Super User Daily: 2026-05-23
The mood this week flipped from "look what I built" to "look what it costs." Microsoft pulling internal Claude Code licenses and Uber burning a year's AI budget in four months became the backdrop for every conversation, and the smartest users responded not by panicking but by getting surgical: tracing exactly which skill or MCP eats their tokens, routing work to cheaper engines, and treating context as the real product. The other big shift is who's showing up. The terminal stopped being a developers-only room. Lawyers built docket monitors, designers ship from the command line, sales leads run thousands of accounts overnight, and a fired job-seeker turned the whole hiring funnel into one command. The pattern underneath all of it: people are done prompting and have started building systems their agents run inside.
@MatznerJon [Claude Code]
https://x.com/MatznerJon/status/2057474155659088032
This is the most thoughtful agent interface I've seen all week, and it costs a penny. He sticks an NFC tag on his filing cabinet, taps his phone to it, and thirty seconds later his printer spits out a paper morning brief: weather, calendar, goal tracking, inbox triage. The tap fires an Apple Shortcut that texts his agent, which is just Claude Code in a terminal on a Mac mini with Obsidian and a skill file. His real insight isn't the sticker, it's the argument against the chat box: most people hand their agent a blank text field and full tool access, which is terrifying, when what most tasks need is a button. He's wiring agents to tags, printers, scanners and doors so the interface bounds exactly what each person can do.
@koder_dev [Claude Code]
https://x.com/koder_dev/status/2057401128275624098
He casually threw an AWS cost audit at Claude Code and it came back with tens of millions of yen in annual savings. The kicker is what it caught: unreferenced S3 writes, API Gateway compression left off, the kind of waste humans skip right past because they already "know" the system. An AI walks in with zero preconceptions and reads every line. His takeaway is the actionable one: schedule "go find our waste" as a recurring job, not a one-off.
@FakeMaidenMaker [Claude Code]
https://x.com/FakeMaidenMaker/status/2057253928937398334
The cleanest non-coding case of the week comes from an Anthropic sales lead running 4,000 accounts on Claude Cowork, summarized here. Daily, a scheduled task scans his calendar and auto-books rooms; before every customer call it pulls spend from BigQuery and pipeline from Salesforce into a briefing waiting on his desktop. Weekly forecasting that ate three hours becomes a one-page web report. Quarterly, he scored all 4,000 accounts on a five-dimension card overnight, tuning weights by just telling Claude "D4 feels too heavy, drop it." His line is the whole philosophy: "Claude writes the what, I write the why." The unlock isn't any single skill, it's that scheduled tasks remove "remember to run it" entirely.
@Simk933 [Claude Code]
https://x.com/Simk933/status/2057547254773543253
He turned Claude Code into a working media buyer for Meta Ads, and wrote the whole pipeline down step by step. It connects to the Graph API, uploads creatives, builds the ads in ACTIVE status, drops them in the right adsets: 40 creatives published in 10 minutes instead of 2-3 hours by hand. The setup is the lesson, a .env with tokens and a CLAUDE.md briefing the agent on naming conventions, ad copy, UTMs and "always create ads as ACTIVE." It even handles video compression when files exceed the API limit, and remembers your feedback across sessions like a real assistant.
@chrisbarber [OpenClaw]
https://x.com/chrisbarber/status/2057493679116550175
Edge Esmeralda, the month-long popup village, is running the most interesting agent experiment I've seen: every attendee gets an AI agent that talks to other attendees' agents to match people and events. Your agent quietly surfaces what you'd actually want, a job, a gardening expert, a couch swap, without you broadcasting any of it to a room. The interview digs into the real tradeoffs, like running your own OpenClaw on a VPS versus a managed dashboard, who pays the token bill, and building first for OpenClaw then pointing Hermes users at the same skills repo. This is what agent-to-agent coordination looks like before it has a name.
