Super User Daily: May 25, 2026
Two themes ran through everything people shipped on May 23. First, the agent stopped being a coding tool and became a coworker you hand whole jobs to, run a portfolio of, and check on from your phone, an electrician reading breaker panels, a scientist writing crystallography code, a Korean trader running an investment committee of agents while he sleeps. Second, the bill arrived. Microsoft pulling internal Claude Code licenses, Uber torching its yearly AI budget in four months, and OpenClaw's own creator dropping $1.3M on tokens in a month all landed the same day, and the loudest practical skill of the day turned out to be not prompting but cost control: stripping plugins, routing to cheaper models, and auditing where your tokens actually go. Here are the people doing the most interesting work with it.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2058216140523163886
A CPO built a product-manager agent from a single Claude Code prompt: it pulls every GitHub issue, scores each one by priority, and writes a daily build-next report, fully instrumented with one command and no IDE opened. The interesting part is the self-improvement loop she bolted on, the agent grades its own scoring accuracy, notices it was ranking bugs too low, and feeds the correction back into itself on a cron while she sleeps. Her analogy was Nadal studying his own film to get 1% better daily. The PM's job collapsed from processing information to curating taste, defining what "good" means and letting the agent consume every issue, Gong call, and Slack thread. Same-day shipping is the result: issue in, prototype out, deployed that afternoon.
@Morett_the_best [Claude Code]
https://x.com/Morett_the_best/status/2058170249229324619
This is the most complete production workflow anyone posted: at Monest, over 80% of the codebase is now Claude-generated, and it runs through a real pipeline, not vibes. Every feature starts as an RFC approved by two tech leads, spread into frontend/backend repos via git submodules, and used as the basis for Linear tickets, so CLAUDE.md always pulls context from both the RFC and the issue. A skill sits in an infinite feedback loop reading CodeRabbit's review comments and auto-applying the valid fixes. The guardrails are the real lesson: PRs capped at 500 lines (because internal data showed bigger PRs got 4x fewer comments and nobody actually reads them), mandatory rabbit-plus-human approval, 16 e2e test shards. His point is brutal and correct, the bottleneck isn't generating code, it's keeping humans able to actually understand what shipped.
@v81093933 [OpenClaw]
https://x.com/v81093933/status/2058129233969381559
A Korean investor wired up a stock-advice system as an actual org chart of agents: the Korean market goes to an OpenClaw agent, the US market to a Hermes agent, and Codex acts as the final coordinator. Each runs as a team with sub-agents handling market conditions, info gathering, and stock analysis, then they hold an "investment committee" meeting and synthesize advice on his portfolio. When the first version gave generic output, he layered in geopolitics, bond-rate moves, China's US-treasury selling, oil, short-seller reports, and retail sentiment, explicitly tuning it to hunt undervalued names rather than chase trends. He keeps the final trade decision human, but the research committee runs itself.
@RealYDT [Claude Code]
https://x.com/RealYDT/status/2058026908336701825
He asked Claude Code to optimize his network the way he'd seen someone do with Codex, and it just started running diagnostics, ping tests, DNS resolution timing, MTU detection. It found 10% packet loss and a foreign domain taking a slow detour, then caught the real bug: his IPv4 MTU was set to 1500, but PPPoE dial-up eats 8 bytes, so anything over 1492 was silently dropped, while IPv6 was correctly at 1492 and IPv4 had been wrong for who knows how long. One netsh command later, packet loss hit zero. His takeaway is the sharp one, the AI didn't know his network in advance, it ran commands, read output, and reasoned to the cause exactly like a human would, just without getting bored.
@walls_jason1 [Claude Code]
https://x.com/walls_jason1/status/2058240168855752812
A master electrician is running production inference on Vertex with Gemini Vision reading homeowner electrical panels, inside a tool he built himself with Claude Code. His own words capture why this matters: the same infrastructure the frontier labs ship on, in the hands of a guy who pulls wire for a living. This is the real story of the day in one tweet, the capability has diffused far past software people.
