Super User Daily: May 27, 2026
Two stories ran side by side this day and they point in opposite directions. One is the cost panic: the Microsoft-cancels-Claude-Code, Uber-burned-its-whole-year-budget, Nvidia-says-compute-costs-more-than-payroll news cycle that dominated every timeline. The other is quieter and far more interesting. While enterprises freak out over the meter, solo operators are pointing these agents at things that have nothing to do with writing code, and walking away with real money. Accountants clearing 8,000 receipts in two days. Faceless YouTube channels printing $9,000. People who can't code rooting handheld consoles. The pattern is clear: the value is moving away from "help me write a function" and toward "run my business while I sleep." Here are the people actually doing it.
@LoganTGott [Claude Code]
https://x.com/LoganTGott/status/2058890124545466860
He spent six months killing his manual LinkedIn workflow and rebuilding the whole thing inside one Claude Code project. Profile rewrites, weekly posts, lead-magnet funnels, outbound DMs, and the clever part: he feeds in 10 sales-call transcripts and gets 7 posts out per session, so the content is mined from real conversations instead of invented. He claims 500+ calls booked off the system. This is the non-coding play in its purest form, a marketing department collapsed into a prompt library.
@Jeanscpa [Claude Code]
https://x.com/Jeanscpa/status/2058805590571483632
A certified accountant launched an "AI implementation advisor for accounting firms" service built around Claude Code, and the proof case is the kind of number that ends arguments. freee MCP plus Taxsys plus Claude Code chewed through three cardboard boxes of receipts, roughly 8,000 of them, doing data entry and duplicate-checking in about two days. This is the work nobody wants to do, done by an agent, and he's now selling the setup to other firms. The most boring corner of finance is exactly where agents pay for themselves.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2059000366721577189
A faceless cat channel pulled around $9,000 in three months. No camera, no editing timeline, no skill. A Mac Mini, Claude Code, and one weekend of setup: stock footage stitched to a voiceover, long-form animated niche content, ChatGPT used to validate names. The honest catch he includes is that it took weekly shipping for five months before it hit. That's the part the hype tweets cut, and it's the part that actually matters.
@ComagerTon79278 [Claude Code]
https://x.com/ComagerTon79278/status/2058846332262215996
He's running CrowdWorks, a Japanese freelance platform, on semi-autopilot with Claude Code. Landed 12 gigs in a month, pushed earnings past 80,000 yen, and says he just tells Claude Code "handle it" and all twelve jobs together take under an hour. He keeps it semi-automatic for now out of caution but is moving toward full automation. The freelancer who outsources the freelancing to an agent is a genuinely new economic creature.
@Maruo_0314 [Claude Code]
https://x.com/Maruo_0314/status/2058836040677560415
He built a YouTube Shorts factory on Claude Code's parallel sessions: eight subagents running at once handle script, image, narration, video, and scheduled posting end to end. He claims about seven minutes per video and ten a day, against the roughly five hours a human editor burns per clip. Zero filming, zero editing skill. Whether or not every number holds, the architecture, a swarm of subagents each owning one stage of a content pipeline, is the real lesson.
@hikarun_agi [Claude Code]
https://x.com/hikarun_agi/status/2059048225395671450
A video-automation practitioner reports the hard outcomes: a 90% cut in editing labor, fully automated clip production that has repeatedly cleared 100k+ views, and 10k TikTok followers in a single month from niche AI video creatives. He pairs Claude Code skills with an agent harness and an Obsidian memory system so the video operation runs itself. The numbers are the point here, this is creative output measured like a factory line.
@ZayvenKnox [Claude Code]
https://x.com/ZayvenKnox/status/2058750032418164812
A Polymarket trading bot built across four repos on roughly $25/month. Claude Code wrote a function that scans 14,000 wallets in four minutes to surface 47 high-win-rate whales, then runs three consensus agents (arbitrage, convergence, whale-copy) sharing one wallet and exiting before the whales do. He claims $200 to $14,300 in 27 days at a 74% win rate, hands-off. Treat the returns with suspicion, but the design, an autonomous loop of agents trading real money, is exactly where the frontier is pushing.
@Meta8Mate [OpenClaw]
https://x.com/Meta8Mate/status/2058912037284319492
This is what heavy looks like. Two months on OpenClaw across four VPS in Singapore, Melbourne, Seoul and Sydney, fixed spend around $1,200/month, regularly burning 60%+ of a 20x MAX quota each week and over 300% combined across four accounts. Out of that he built two websites from scratch, trading tools (support/resistance, K-line, meme-coin radar), auto-trading strategies, and daily research reports and morning briefings. When someone is voluntarily spending this much, the token bill isn't a cost, it's the engine.