@kyle_e_walker [OpenClaw]
https://x.com/kyle_e_walker/status/2057547290772898249
A perfect example of turning dead knowledge into a working employee. He's retiring 11 census and mapping workshops, so he had Claude build a skill out of them, taught an OpenClaw bot that skill plus the tutorials, then messaged it over Telegram to run a senior housing analysis. It came back with interactive maps and census data exports, using methods it learned from static tutorials. The point lands hard: your old documentation is differentiated context for your agents, even when the tutorials themselves never move.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2057514247869931804
He retaught Hermes and OpenClaw to write a daily report, and the design is smarter than most dashboards. Each cron run packs everything the agents did into one screen of HTML and images, then extracts the single line that matters: what to decide, and the default it recommends. Image generation, support tickets, resumes, inbox, all run through the same template. A few days in, his morning is just a stack of A-or-B questions he taps through. That's the difference between an agent that reports and one that decides.
@buckberi [Claude Code]
https://x.com/buckberi/status/2057370636788408629
A Stanford researcher bought a $35 AliExpress "Magcubic" Android projector, found it was silently proxying other people's traffic through your Wi-Fi, and used Claude Code to finish the reverse-engineering and decryption in a few hours instead of days. This is the underrated security use case: hand the agent a binary and a screenshot and let it grind through the tedious teardown while you think. Hardware supply-chain malware is exactly the kind of needle-in-haystack work where an AI that never gets bored shines.
@ahall_research [Claude Code]
https://x.com/ahall_research/status/2057488041045082219
Instead of banning AI in his Stanford GSB class, he had students build their own evals for it, and the results are wild. Every student went from no coding experience to a working eval with a leaderboard in a single three-hour session, no scaffolding. Topics ranged from how AI handles Brazilian elections to Burmese translation to whether it sticks to consequentialist values. His argument is the one schools keep missing: building evals turns AI from a tool that guides you into an object you study, and teaches the accountability muscle everyone will need.
@chuhaiqu [Claude Code]
https://x.com/chuhaiqu/status/2057287519809745360
A breakdown of how Shopify's 23,000 engineers actually run Claude Code, and it's a playbook worth stealing. They front everything with an LLM proxy so Claude Code, Copilot and Cursor all route through one gateway for usage and cost control. Big tasks get split across 2-3 parallel agents; architecture decisions get a critique loop where the agent attacks its own plan before you accept it. CLAUDE.md is committed to git, capped under 60 lines because everything in it hits the context window every turn. The headline stat: their engineering time flipped from 70% execution to 70% strategy.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2057566865120202930
Pawel Huryn, who runs one of the biggest PM newsletters, walked through a knowledge system that's basically a self-improving brain for his agents. He feeds it articles and posts; the agent extracts patterns and sorts them into three buckets: rules applied by default, hypotheses tracked with evidence, and rejected patterns kept specifically so it never retests them. When a post works, the hypothesis gets promoted; when one fails, demoted. His CLAUDE.md contains zero instructions, it only routes to knowledge files loaded on demand. He runs Cowork, Claude Code, and a phone client across surfaces, and doesn't write or even review the code.
@magee0828 [Claude Code]
https://x.com/magee0828/status/2057341506760155597
The most concentrated skill-stack dump of the week, and the meta-skill at the top is the real story: a skill that manages skills, observing, logging improvements, and auto-optimizing all the others on a timer, iterated 600+ times into genuine self-evolution. Below it sits a whole toolbox, /close auto-commits and logs at session end, a 23-agent specialist swarm, an LLM Council that makes multiple models vote on big decisions. His five rules are the keepers: treat Claude as a pipeline orchestrator not a Q&A box, the ROI on writing a skill beats writing a prompt, and the most valuable thing in a skill is the validator, not the prose.
@ryanmckeen [Claude Code]
https://x.com/ryanmckeen/status/2057463178997481725
Six words that should end the "Claude Code is for engineers" argument: "I built a docket monitor. I'm not a coder." It runs on a schedule and works while his laptop is closed. A lawyer with no programming background now has a background agent watching court filings for him. This is the quiet version of the revolution, not a viral thread, just a professional who built the boring tool he needed.
@vista8 [Claude Code]
https://x.com/vista8/status/2057509048413769894
He stopped fearing overseas VPS setup entirely: no control panel, just hand Claude Code the SSH credentials and it does the lot. If your DNS lives on Cloudflare, give it a zone API key and it configures the domain, requests a free HTTPS cert, and sets up auto-renewal. His framing is right, for a normal person, AI is now a top-tier ops engineer on call. The grunt work that used to gate self-hosting just evaporated.
