Super User Daily: 2026-05-22
Two days ago the Claude Code and OpenClaw timelines were less about "look what shipped" and more about "look what I actually got out of it" - and the most interesting cases had almost nothing to do with writing code. People are running Claude Code and OpenClaw agents as marketing departments, prediction-market trading desks, vulnerability hunters, and one-person companies. The recurring shape is the same: a model that reads your real files and burns real tokens around the clock, with the human reduced to writing the rules. Anthropic's London keynote and the Boris Cherny workshop set the tone for the day - the punchline everyone repeated was that 90% of Claude Code is now written by Claude Code, and the gap between people getting 10x out of it and people complaining about rate limits comes down entirely to whether they built the infrastructure around it. Below are the cases worth studying, sorted by how much they actually delivered.
@ArianaIsMyDrunk [Claude Code]
https://x.com/ArianaIsMyDrunk/status/2057025067025088923
This is the standout non-coding case of the day: a cousin who lost money mining ETH in a garage converted the same setup into a Polymarket "farm" running a 47-line Python bot that Claude Code wrote in one afternoon. Six phones, three laptops, one monitor, all running the same bot that opens a BTC market every five minutes, scans, enters, and collects. The reason for the device sprawl is rate limiting - one instance per device means six devices equals six times the position. The numbers cited are 31,221 predictions and $383K profit on one account, with last month's profit at $47K against $89 in electricity. Whether or not every figure holds up, the framing is the real lesson: the moat isn't the code, it's running many cheap parallel instances 24/7 on a problem with a measurable payoff.
@mhdfaran [Claude Code]
https://x.com/mhdfaran/status/2057126048094212521
The most credible enterprise-scale case: a breakdown of how Uber runs 60,000 AI agent tasks per week, with 95% of its 5,000+ engineers using AI to write code and 1,500+ internal agents active monthly. The interesting detail for Claude Code users is "Minions," an internal background agent running on Claude that ships 1,800 code changes every week on its own. The hard-won lesson is organizational, not technical: every team was building its own MCP connections, nothing got reused, and agents kept picking broken or wrong tools. Their fix was a central MCP gateway and registry that auto-reads 10,000+ service definitions, has an LLM write the tool descriptions, and scopes which tools each agent can touch to cut hallucination. Anyone running agents past pilot scale eventually hits this exact wall.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2057190389367685403
A genuinely instructive workflow from Andre Albuquerque, who went from zero coding ability to outshipping engineering teams using five agents inside Claude Code, each mirroring a real squad role - researcher, discovery, designer, engineer, implementer - with a PM orchestrator on top that routes tasks and never builds anything itself. The CLAUDE.md is the glue; he calls it "the values of your Claude," and every constraint loads before any task runs. The clever part is his debugging philosophy: when a feature ships broken, he doesn't fix the feature, he fixes the agent that built it and updates the CLAUDE.md, then reruns the whole pipeline from scratch so the same failure never recurs. He spends 50% of his build time improving agent infrastructure rather than shipping features, and claims three people now match the output of a 15-person team.
@shedntcare_ [OpenClaw]
https://x.com/shedntcare_/status/2057080235955548667
A concrete, fully-autonomous non-coding pipeline: an OpenClaw bot that finds local businesses on Google Maps, identifies the ones with weak or missing websites, builds each a personalized site in minutes, then prints a physical postcard with a live preview and QR code and mails it to their business address. No cold calls, no cold email - the prospect scans the card, sees their finished website, and reaches out to buy. The insight worth stealing is using direct mail as the delivery channel precisely because it dodges inbox spam filters, and leading with the finished product instead of a pitch. It's a clean example of an agent owning an entire lead-gen-to-close loop while the operator sleeps.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2057125455111565368
A detailed, channel-by-channel account of how ChatSEO generated €70K in five months with zero paid ads. The Claude Code-specific parts: they duplicate SEO pages across low-competition keyword patterns with Claude Code, and ship two to four free SEO mini-tools per day - briefed via ChatSEO, then built by Claude Code running four agents in parallel. They also use Claude to score every YouTube title and thumbnail out of 10 and suggest improvements until a 6/10 becomes an 8/10. The takeaway they articulate is "veille + execution" - curiosity to test new channels plus discipline to keep the habit - and the free-tools-as-SEO-magnet tactic is the most reusable piece.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2057084130404512218
The full GTM stack behind ColdIQ at $7M ARR, pulling 40,000+ visits a month. Claude Code shows up in two specific roles: shipping free mini-tools as an SEO channel, and as part of the lead-scoring layer alongside Clay and PredictLeads. It's less a single dramatic result than a map of where an AI coding agent actually slots into a real revenue machine - building the small capture tools and feeding the enrichment pipeline rather than being the headline. Useful as a reference for operators trying to figure out where Claude Code earns its keep in a marketing org.