@gagarot200 [Claude Code]
https://x.com/gagarot200/status/2058107764942287062
Someone ran Hermes Agent plus Claude Code on a stock 1998 iMac G3 with 32MB of RAM and Mac OS 8, no modern upgrades. They ported Karpathy's llama2.c to classic Mac OS, compiled it with Retro68, used a Python script to convert the model data to big-endian for the PowerPC chip, and got a 260K-parameter model generating text locally on a 233MHz machine most people would call e-waste. It's a stunt, but it's the kind of stunt that shows how thin the floor for "can run an agent" has gotten.
@dadbodshuffl [OpenClaw]
https://x.com/dadbodshuffl/status/2058122167506886686
He built an OpenClaw-style agent for local Indian languages using Sarvam: a voice note in Gujarati, Hindi, or English goes through Saarika STT and Saaras translate, into a local runtime, which drives a logged-in Blinkit browser session via Playwright and updates the cart only after explicit approval. His design philosophy is worth stealing, don't make the LLM do everything, let the runtime handle deterministic state, approvals, browser actions, and language consistency, with the agent as fallback only when needed. Replies stay in the user's language via Sarvam translate. This is what agent UX for non-English, non-technical users actually looks like.
@regent0x_ [Claude Code]
https://x.com/regent0x_/status/2058079320405332047
The cleanest cost lesson of the day, framed as a generational contrast: a dad built a $2M company on a 1986 Macintosh Plus with one printer and zero configuration, while his son's Claude Code setup burned 62,000 tokens, 31% of the context window, before typing a single prompt, eaten by 23 plugins, 8 skills, and 5 MCP servers sitting idle. Sessions died at 30 minutes and outputs degraded as context filled. He ran /context, saw the damage, deleted everything, and dropped to 6,000 tokens of overhead using only built-ins, /compact, /clear, /resume. Sessions went from 30 minutes to 3+ hours on the same work. His punchline: most plugins are just markdown files you could write yourself in .claude/commands with zero overhead.
@eeuoss [Claude Code]
https://x.com/eeuoss/status/2058249221690863884
He's seeing how far Claude Code can go building a graphics renderer from scratch, pure Metal4, a basic material system, no libraries beyond ABSL utilities. The detail that makes it work: he gave the agent a way to render its output to an image so it could look at the result and judge it, and it would go fetch textures on its own when needed. This is the self-verifying loop applied to a genuinely hard, low-level domain where you can't fake correctness.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2058003129585860683
A dead-simple, high-leverage trick: add "E2E test it" to the end of any Claude Code request and it runs its own verification loop, finishes the work, cuts screenshots or frames, looks at them with its own eyes, and fixes what's wrong before reporting done. No special MCP needed since Claude Code can already execute commands and view images it produces. He says it's most powerful for video editing, design, and landing-page checks, the visual work where "done" usually hides a subtitle clipping off the screen. The deeper point: hand the AI a way to check its own answer instead of trusting "it's finished."
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2058094516255772977
Same builder, separate proof: he made an entire video commentary on Google I/O using only Claude Code, the title card, two-column subtitles, subtitle animation, and BGM are near-perfect, with transcription and music handled via Gemini 3.5 Flash and Lyria 3 Pro APIs. He's turned the whole thing into a skill, which means his video editing is now fully automated. His note that auto-captioning in CapCut produces so many typos it's actually slower is the kind of real-world detail that separates someone shipping from someone speculating.
@Money_plus_ai [Claude Code]
https://x.com/Money_plus_ai/status/2058178425840439662
Someone built an AI job-search system with Claude Code that sent over 700 tailored applications and actually landed him a job, then open-sourced it. It automates nearly the whole funnel: reviews job pages across companies, customizes your resume per role, and auto-fills application forms. The repo ships with 14 modes (evaluation, scraping, PDF handling), a Go-based terminal dashboard, ATS-optimized PDF resume generation via Playwright, and 45+ pre-configured companies including Anthropic, OpenAI, ElevenLabs, and Stripe. The outcome, not the demo, is the point: it got a human hired.
@harjtaggar [OpenClaw]
https://x.com/harjtaggar/status/2058273638332404043
A snake showed up in his backyard, so instead of panicking he sent the photo to his OpenClaw, which identified it as a harmless gopher snake, gave him a local number to call to confirm, and suggested dealing with the gophers that were attracting it. His framing is the quietly important part, he has small moments like this every day now that make daily life feel a little more magical. This is the non-coding, ambient-assistant use case that OpenClaw keeps proving out, and it's more compelling than most of the dev demos.