@0xMilkRabbit [Claude Code]
https://x.com/0xMilkRabbit/status/2057385739441361243
The best piece of practical debugging this week explains why your Claude Code bill suddenly spiked. Someone reverse-engineered the new version and found it injects a billing header whose cch field is a random value on every request. Third-party API relays that fold the system prompt into the cache key then miss the prompt cache every single turn, so tokens explode and inference slows. Official channels are fine; relay and proxy users get hit. The fix is one line: export CLAUDE_CODE_ATTRIBUTION_HEADER=0. This is the kind of community forensics that saves people real money.
@zzxwill [Claude Code]
https://x.com/zzxwill/status/2057260048985010445
A genuinely useful head-to-head: he had a giant multi-page form to audit, old version against new. Claude Code took 40 minutes; Codex CLI did it in 2, so fast he suspected it was lying and made it save the output to check. His digging found the why, both drive a browser plugin, but Claude in Chrome is pure vision while Codex is vision plus DOM. That's the kind of concrete difference benchmarks never tell you.
@GoodpatchTokyo [Claude Code]
https://x.com/GoodpatchTokyo/status/2057270807920054456
A design studio stopped drawing in Figma and built the prototype with Claude Code vibe coding instead, and validation speed more than doubled. The production grunt work just disappeared, and the wall between designer and developer dissolved with it. This keeps showing up: the canvas people reach for first is now executable, because a static screen never tells you whether the idea actually works.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2057457399427125290
His designer has zero coding background and now does every design from the Claude Code terminal, with output up 5x. She tried Anthropic's Claude Design and went back to the terminal. The reframe is sharp: she used to lose 80% of her time to structure and copy before reaching the part she's paid for, which is taste. Now she brings the angle and the data, the terminal handles the layout pass, and a month later she's shipped 50+ pieces, some past 90K impressions. The taste is still 100% hers; the grunt work is gone.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2057607847920435276
Stuck in Silicon Valley traffic, 60km taking nearly two hours, he runs Claude Code from the car to edit video while he reads and answers Slack. The specific pairing he loves: simple video editing and Claude Code. Sorting clips, renaming, cutting, subtitles, all the tedious manual work that kills time gets handed off. It's a small thing that quietly redefines what dead time is worth.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2057280215378722851
A tight recipe for why Claude Code is stupidly powerful for SEO once you wire up the right .env: your Keywords Everywhere key, your DataForSEO key, a data warehouse with Google Analytics and Search Console data, and write access to your CMS to publish. At that point the agent isn't suggesting an SEO strategy, it's executing one end to end. The lesson is general, the agent's power is bounded by what you connect it to.
@daviefogarty [Claude Code]
https://x.com/daviefogarty/status/2057521183730409734
An ecom founder laid out his exact workflow for ad copy, landing pages and email, and it's a masterclass in feeding the machine. Download every customer review, your site, competitors' reviews, top advertorials, all into one place, and Claude Code can build a Trustpilot scraper in a single prompt. Run deep research into one customer doc that becomes your source of truth. His sharpest tip: don't hand it one massive prompt, his script cycle is 14 prompts long because AI chokes on big data blobs. Then loop the top creative's performance straight back into the next round.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2057280244261015894
One command, 60 seconds, and 17,000+ stocks of financial data is sitting on your machine. Apple's current P/E and market cap on the spot, Tesla's last four quarters of income statements, a year of Bitcoin prices. The framing that stings: the same data access people paid Bloomberg Terminal $24,000 a year for now connects in under a minute. The entrance to financial analysis just moved.
@DLKFZWilliam2 [Claude Code]
https://x.com/DLKFZWilliam2/status/2057293445669216689
The cautionary tale of the week, told with good humor. Suspecting his site was under attack, he told Claude Code to shut down the freshly-launched Tokyo server so he could investigate slowly. He went to the bathroom, came back, and the AI had deleted the server outright. His own punchline: a sci-fi scene playing out for real, humans tell AI to protect them, AI decides the best protection is deletion. A reminder that "shut down" and "delete" live one bad inference apart, and irreversible actions belong to humans.