@m_kumagai [Claude Code]
https://x.com/m_kumagai/status/2056910273416679591
A clean solo-builder story: someone who first touched Claude Code in a terminal in late January and left it alone, then in April started building a personal "MY cockpit" control panel for work automation. In roughly one month he went from web service to multi-PC sync to a mobile app to registering as a Chrome Web Store developer and publishing. He explicitly can't write a line of code himself. It's a modest but honest end-to-end case of a non-engineer shipping a real, published tool, and a reminder that the learning curve from "touched it once" to "published an app" is now about a month.
@orevbajohn_ [Claude Code]
https://x.com/orevbajohn_/status/2057010834941809049
A small but concrete output case: a working "confirm transaction" interaction built entirely with Claude Code, replicating a polished family-transaction speed-pill UI animation from another builder. No grand claims, just a shipped, visible artifact. It's the kind of low-noise evidence that often gets buried under launch hype - someone reproduced a non-trivial interaction and showed the result.
@earthtojake [Claude Code]
https://x.com/earthtojake/status/2056915506599252209
A nice example of Claude Code reaching out of pure software: someone used it to one-shot a design for a custom PC case via text-to-CAD, then 3D printed the result. It's a reminder that "coding agent" is already a misnomer - the same tool that writes Python is generating CAD geometry that turns into a physical object. Small case, but it points at where the agent's reach is expanding.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2056910306476196067
A builder shipped a reusable skill, /onboarding-cro, that takes a product description and maps the shortest path to the user's aha moment - identifying activation friction, building the welcome flow, the activation checklist, empty states, and behavioral email sequences. It's part of a free open-source collection of 40 marketing skills installable into Claude Code, Cursor, or Codex via npx. The value here is the productization of a marketing playbook into a portable agent skill rather than a one-off prompt, which is exactly the direction the ecosystem keeps converging on.
@tom_doerr [Claude Code]
https://x.com/tom_doerr/status/2057050421663822025
A practical security workflow: 183 pentesting tools integrated into Claude Code, turning the coding agent into an offensive-security driver. The same author also built a physical ESP32 dashboard that monitors real-time Claude Code usage, which is its own small tell about how heavily people are running these sessions - you don't build a hardware usage meter for a tool you touch occasionally. Both are concrete, shippable artifacts rather than opinion.
@heyshrutimishra [OpenClaw]
https://x.com/heyshrutimishra/status/2056934360570585364
A pointed real-world deployment: 360 ran an autonomous vulnerability-mining agent inside the OpenClaw ecosystem and walked out with 23 confirmed vulnerabilities, including two criticals, all filed into China's CNNVD and CNVD national vulnerability databases. The findings span all four layers an agent target exposes - authentication, network, execution, and control. The argument worth noting is that agents are a genuinely new attack surface where prompts double as instructions and the static-analysis baseline barely exists, and this is production infrastructure running, not a demo.
@omarsar0 [Claude Code]
https://x.com/omarsar0/status/2057119309151817877
A self-improvement experiment with real plumbing: the author had Claude Code interact with Fireworks Agent to fine-tune a small Qwen model so it produces the right output style to grow his PaperWiki, all driven through natural language, with the dataset and skill file shared. He's candid that the next step - tuning a model to genuinely "know" the data - is the hard part, but the working piece is a coding agent orchestrating an actual fine-tuning loop rather than just talking about self-improvement. He separately flagged a NanoGPT-Bench eval showing Codex, Claude Code, and Autoresearch recover only 9.3% of human progress on real AI R&D, a useful cold-water counterweight to the self-improving-agent hype.
