Super User Daily: 2026-05-19
Karpathy joined Anthropic this week and the timeline lost its mind, but underneath the noise the actual builders kept shipping, and one pattern jumps out: Claude Code stopped being a code tool. People are running it as a financial analyst, a video editor, an academic writing partner, a 24/7 ops crew, even a lab instrument for visualizing protein binding. The other loud thread is money. Everyone is suddenly counting tokens out loud, because the bill is exactly where the magic stops being free, and the people getting the most out of these agents are the ones who treat token spend as a budget to engineer, not an accident to absorb. Here is what people actually did this week.
@hqmank [OpenClaw]
https://x.com/hqmank/status/2056585395182420084
OpenClaw founder Peter Steinberger says he is burning roughly $1.3M a month on tokens to keep about 100 Codex agents running. That is not a typo and it is the most honest data point of the week: this is what it actually costs to run a swarm at full tilt right now. His bet is that once tokens get cheap, code gets cheap too, and the real bottleneck becomes taste, tests, and stable releases. Worth sitting with if you think agent ops is a hobby budget.
@sadek [Claude Code]
https://x.com/sadek/status/2056691948975460482
Brand24 got a trophy from OpenAI for burning 10 billion tokens, and the founder figures they have done about the same again inside Anthropic and Claude Code. The plan is for those trophies to end up as monitor stands and paperweights. The real story is underneath: his whole team now builds automations, internal mini-products and external tools as a daily habit. Token spend at this scale is not waste, it is the physical receipt of a team that turned every employee into a builder.
@DealsDhamaka [Claude Code]
https://x.com/DealsDhamaka/status/2056804088181076324
Stuck in three hours of Bengaluru traffic, this guy built an end-to-end agentic workflow for a Pharma R&D use case entirely inside Claude Code. He just collaborated with the AI, assembled artifacts and agents, and plugged in the right skill packs. What he says would normally take a three-person team three weeks of coordination, development and review got compressed into one evening in a car. This is the non-coding domain story that matters: a regulated, jargon-heavy field collapsed into a single commute.
@adriano_viana [Claude Code]
https://x.com/adriano_viana/status/2056708177778135259
He spent real time testing Claude Code's /goal command, the one that works while you sleep, and his biggest takeaway is architectural: it is not a dumb loop, it is two models. The main Claude executes, and a Haiku acts as an external referee that reads the success condition you wrote and decides each turn whether it has been met. If not, Claude keeps going. That referee pattern is the quiet unlock for trusting an agent to run unattended.
@seelffff [Claude Code]
https://x.com/seelffff/status/2056742772879622350
One plugin install, and Claude worked autonomously for two hours without drifting off the plan. The plugin is obra/superpowers. What it actually does is force discipline before code: it stops before writing a single line, asks what you are really building, writes a spec, shows it in readable chunks and waits for sign-off, then breaks the work into 2-5 minute tasks with exact file paths and verification steps. The lesson keeps repeating this week: the model is fine, the harness is the difference.
@AYi_AInotes [Claude Code]
https://x.com/AYi_AInotes/status/2056685348156055980
She fed Claude Code one prompt in xhigh mode, let it run for two minutes, and got back a complete interactive HTML page for a 2026 org-and-talent review. Dark theme, horizontal timeline, collapsible risk tables, a team avatar wall, progress bars, all interactions working, no polish skill needed, just open it and send it to the boss. The same job in Notion used to take her 30-40 minutes of template wrangling. The kicker: a single self-contained HTML file, all CSS and JS inline, under 700 lines.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2056657199070695790
He fully automated video editing with Claude Code. Drop the raw footage into a folder and it sorts the clips, cuts the dead parts, adds jump cuts to pick up the pace, and burns in subtitles. He went off to another meeting and came back to a finished edit. The part he praises most: it understands the content well enough to make sensible cuts and even paraphrase the captions. This is Claude Code earning the nickname people keep giving it, Claude Work.
@AliAlkhuzaee_ [Claude Code]
https://x.com/AliAlkhuzaee_/status/2056607635156181455
In under sixty seconds he turned Claude into a hardcore financial analyst by wiring in the Financial Datasets MCP covering 17,000+ stocks. Now his coding assistant analyzes company balance sheets, watches crypto prices, and spits out precise earnings reports on a keystroke. Three steps: add the MCP, do the OAuth in the browser, start asking about any ticker. This is the quiet pattern of the week, plugging one MCP into Claude Code and walking out with a domain tool.