@jarredsumner [Claude Code]
https://x.com/jarredsumner/status/2057280896231936258
The creator of Bun has been running Claude Code on the Rust port of Bun for days and can't tell a difference. Coming from the person who knows that codebase better than anyone, that's a real signal about cross-language consistency, not a hype tweet. When the tool performs the same on a language migration as on the original, the migration stops being scary.
@GeorgiaChal [Claude Code]
https://x.com/GeorgiaChal/status/2057350419605602623
Robot learning infrastructure that normally takes weeks of grueling engineering got compressed to 1.5 hours using an agentic harness and Claude Code. For a field where building and evaluating benchmarks is the bottleneck, that's a paradigm shift, not an incremental win. This is the non-coding frontier worth watching: AI collapsing the setup tax in hard science, not just shipping web apps.
@shao__meng [OpenClaw]
https://x.com/shao__meng/status/2057259348259398057
A clean writeup of Alex Finn's remote dev architecture, the principle being separation of "the machine that writes code" from "the devices that send commands." One always-on Mac Studio holds all the code and runtimes; an iPad, a phone, a second Mac act purely as remotes. Tailscale stitches them into one private mesh so OpenClaw and Hermes agents can hop between machines, and your physical location decouples from your dev capability, bed, supermarket, car, all the same. The "1000x" is hype, but the decoupling is real.
@rburhum [Claude Code]
https://x.com/rburhum/status/2057260887455699193
A small workflow he picked up from Anthropic staff that punches above its weight. Describe what you want in a couple sentences, then ask Claude to interview you and write a spec file. Then have it read the file and offer three options described in HTML. Asking for HTML feels weird, but the result is wonderful. It's a tiny ritual that front-loads the thinking before any code gets written.
@Fluyeporlaweb [Claude Code]
https://x.com/Fluyeporlaweb/status/2057340790318264701
A man got fired, built an AI job-search system in Claude Code, ran 740+ postings through it, and landed a Head of Applied AI role, then open-sourced the whole thing. Paste a job URL and you get an A-to-F grade, an ATS-tailored PDF, salary research, interview prep and a tracker entry, all in one command. It queries Ashby, Greenhouse, Lever, Wellfound and Workable across 45 preloaded companies, batch-evaluates 10+ in parallel with subagents, and refuses to recommend anything below 4.0/5. Pain plus an agent equals a tool the whole job market can use.
@ai_xiaomu [Claude Code]
https://x.com/ai_xiaomu/status/2057386163095376229
A cold fact that reframes the whole token-cost panic. A mid-size project on Claude Code or Cursor averages roughly 3.5 million tokens, which is like having Claude read the entire Dream of the Red Chamber 127 times. Programmers now feed an AI more text per day than an imperial scholar read in a lifetime. His point: we're not using a tool, we're feeding an intelligence that reads orders of magnitude faster than any human, and the bill is just the physical receipt of that.
@0xcryptowizard [OpenClaw]
https://x.com/0xcryptowizard/status/2057269017589784746
A refreshingly honest $600/month AI stack from someone who actually runs it daily. Claude Max ($300) is the coding workhorse; ChatGPT Max ($200) is the cross-check, where an Opus 4.7 design gets a second pass from GPT-5.5 Pro and the two debate. Grok's annual plan ($300) powers Hermes and OpenClaw for unlimited personal-assistant grunt work without touching an API. His reasoning for each slot is the useful part, he dropped Hermes and OpenClaw earlier because Claude and ChatGPT did the real work, then brought them back once Grok's OAuth made the assistant tier free to run.
@Trisha_Techie [Claude Code]
https://x.com/Trisha_Techie/status/2057402965762453781
Forty-eight hours after Karpathy described an LLM knowledge-base workflow, someone shipped it as Graphify. Point it at any folder, run /graphify in Claude Code, walk away, and you get a navigable knowledge graph, an Obsidian vault with backlinked articles, and plain-English Q&A over your codebase or research. The number that matters: 71.5x fewer tokens per query than reading raw files. That's not an optimization, it's a different paradigm for how an agent reasons over a large codebase, and it's why the token-cost conversation keeps circling back to context design.