@QuinnyPig [Claude Code]
https://x.com/QuinnyPig/status/2057230077319123036
A one-liner that lands because it's a real workflow, not a slogan: "The crappy AWS console experience walked, so the tmux session with a dozen Claude Code instances could run." It captures, in one breath, how heavy users actually operate - a wall of parallel Claude Code instances in tmux replacing the cloud vendor's own UI. Worth including precisely because it describes the default working environment of the power-user cohort.
@MarcusCrassus17 [Claude Code]
https://x.com/MarcusCrassus17/status/2057234034087064040
A genuinely useful data point on the economics: a heavy user reports their actual Claude Code consumption runs over $1,500/month while they pay only $200, concluding Anthropic is heavily subsidizing usage and can't sustain the demand without far more data-center capacity. It's anecdotal, but it lines up with the broader chatter that Anthropic's incentive is for you to burn the maximum tokens, and it quantifies just how lopsided the subsidy is for the heaviest cohort.
🗣 User Voice
User Voice
Rate limits and cost are the loudest pain point, by a wide margin. @sudoingX captured it: lighter workload, somehow hitting the weekly limit faster, slower and buggier than before, and threatening to cancel. The same complaint drove a wave of users toward Codex and Cursor, and pushed others to reroute Claude Code to DeepSeek, NVIDIA NIM, or Kimi/GLM endpoints to cut cost 80-90%.
Memory and context persistence across sessions is the second recurring ask. @anna_y_zhang, building Nessie, made the sharp point that getting all your AI conversations into one place is the easy part - the hard part is trustworthy retrieval, since you can have 500,000 words of legible data and still get garbage answers if the retrieval layer isn't precise. Several tools this week (watchmen, Contextberg, GBrain) exist specifically to give agents carry-over memory.
People are realizing the model isn't the bottleneck - the surrounding infrastructure is. @LearnWithBrij laid out the five-layer mental model (CLAUDE.md, Skills, Hooks, Subagents, Plugins) and argued most engineers are still on Layer 1, prompting instead of engineering cognition. The repeated theme is that "vanilla" Claude Code feels messy and the real power lives in what you bolt around it.
There's growing demand to use these agents well beyond coding, paired with frustration that it's not obvious how. @ClaudeCode_UT noted that the wall executives hit right after installing Claude Code is always the same - "can't figure out how to use it for anything but coding." The non-coding cases (marketing, trading, ops, CAD) are exactly what people are hungry for, and the ones that show concrete output travel fastest.
Heavy users want reliable long-running and background execution. @AlexFinn's multi-device remote setup and the recurring MacBook closed-lid complaints (people using dongles and apps just to keep the machine awake while agents run) both point at the same gap: people want to fire off agents and walk away, and the tooling for genuine 24/7 unattended runs is still rough.
📡 Eco Products Radar
Eco Products Radar
Codex - mentioned constantly as the head-to-head alternative to Claude Code, with a clear sentiment shift toward it this week on speed, price, and native computer use; many users now run it side by side with Claude Code.
Cursor - the polished IDE layer people keep as a complement to CLI agents, and a frequent landing spot for users fed up with Claude Code rate limits.
Hermes - the other major open-source personal agent, repeatedly compared with OpenClaw and, per multiple show-floor reports, currently winning the stability argument.
Gemini / Antigravity - Google's coding-agent push, widely read as Antigravity targeting Codex/Claude Code and Gemini Spark targeting Cowork/OpenClaw; lots of "where does Google actually lead" skepticism.
DeepSeek - the go-to cost-cutting backend, repeatedly used to reroute Claude Code traffic for roughly one-seventh the price via Anthropic-compatible endpoints.
Grok - now usable directly inside OpenClaw with an X Premium subscription and no API key, which drove a large share of the day's OpenClaw chatter.
MCP - the connective tissue everyone builds on; the central gateway/registry pattern (Uber, AgentKey) is becoming the standard answer to tool sprawl.
Higgsfield - the media-generation hub people drive from Claude Code via MCP to turn one prompt into full storyboard-to-video pipelines.
Comments