@shedntcare_ [OpenClaw]
https://x.com/shedntcare_/status/2058807825166668084
An OpenClaw plus Arcads bot running multiple Instagram accounts, each a distinct AI persona, all posting automatically and funneling traffic to his SaaS. The loop: OpenClaw researches trending topics, writes conversion-optimized scripts, Arcads turns them into AI-actor videos, and the system auto-posts across accounts with sign-up CTAs, pulling thousands of views a week. A one-person growth team that runs 24/7 and never asks for a raise.
@AYi_AInotes [OpenClaw]
https://x.com/AYi_AInotes/status/2058849167871443029
A Xiaohongshu (RED) auto-operation skill for OpenClaw and Hermes that deliberately uses browser automation rather than the API, mimicking real human clicks to keep ban risk low. You tell your "lobster" things like "analyze my feed," "give me five post topics today," "replicate this viral note," "reply to the latest comments," and it runs the whole loop: reads account data, deconstructs viral posts, generates content plus covers, auto-publishes and auto-replies, saving everything to a Markdown knowledge base. Non-coding social ops, fully delegated.
@Web3GameMaster [OpenClaw]
https://x.com/Web3GameMaster/status/2058928820313280700
In China you can now order McDonald's directly through OpenClaw via the official McDonald's MCP, and it runs on Phala Network inside a TEE so payment and order data stay in an isolated enclave, verifiable on-chain. It's a small thing, ordering a burger, but it's the first clean example of an agent completing a real consumer transaction with privacy guarantees baked in, off an open-source template anyone can fork.
@nikunj [Claude Code]
https://x.com/nikunj/status/2058783316753686558
A genuinely useful technique: point Claude Code at a browser harness or Playwright, click around a website while it sniffs the network requests, and let it infer the API structure, auth scheme, and rate limits. Suddenly sites with "no public API" become programmable. He uses it for scraping, monitoring, and intel-gathering jobs. This is the kind of thing that used to take a reverse-engineering specialist an afternoon, now done conversationally.
@Muskanjain0401 [Claude Code]
https://x.com/Muskanjain0401/status/2058799213975396814
She built a job dashboard for a friend hunting at YC companies: 1,371 roles ranked across every YC company from the last two years, catching real-time intent signals (posts under 30 days old, fresh funding, founder hiring tweets, 407 caught in five days), 853 roles with comp listed, all tiered by reply probability. The stack is a tour of the current toolbox: Claude Code, Perplexity API, Firecrawl, MiniMax 2.7, Unipile. A finished, genuinely useful product built as a favor.
@bodryachog [Claude Code]
https://x.com/bodryachog/status/2058789975559217420
The enterprise reality check the hype needs. He migrated 100,000 lines from Java GraalVM to TypeScript Bun, in production since February with zero prod incidents. A follow-up task, take Jira ticket DH20-XXXX, implement it, add unit and integration tests, was done through Claude Code with the Jira MCP attached in about 90 minutes, most of which he spent watching infrastructure spin up rather than writing anything. This is what it looks like when agents are actually load-bearing in a real codebase.
@PeterMcCormack [Claude Code]
https://x.com/PeterMcCormack/status/2058715947695743138
The Bitcoin podcaster replaced Squarespace for his media company in three days, working alone. Seven websites were on Squarespace; by end of summer there will be none. It's a one-liner, but it's a media business owner with no dev team deciding the cheaper, more flexible path is to have an agent rebuild his whole web presence. Multiply that decision across every small media company and you see where a lot of SaaS revenue is about to go.
@luizpersechini [Claude Code]
https://x.com/luizpersechini/status/2058916718559265039
On a hungover Sunday he and "my friend Claude Code" hacked an R36S handheld console using an Ethernet-USB adapter, got root access, optimized the system, stripped bloat, installed Final Fantasy VII, and bumped the VRAM. He stresses he had no programming knowledge and no direct device access. Hardware tinkering, the most intimidating corner for non-coders, walked through conversationally by an agent.
@lagerskoy [Claude Code]
https://x.com/lagerskoy/status/2058890392976634144
A walkthrough where Unity 6 goes from an empty scene to working driving physics, a camera system, game objects, materials, and a playable prototype in minutes, all via Claude Code talking to Unity's native MCP server. Game development that used to be a six-hour slog of boilerplate collapses into a few minutes of direction. The MCP server is the quiet hero, it's what lets the agent actually drive the engine instead of just suggesting code.
@clairevo [Claude Code]
https://x.com/clairevo/status/2058937544952496601
She points out that the people building Claude Code and Cowork inside Anthropic do the most creative non-coding things with it. Felix Rieseberg, the eng lead, built a 3D model of his new house to plan a move, set up Claude Code to track promises he makes on X, and made a physical "Claude Code buddy button" for $20. When the builders themselves use the tool for moving house and keeping their word, that tells you the surface area is way bigger than coding.