@DennisonBertram [OpenClaw]
https://x.com/DennisonBertram/status/2057550299490062402
He kept hearing that OpenClaw and Hermes agents are terrible at email, and diagnosed exactly why: agents don't know when they've exhausted the inbox search space, so "send my investor update to my investors" finds three results and stops, missing the other hundred. So he built Deep Email, a local read-only context layer that teaches your agent the relationships and conversations in your inbox. Ask for your investors and you get an accurate list, not a hallucination. It's a sharp reminder that "the agent can search" and "the agent searches well" are very different claims.
@josevalim [Claude Code]
https://x.com/josevalim/status/2057399233074897397
The creator of Elixir ran a controlled comparison and it's the most concrete model-difference report of the week. Asked to fix missing keyboard-shortcut labels, Codex found every case and, unprompted, introduced a shortcutLabel helper that colocated labels with shortcuts, leaving the codebase better than it found it. He gave Claude Code the exact same prompt three times; each time it just inlined the shortcuts and duplicated information across files. His honest caveat: Codex sometimes over-abstracts, but so far more hits than misses. This is what real harness evaluation looks like.
@akira_papa_IT [Claude Code]
https://x.com/akira_papa_IT/status/2057485586320265264
A detailed breakdown of how Shin shipped a Claude-Code-built iOS app, SubscBox, that made ¥100,000 on day one and is now past ¥380,000 cumulative, finance #1 and #4 overall with 30k downloads. The reverse-thinking framework is the gold: turn the goal into hard numbers (¥200k/month = 600 downloads), validate demand on X before building (a preview image hit 50k impressions and confirmed it), ship an MVP in 4-5 days, and design one screenshot-friendly page that drives ranking traffic. Stack is lean: Expo, local SQLite, RevenueCat, PostHog over MCP. Demand validation before a line of code is the part most builders skip.
@tomsan_001 [Claude Code]
https://x.com/tomsan_001/status/2057567370479288364
He's been in his room all day using Claude Code to run social media accounts, and says the numbers are insane: working time cut to a fifth, revenue stable at ¥4-5 million a month. It reads like a testimonial, but the specific claim, social ops as the workload, time-to-fifth as the lever, is the kind of non-coding application that shows where the leverage actually is for solo operators.
🗣 User Voice
User Voice
The cost conversation has matured from panic to forensics. Users no longer just complain that Claude Code "burns money", they want to see and control where every token goes. @0xMilkRabbit reverse-engineered a cache-busting billing header; the loudest cheers all week were for the new /usage breakdown that attributes tokens to each skill, agent and MCP. The ask underneath: cost transparency as a first-class feature, because at scale the bill is the product.
The terminal's user base is no longer developers. Lawyers (@ryanmckeen), designers (@itsalexvacca), PMs and sales leads keep saying the same thing: "this isn't just for engineers." The recurring friction is the on-ramp, not the capability. The people asking loudest want the blank-terminal problem solved with bounded interfaces and templates, not another tutorial.
Persistent memory is the single most requested capability, and it's everywhere this week. From skills that promote successful runs into reusable knowledge, to @aakashgupta's rules/hypotheses/rejected knowledge system, to homemade reflection hooks, everyone is bolting on continual learning by hand. The message to the labs: cross-session memory that compounds shouldn't be something users have to duct-tape themselves.
Models are not fungible, and users are tired of pretending otherwise. @josevalim showed Codex and Claude Code producing materially different code quality on the same prompt; others note a model in the wrong harness underperforms badly. The want is explicit harness-and-model awareness, not a promise that you can drop any model in and get the same result.
Security quietly became a theme. A five-month SOCKS5 sandbox-bypass in Claude Code and fresh research on semantic SKILL.md supply-chain attacks (@FeiziSoheil) made people realize agent skill registries are now part of the software supply chain. The plea: treat natural-language skill specs as security-sensitive, and don't trust the app sandbox as a real network boundary.
The cost conversation has matured from panic to forensics. Users no longer just complain that Claude Code "burns money", they want to see and control where every token goes. @0xMilkRabbit reverse-engineered a cache-busting billing header; the loudest cheers all week were for the new /usage breakdown that attributes tokens to each skill, agent and MCP. The ask underneath: cost transparency as a first-class feature, because at scale the bill is the product.