@MelihKarakelle [Claude Code]
https://x.com/MelihKarakelle/status/2058308161417711679
He asked Claude Code to write a user-manual PDF for a hardware device by reading its code. When it produced ASCII art instead of screenshots, he told it to generate the images from the Arduino code, and it sat down and wrote Python to fully emulate the Arduino code, render the LCD screen, and output actual PNGs. His reaction lands: a human would have refused and drawn it pixel by pixel in Illustrator, which is exactly what he used to do. The agent solved a visual problem by building a simulator, which is not how most people think to use it.
@eternityspring [Claude Code]
https://x.com/eternityspring/status/2058106975599403187
He stopped watching the screen while coding and started watching a light. He built a physical status lamp for Claude Code: solid green for idle, blinking yellow for thinking, blinking red for running tools, wired through Claude Code hooks to a Node serial bridge to an Arduino UNO and a three-color LED. It's a tiny build, but it's a smart read on what agentic work actually changes, you're no longer staring at output, you're supervising a process, and ambient signals beat a wall of text.
@dcfgod [Claude Code]
https://x.com/dcfgod/status/2057992546404769794
He uses Claude Code directly for on-chain analysis and it just figures itself out, finds an RPC, writes scripts, and digs. He asked it to investigate etherfi token flows to find who's selling and then drill into individual wallets, and got genuinely useful findings: team wallets not selling, ETHFI staking only going up, CEX inflows and outflows roughly balanced, and a few early investors who waited for the all-time lows finally capitulating but none at large size. His read on the market, that it's just illiquid, came out of an analysis he didn't have to write by hand.
@xjuntaro [Claude Code]
https://x.com/xjuntaro/status/2058122894815924681
A founder he met is running an AI-powered rollup play: buy a sleepy membership business (an association, a mutual-aid society) where a clerk manually manages the member roster, dues collection, chasing payments, and event planning, then rebuild it with Stripe and Supabase to create a member portal, cut the clerical headcount, and sell once margins improve. His observation is the sharp one, this is a one-day implementation with AI, and the engineers good enough to use Claude Code are often too sophisticated to notice these unglamorous businesses exist. The arbitrage isn't the tech, it's seeing where the boring money is.
@rohit4verse [Claude Code]
https://x.com/rohit4verse/status/2058272712653746581
The pitch: a marketing team used to cost a payroll, now it costs $200 a month. A veteran who spent $100M+ running ads builds an entire funnel solo in Claude Code in one sitting, positioning, landing page, lead magnet, ads, SEO. He's honest about the tax, though, the engineer's road means wiring up Perplexity, Playwright, and Firecrawl plus 17 hand-trained skills before any of it earns its keep, which is a real setup cost, not a free lunch. That candor about the upfront work is what makes it credible.
@mvanhorn [Claude Code]
https://x.com/mvanhorn/status/2058189714088456687
He built a Tesla skill that works across Claude Code, OpenClaw, and Hermes: "unlock the car" and "turn on dog mode" as one-line commands callable from phone or laptop, an agent that defrosts the car at 7:50 every winter school morning before dropoff, a charging-cost ledger, and a Supercharger-queue watcher his agent can page him about. The signing key stays on his own host. It's a clean example of agents reaching out of the terminal and into physical daily routines.
@yutosuzuki [Claude Code]
https://x.com/yutosuzuki/status/2058135863436198322
He's been building a 60-plus-file set of SKILLS, RULES, and HOOKS he calls his "initial harness," and he just did the work to make it usable by other people. Two moves: he catalogued all 60+ files by purpose and category into an infographic (and discovered he hadn't fully grasped his own system until he organized it), and he built a distribution skill so a recipient can bulk-download via Google Drive and run with minimal setup. His insight is worth keeping, executable knowledge and documentary knowledge are fundamentally different things, and making a skill handoff-ready forces dependency cleanup, secret separation, and a re-audit of your own architecture.