@obsidianstudio9 [Claude Code]
https://x.com/obsidianstudio9/status/2058843041629430043
A fully self-running "personal OS": Obsidian plus Claude Code plus n8n, where the morning briefing, daily capture processing (an inbox.md that gets auto-classified), and the weekly review all fire on their own from the vault, with errors logged to an errors.md. He says the LLM cost is minimal, the initial build took half a day, and from week two it essentially runs itself. This is the dream a lot of people are circling, your notes app quietly becoming an operating system for your life.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2058768136921616761
A content-syndication loop using nothing but an X subscription: Claude Code calls Hermes Agent, which searches X for his own articles, pulls the full text, saves it as Markdown, then reposts and adapts it to Substack and Note. Every step is wrapped as a Skill, so no Chrome extension and no scraping headaches. The interesting move is Claude Code orchestrating a second agent (Hermes) as a tool, agents calling agents to get around platform walls.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2058940930447867951
He collapsed AI-agent eval setup from a two-week chore into an afternoon: build the agent, instrument it with a skill, ask Claude to propose the eval, and under an hour later you have a measured, traceable PM agent with priority-scoring evals. Then the good part, he feeds eval failures into a loop skill on a cron job, so every day the agent rewrites its own prompts, generates new traces, and improves its own evals without him touching it. The self-improving loop, pointed at a real product job.
@mfigge [Claude Code]
https://x.com/mfigge/status/2058706408569438612
The honest counterweight to all the magic. He tried to edit a video entirely through Claude Code, no Premiere, no After Effects, just natural-language prompts. A 15-minute traditional edit became a four-hour, token-heavy, "terrible UX" ordeal. Worth posting loudly: the non-coding frontier has hard edges, and timeline-based creative work is still one of them. Knowing where the agent fails is as valuable as knowing where it wins.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2058911940613665181
A clean before/after on context efficiency. Scanning a large project with Claude Code took 52 tool calls; after dropping in CodeGraph, which indexes functions, variables and dependencies into a knowledge graph, it took 3. On a 4,000-file project that's 46% faster and 35% fewer tokens, fully local. With the cost panic raging this day, this is the practical answer, stop making the agent re-read your codebase and give it a map.
@svpino [Claude Code]
https://x.com/svpino/status/2058888825284513842
An underrated context trick: instead of /compact, hit the checkpoint menu, pick a checkpoint right after the early context you want to keep, and choose "Summarize from here." It preserves the valuable early specs and decisions while collapsing the later noise, rather than blindly compressing everything. Small, but if you live in long Claude Code sessions this is the difference between keeping your architecture decisions and losing them.
@mvanhorn [Claude Code]
https://x.com/mvanhorn/status/2058898044926775390
He built a set of Claude Code skills (ppress) that auto-generate a high-quality CLI plus skill for Hermes, OpenClaw or Claude Code in 30 to 60 minutes, and racked up 150+ tools in his library within two weeks. Favorites include a combined Google Flights plus Kayak CLI, a Tesla integration, and ESPN. This is the meta-move, using the agent to manufacture the agent's own tools, and the library compounds faster than anyone could build by hand.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2058887184040554833
The safety pattern everyone giving agents real power should copy. He uses n8n as a proxy layer around OpenClaw so the agent never directly holds API keys. OpenClaw does the reasoning, then calls an n8n webhook that pauses on a Wait node and sends him an approval form before anything sensitive happens, refunds, billing changes, DB writes, posting content. Human-in-the-loop done right, the agent thinks freely but a person gates the actions that can hurt you.
🗣 User Voice
User Voice
The most striking thing across all of these is the gap between what people are building and what they're complaining about. The complaints are getting sharper and more specific, which is itself a sign of maturity, these are power users, not tourists.
Cost is the wound everyone is poking. The entire day was colored by the Microsoft/Uber/Nvidia "AI costs more than the humans" narrative, and at the individual level @Meta8Mate is openly tracking $1,200/month while others report canceling after the bills spiked. The ask is consistent: predictable, controllable, outcome-based spend, not a meter that spins faster the better the agent gets.
Memory and context loss is the second wound. @anujcodes_21 nails the daily ritual nobody notices, re-explaining your stack to Claude, Cursor, ChatGPT, Codex, Gemini and OpenClaw because none of them share context. It's why half the workflows here are really memory hacks: CLAUDE.md, Obsidian vaults, knowledge graphs like @VincentLogic's CodeGraph. People are bolting persistence onto tools that forget.
Skill and config bloat is the newer complaint. Power users report that installing every recommended skill actually makes the agent dumber and quietly inflates token use, and they want one-job skills and a "skill-cleaner" to trim the prompt budget. The first wave was "give my agent more abilities." This wave is "stop drowning it."