@kirubaakaran [Claude Code]
https://x.com/kirubaakaran/status/2056576427974349105
A few prompts in Claude Code and he had a working Trade Journal Portal. He uploads his broker CSV from DhanHQ and it generates P&L broken out by date, day, week, month and DTE, plus a calendar heatmap, cumulative returns and a drawdown chart. He even shared the exact prompt he used. This is the kind of thing that used to be a paid SaaS subscription, now it is an afternoon and a CSV.
@WiseInvest513 [Claude Code]
https://x.com/WiseInvest513/status/2056719647555399887
His morning used to be 35-45 minutes of low-value tab-hopping: Yahoo Finance, Xueqiu, a Twitter market digest, three broker apps to check positions, TradingView. Now he is using vibe coding to compress that and has wired Airtap into his investing flow, an AI that taps, scrolls, types and navigates his phone like a real person. Write the prompt once and it runs every day for him. Set it once, run it forever, his words.
@levelsio [Claude Code]
https://x.com/levelsio/status/2056705090073870460
Levelsio's safety tip is dead simple: ask Claude Code to security-audit your devices. He already does it on his VPS servers and tried it on his MacBook Pro, where it did a solid job. It caught a pile of unsecured stuff, including FileVault he forgot to turn on, plus some local networking issues. The prompt is literally just can you security audit my computer. No tooling, no setup, real findings.
@oikon48 [Claude Code]
https://x.com/oikon48/status/2056571858242404578
He spotlights how NOT A HOTEL deployed Claude Code across an enterprise org, and it is a reference playbook. Access control via domain claim plus mandatory Okta SSO and automated IdP provisioning so shadow tenants cannot structurally exist. Guardrails through Managed Settings pushed over MDM, with API keys held in 1Password CLI to avoid plaintext files and shrink the prompt-injection surface. Cost is metered per usage through the Anthropic Console with seats handed out only to proven users. This is what serious enterprise Claude Code looks like.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2056562488066359494
His job today is almost 90% building agent harnesses for enterprises, and the other 10% is eval, testing, observability and improvement. He is blunt about the state of it: Anthropic's enterprise plan lets you centralize how Claude Code runs across whole dev teams, but there is no manual and no book that teaches this at scale. They are writing the best practices as they deploy each new configuration. That is what a brand-new profession looks like before the textbooks exist.
@hinabe_ch [Claude Code]
https://x.com/hinabe_ch/status/2056621517216350620
He took a sizable AI game-world project where he had estimated 420 person-months, and AI coding compressed it to 150. Then reality bit: a problem hit a critical component, and of the 15 outsourced engineers he hired, only 2 could actually tell the AI specifically how to fix what it broke. So 80% of the AI-coding fixes piled onto those 2 people. He sent 10 of the other 13 back to the agency and still finished the pilot half a year early. The honest footnote: this feels like a bargain-sale window, not a free lunch.
@altudev [Claude Code]
https://x.com/altudev/status/2056711553211158591
At an ASELSAN event for high school students, he and a colleague built a Suno-style music generation tool from scratch live, pure vibe coding in Claude Code. They used MiniMax Music 2.5 to generate the tracks and GPT-5 to write the lyrics, wiring multiple AI models together on the access layer. The point was not the tool, it was showing teenagers that you can stand up a real multi-model product in front of an audience. That is the demo that recruits the next cohort.
@showheyohtaki [Claude Code]
https://x.com/showheyohtaki/status/2056543858977304902
He built a skill that generates commercial video with Claude Code plus Higgsfield. You type in the company name and product name, answer a few questions, and it produces the video. His reaction lands hard: he once spent millions of yen making a TV commercial, and now it feels like he can make one for just the cost of the AI. The whole creative production chain from concept to motion is collapsing into a guided prompt.