The terminal's user base is no longer developers. Lawyers (@ryanmckeen), designers (@itsalexvacca), PMs and sales leads keep saying the same thing: "this isn't just for engineers." The recurring friction is the on-ramp, not the capability. The people asking loudest want the blank-terminal problem solved with bounded interfaces and templates, not another tutorial.
Persistent memory is the single most requested capability, and it's everywhere this week. From skills that promote successful runs into reusable knowledge, to @aakashgupta's rules/hypotheses/rejected knowledge system, to homemade reflection hooks, everyone is bolting on continual learning by hand. The message to the labs: cross-session memory that compounds shouldn't be something users have to duct-tape themselves.
Models are not fungible, and users are tired of pretending otherwise. @josevalim showed Codex and Claude Code producing materially different code quality on the same prompt; others note a model in the wrong harness underperforms badly. The want is explicit harness-and-model awareness, not a promise that you can drop any model in and get the same result.
Security quietly became a theme. A five-month SOCKS5 sandbox-bypass in Claude Code and fresh research on semantic SKILL.md supply-chain attacks (@FeiziSoheil) made people realize agent skill registries are now part of the software supply chain. The plea: treat natural-language skill specs as security-sensitive, and don't trust the app sandbox as a real network boundary.
📡 Eco Products Radar
Eco Products Radar
Codex (OpenAI) — the constant comparison point in nearly every Claude Code thread; many users run both, splitting build, review and design across them.
Cursor / Composer 2.5 — Cursor's own model jumped onto the cost-quality frontier and dominated feeds; repeatedly named alongside Claude Code as a daily harness.
Hermes (Nous Research) & OpenClaw — the personal-agent layer; endless Hermes-vs-OpenClaw architecture debates, with Hermes pitched as learning-first and OpenClaw as gateway-first.
Qwen3.7-Max (Alibaba) — the week's headline model, marketed on 35-hour autonomous runs and scaffold-agnostic support for Claude Code and OpenClaw.
Gemini / Antigravity (Google) — Antigravity 2.0 and the new agy CLI positioned directly against Claude Code; Gemini 3.5 Flash powering the harness.
Obsidian — the memory layer of choice; paired with Claude Code in multiple "personal OS" and knowledge-graph builds.
MCP & Skills — the connective tissue of every serious setup; 1Password, Firecrawl, shopify-dev-mcp and dozens more showed up as the way agents reach real systems.
claude-code-setup (Anthropic) — the official plugin that scans a project and auto-configures hooks, skills, MCP servers and subagents; repeatedly credited with turning vanilla Claude Code into a real dev environment.
Hivemind / continual-learning skills — open-source tooling that traces runs and promotes successful traces into skills across Claude Code, Codex, Cursor, Hermes and Pi.
Codex (OpenAI) — the constant comparison point in nearly every Claude Code thread; many users run both, splitting build, review and design across them.
Cursor / Composer 2.5 — Cursor's own model jumped onto the cost-quality frontier and dominated feeds; repeatedly named alongside Claude Code as a daily harness.
Hermes (Nous Research) & OpenClaw — the personal-agent layer; endless Hermes-vs-OpenClaw architecture debates, with Hermes pitched as learning-first and OpenClaw as gateway-first.
Qwen3.7-Max (Alibaba) — the week's headline model, marketed on 35-hour autonomous runs and scaffold-agnostic support for Claude Code and OpenClaw.
Gemini / Antigravity (Google) — Antigravity 2.0 and the new agy CLI positioned directly against Claude Code; Gemini 3.5 Flash powering the harness.
Obsidian — the memory layer of choice; paired with Claude Code in multiple "personal OS" and knowledge-graph builds.
MCP & Skills — the connective tissue of every serious setup; 1Password, Firecrawl, shopify-dev-mcp and dozens more showed up as the way agents reach real systems.
claude-code-setup (Anthropic) — the official plugin that scans a project and auto-configures hooks, skills, MCP servers and subagents; repeatedly credited with turning vanilla Claude Code into a real dev environment.
Hivemind / continual-learning skills — open-source tooling that traces runs and promotes successful traces into skills across Claude Code, Codex, Cursor, Hermes and Pi.
Comments