@d01490201 [Claude Code]
https://x.com/d01490201/status/2058244269429965130
A scientist points out that ordinary programmers can't write software that accounts for crystallography and electron-beam behavior, and the commercial packages that do cost in the high tens to hundreds of thousands of yen. With Claude Code, his own domain knowledge turns directly into working programs, and he says it's genuinely fun. This is the underrated frontier, deep-domain experts who were never going to hire a dev now shipping specialized scientific tooling themselves.
@leafmeta [Claude Code]
https://x.com/leafmeta/status/2058033615142068599
He reverse-engineered Claude Code and found the AI decision logic is only 1.6% of the code, the other 98.4% is permission, session, tool, and context management, all sitting under a five-layer stack whose real core is the agent loop, the model calling a tool, getting a result, and judging again. He connects it to Anthropic renaming the "Claude Code SDK" to the "Claude Agent SDK" last year, a declaration that the loop and the boundaries, not the model, make the agent. His conclusion is the meta-lesson of the whole day: people who just swap models lose time, people who can design the loop and the guardrails pull several times the output from the same model.
@Mnilax [Claude Code]
https://x.com/Mnilax/status/2058283663805047224
While everyone celebrated Claude prompting itself on stage, he went and audited the settings.json that self-prompting actually runs on, 125+ keys, only ~40 documented, four of his own not in the docs anywhere. He found 18 settings that actually move the bill, including a known deny-rule bug where your config says block but the binary reads anyway, and a 10% Opus tax most apps pay by default through inference_geo. This is the unglamorous, money-saving work the hype skips over, and it's exactly aligned with the day's cost panic.
@svpino [Claude Code]
https://x.com/svpino/status/2058167324130398408
He's letting Claude Code run on his Omarchy install and edit the OS config files directly, and notes how neat it is that config files drive most of the OS and current models handle them really well. His waybar was misbehaving and he couldn't figure out why; Opus did, and fixed it. A small but telling case of the agent moving from writing apps to administering the machine itself.
🗣 User Voice
User Voice
The mood shifted hard from capability to economics this week, and the asks are getting specific. A few clear signals from how people talked:
Cost is now the number-one constraint, not capability. Microsoft canceled internal Claude Code licenses, Uber burned its 2026 AI budget in four months, and tens of thousands of devs are routing Claude Code through DeepSeek and Kimi via free proxies to zero out their bill. Users want real cost controls and visibility, and they're celebrating /usage and gateway tools that expose per-skill spend. (@AskYoshik, @rdd147)
Plugin and context bloat is the new self-inflicted wound. Multiple heavy users independently discovered that their carefully assembled plugin stacks were eating 30-40% of the context window before work even started, and that stripping back to built-ins felt like upgrading the model. The ask is leaner, smarter defaults. (@regent0x_, @L1vsun)
The structure beats the model, so make the structure better. The loudest technical insight was that the harness, the loop, permissions, context management, is 98% of what makes Claude Code work, and people want native, first-class multi-agent orchestration instead of bolting it on. (@leafmeta, @_vmlops)
Latency is quietly driving defection to Codex. Several users said Claude Code's response time makes them hesitate to even ask a quick question, and that the moment switching costs drop, they'll move. Speed is now a retention issue. (@suna_gaku, @akihiro_genai)
Persistent, portable memory is still unsolved enough that people are bolting on third-party layers and comparing Claude's project memory against Codex's AGENTS.md model. They want memory that survives sessions and that they own. (@CMhOeNnExY)
And underneath it all, a pull toward agent-native everything: tools designed for agents to talk to directly rather than for humans to click, and a refusal to be locked to one model provider. (@davidyinai)
The mood shifted hard from capability to economics this week, and the asks are getting specific. A few clear signals from how people talked:
Cost is now the number-one constraint, not capability. Microsoft canceled internal Claude Code licenses, Uber burned its 2026 AI budget in four months, and tens of thousands of devs are routing Claude Code through DeepSeek and Kimi via free proxies to zero out their bill. Users want real cost controls and visibility, and they're celebrating /usage and gateway tools that expose per-skill spend. (@AskYoshik, @rdd147)
Plugin and context bloat is the new self-inflicted wound. Multiple heavy users independently discovered that their carefully assembled plugin stacks were eating 30-40% of the context window before work even started, and that stripping back to built-ins felt like upgrading the model. The ask is leaner, smarter defaults. (@regent0x_, @L1vsun)
The structure beats the model, so make the structure better. The loudest technical insight was that the harness, the loop, permissions, context management, is 98% of what makes Claude Code work, and people want native, first-class multi-agent orchestration instead of bolting it on. (@leafmeta, @_vmlops)
Latency is quietly driving defection to Codex. Several users said Claude Code's response time makes them hesitate to even ask a quick question, and that the moment switching costs drop, they'll move. Speed is now a retention issue. (@suna_gaku, @akihiro_genai)
Persistent, portable memory is still unsolved enough that people are bolting on third-party layers and comparing Claude's project memory against Codex's AGENTS.md model. They want memory that survives sessions and that they own. (@CMhOeNnExY)
And underneath it all, a pull toward agent-native everything: tools designed for agents to talk to directly rather than for humans to click, and a refusal to be locked to one model provider. (@davidyinai)
📡 Eco Products Radar
Eco Products Radar
Codex: the constant comparison point and increasingly the daily driver for "raw speed and final 20%" while Claude Code holds design and planning. Mentioned in nearly every workflow post.
