty about limits matters. @MLWhiz publishing a local setup that produces two tokens in six minutes is more useful than ten hype threads, and the community rewards that candor.
The frontier of desire is parallelism and autonomy: dynamic workflows, slash-loop, @fkysly casually running 151 agents. People don't want a better chatbot, they want more agents running longer with less babysitting.
📡 Eco Products Radar
Eco Products Radar
Codex — the permanent reference point. Mentioned in nearly every workflow, increasingly the cheaper second engine people pair with Claude Code rather than abandon it for.
Hermes Agent — Nous Research's local-first agent. Shows up across agency stacks, Jira automations and voice setups; the open, runs-on-your-machine alternative people keep reaching for.
Ollama + Mac Mini (local) — the cost-rebellion stack. One env var redirects Claude Code to local hardware; cited constantly, though @MLWhiz 's honest benchmark shows consumer machines aren't there yet.
MiniMax M3 — the budget model of the week, swapped in under Claude Code for routine coding at a fraction of Opus 4.8's cost.
Dynamic Workflows — Anthropic's on-the-fly harness generator; treated as the biggest primitive since Skills, powering the 100-plus-agent runs.
Memanto / Micro — memory layers bolted onto Claude Code, Codex and OpenClaw to fix the forgets-everything problem.
Codex — the permanent reference point. Mentioned in nearly every workflow, increasingly the cheaper second engine people pair with Claude Code rather than abandon it for.
Hermes Agent — Nous Research's local-first agent. Shows up across agency stacks, Jira automations and voice setups; the open, runs-on-your-machine alternative people keep reaching for.
Ollama + Mac Mini (local) — the cost-rebellion stack. One env var redirects Claude Code to local hardware; cited constantly, though @MLWhiz 's honest benchmark shows consumer machines aren't there yet.
MiniMax M3 — the budget model of the week, swapped in under Claude Code for routine coding at a fraction of Opus 4.8's cost.
Dynamic Workflows — Anthropic's on-the-fly harness generator; treated as the biggest primitive since Skills, powering the 100-plus-agent runs.
Memanto / Micro — memory layers bolted onto Claude Code, Codex and OpenClaw to fix the forgets-everything problem.
Comments