And there's a quieter structural gripe: the CLI-versus-app and remote-access friction. People want to run Claude Code 24/7 on a Mac Mini and drive it from their phone (@yanhua1010), and they hit single-device and session-recovery limits doing it. The agent is becoming an always-on coworker, and the tooling still assumes a developer at a laptop.
The most striking thing across all of these is the gap between what people are building and what they're complaining about. The complaints are getting sharper and more specific, which is itself a sign of maturity, these are power users, not tourists.
Cost is the wound everyone is poking. The entire day was colored by the Microsoft/Uber/Nvidia "AI costs more than the humans" narrative, and at the individual level @Meta8Mate is openly tracking $1,200/month while others report canceling after the bills spiked. The ask is consistent: predictable, controllable, outcome-based spend, not a meter that spins faster the better the agent gets.
Memory and context loss is the second wound. @anujcodes_21 nails the daily ritual nobody notices, re-explaining your stack to Claude, Cursor, ChatGPT, Codex, Gemini and OpenClaw because none of them share context. It's why half the workflows here are really memory hacks: CLAUDE.md, Obsidian vaults, knowledge graphs like @VincentLogic's CodeGraph. People are bolting persistence onto tools that forget.
Skill and config bloat is the newer complaint. Power users report that installing every recommended skill actually makes the agent dumber and quietly inflates token use, and they want one-job skills and a "skill-cleaner" to trim the prompt budget. The first wave was "give my agent more abilities." This wave is "stop drowning it."
And there's a quieter structural gripe: the CLI-versus-app and remote-access friction. People want to run Claude Code 24/7 on a Mac Mini and drive it from their phone (@yanhua1010), and they hit single-device and session-recovery limits doing it. The agent is becoming an always-on coworker, and the tooling still assumes a developer at a laptop.
📡 Eco Products Radar
Eco Products Radar
Codex (OpenAI) is the permanent control group, mentioned in nearly every workflow as the speed-and-finish counterpart while Claude Code holds planning and design. Cited constantly.
Hermes Agent (Nous Research) is the open, self-hosted partner that keeps showing up, orchestrated by Claude Code as a sub-tool (@tetumemo), running social ops (@AYi_AInotes), and powering local-first setups. The "run it on your own machine" alternative everyone reaches for.
n8n is the connective tissue of the day, the proxy/guardrail layer (@SimonHoiberg), the automation backbone behind the self-running personal OS (@obsidianstudio9), the thing that turns a $20 subscription into client deliverables.
Obsidian has quietly become the memory layer of choice, plugged into Claude Code via the filesystem MCP to give agents persistent, queryable context (@obsidianstudio9, @ObsidianOtaku).
CodeGraph / Understand Anything are the context-efficiency tools of the moment, pre-indexing a codebase into a knowledge graph so the agent stops re-reading files, cutting tool calls and tokens dramatically (@VincentLogic, @DivyanshT91162).
Google Antigravity (and Gemini CLI / Grok Build) draw heavy comparison traffic as the rival Claude-Code-style harnesses, frequently in head-to-head workflow tests.
Higgsfield / Arcads are the video-generation engines bolted onto these agents for the content-factory plays (@shedntcare_), turning agent-written scripts into finished AI-actor video.
Nango is the open-source unified integrations layer that surfaced loudly this day as the self-hosted replacement for the $50k/year integration platforms, fed by Claude Code/Cursor/Codex prompts.
Codex (OpenAI) is the permanent control group, mentioned in nearly every workflow as the speed-and-finish counterpart while Claude Code holds planning and design. Cited constantly.
Hermes Agent (Nous Research) is the open, self-hosted partner that keeps showing up, orchestrated by Claude Code as a sub-tool (@tetumemo), running social ops (@AYi_AInotes), and powering local-first setups. The "run it on your own machine" alternative everyone reaches for.
n8n is the connective tissue of the day, the proxy/guardrail layer (@SimonHoiberg), the automation backbone behind the self-running personal OS (@obsidianstudio9), the thing that turns a $20 subscription into client deliverables.
Obsidian has quietly become the memory layer of choice, plugged into Claude Code via the filesystem MCP to give agents persistent, queryable context (@obsidianstudio9, @ObsidianOtaku).
CodeGraph / Understand Anything are the context-efficiency tools of the moment, pre-indexing a codebase into a knowledge graph so the agent stops re-reading files, cutting tool calls and tokens dramatically (@VincentLogic, @DivyanshT91162).
Google Antigravity (and Gemini CLI / Grok Build) draw heavy comparison traffic as the rival Claude-Code-style harnesses, frequently in head-to-head workflow tests.
Higgsfield / Arcads are the video-generation engines bolted onto these agents for the content-factory plays (@shedntcare_), turning agent-written scripts into finished AI-actor video.
Nango is the open-source unified integrations layer that surfaced loudly this day as the self-hosted replacement for the $50k/year integration platforms, fed by Claude Code/Cursor/Codex prompts.
Comments