Hermes Agent: Nous Research's local-first agent framework, paired with Claude Code/OpenClaw across investment bots, retro hardware, and Tesla skills. The open-source-runs-on-your-machine alternative people keep reaching for.
Google Antigravity: Google's Claude Code-alike CLI/IDE, heavily discussed, including hands-on reviews calling it underwhelming versus Codex and Cursor, plus Pichai's "12-hour OS build" demo.
codegraph: pre-indexed code knowledge graph cutting tool calls 70-94% and exploration cost ~35%, works across Claude Code, Codex, Cursor, opencode, Hermes. The most-cited context-efficiency tool of the day.
free-claude-code / 9router: proxies routing Claude Code traffic through 10-40 free providers (DeepSeek, Kimi, etc.), 20,000+ devs, with auto-fallback and ~40% token compression. The center of the cost rebellion.
agentmemory / SIBYL: persistent semantic memory layers via MCP, pitched as replacing paid memory APIs, with benchmark claims on hallucination and long-horizon retention.
opencode: open-source terminal agent, 157K+ GitHub stars, bring-your-own-model across 75+ providers, repeatedly framed as the free Claude Code.
Cowork: Anthropic's document automation (Word/Excel/PDF) and the wedge for non-technical professional adoption.
NotebookLM: paired with Claude Code/Hermes for research offloading to cut token spend.
Pi / oh-my-pi: the Pi harness ecosystem, positioned as the Arch-vs-Ubuntu alternative to Claude Code/Codex, strong on browser use.
Codex: the constant comparison point and increasingly the daily driver for "raw speed and final 20%" while Claude Code holds design and planning. Mentioned in nearly every workflow post.
Hermes Agent: Nous Research's local-first agent framework, paired with Claude Code/OpenClaw across investment bots, retro hardware, and Tesla skills. The open-source-runs-on-your-machine alternative people keep reaching for.
Google Antigravity: Google's Claude Code-alike CLI/IDE, heavily discussed, including hands-on reviews calling it underwhelming versus Codex and Cursor, plus Pichai's "12-hour OS build" demo.
codegraph: pre-indexed code knowledge graph cutting tool calls 70-94% and exploration cost ~35%, works across Claude Code, Codex, Cursor, opencode, Hermes. The most-cited context-efficiency tool of the day.
free-claude-code / 9router: proxies routing Claude Code traffic through 10-40 free providers (DeepSeek, Kimi, etc.), 20,000+ devs, with auto-fallback and ~40% token compression. The center of the cost rebellion.
agentmemory / SIBYL: persistent semantic memory layers via MCP, pitched as replacing paid memory APIs, with benchmark claims on hallucination and long-horizon retention.
opencode: open-source terminal agent, 157K+ GitHub stars, bring-your-own-model across 75+ providers, repeatedly framed as the free Claude Code.
Cowork: Anthropic's document automation (Word/Excel/PDF) and the wedge for non-technical professional adoption.
NotebookLM: paired with Claude Code/Hermes for research offloading to cut token spend.
Pi / oh-my-pi: the Pi harness ecosystem, positioned as the Arch-vs-Ubuntu alternative to Claude Code/Codex, strong on browser use.
Comments