@adityarao310 [Claude Code]
https://x.com/adityarao310/status/2056687979037405206
The interesting move in Higgsfield's new workflow is not the video generation, it is everything before it. Claude Code pulls visual references through Pinterest, builds a mood and style direction, then generates a six-scene storyboard using GPT Images through the Higgsfield MCP, complete with on-screen text. Only after all that does Seedance animate the scenes. The pipeline treats generation as the last step, not the first, which is exactly how a real creative team works.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2056800169866441207
He thinks people are underrating AI analytics tools, and he has the workflow to back it. With Claude Code he connected LogRocket and PostHog to his site in minutes, then ran their AI across about ten prompts to dig into user behavior: the most common path from homepage to exit, where users hesitate longest, the last 30 seconds before a bounce. He says the results were wild and the AI nailed 4 or 5 out of 5 on every query. Analytics without a data team, assembled in an afternoon.
@JeremyNguyenPhD [Claude Code]
https://x.com/JeremyNguyenPhD/status/2056577908823785495
He surfaces a Claude Code use case from outside engineering entirely: Paul Goldsmith-Pinkham, an economist at Yale, using Claude Code for academic writing and thinking. There is a blog post and a 45-minute video walking through it. This is the corner of the ecosystem that gets drowned out by the coding noise, a serious researcher treating the agent as a thinking partner for papers, not a code generator. Expect more of this from academia as the year goes.
@bearliu [Claude Code]
https://x.com/bearliu/status/2056632703659684236
Two-plus months into Obsidian and his biggest win is wiring it together with Claude Code, using the vault as the container for an externalized knowledge and content base that Claude Code can call. His daily life records and data flow into Obsidian automatically, and the AI then proofreads its own output against that ground truth, getting sharper the more he uses it. How far does it go? It even auto-captures the things he mumbles in his sleep into the database. A personal memory layer that actually compounds.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2056823220922433759
He built a skill that applies behavioral psychology to marketing: cognitive biases, persuasion principles, pricing psychology and decision architecture, with ethical guardrails baked in. You describe your marketing challenge and it identifies which psychological principles apply, how to implement them ethically, and where they will move conversion most. His framing is the real lesson: understanding why people buy is more valuable than another generic copy generator. Skills are becoming where domain expertise lives.
@op7418 [Claude Code]
https://x.com/op7418/status/2056575297810751506
He had finished a skill that turns a frontend into an explainer video, then wanted to optimize cost, so he had Claude Code run a 40-second case and analyze exactly where the tokens went. The numbers: a 40-second video burned 1.45 million tokens, input dominated, output was just 0.7%, and the video-composition and skills steps ate the biggest share. With Claude Code, 92% of that hit cache, so it was survivable. This is the kind of token forensics that separates people who run agents from people who just poke them.
@defileo [OpenClaw]
https://x.com/defileo/status/2056754430100574640
He built an OpenClaw command center that looks like Iron Man's war room, an AI team running in real time on one dashboard. Every node is a 24/7 autonomous agent: a content command, the wire, a videographer, a ProductForge, an Instagram buddy, all running and reporting in. The build took one afternoon: export a PDF of his OpenClaw agents, feed it into Lovable, push to GitHub, and have Claude Code handle the rest. The hate-it-because-it-is-free energy is real.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2056785706308026509
He rewired the cron output on his OpenClaw and Hermes setup. The old way: a dozen-plus cron runs a day, each dumping 2000 words into Telegram, and his brain was mush by evening. Now each run delivers three things, Markdown for the next agent, HTML for him, and just 5 lines into Telegram, costing 51ms extra render. More crons every day, less to read every day. A few weeks in, the AI employee crew runs on its own and he just scans 5 lines to check on it.
@jordymaui [OpenClaw]
https://x.com/jordymaui/status/2056700108306043032
Most people think their OpenClaw subagents are broken; his weren't, they were timing out. Every long research or build job kept dying with request timed out before a response was generated. The fix was one config setting, agents.defaults.timeoutSeconds set to 7200. Important caveat from his trial and error: do not set it to 0, OpenClaw won't accept that. A tiny, hard-won knob that is the difference between a long-running agent finishing and quietly dying.
@araichuu [Claude Code]
https://x.com/araichuu/status/2056855320576819667
She hit a wall, in the best way. Bringing in Claude Code exploded the amount of work she can and wants to do, even opening a new revenue line, an AI-plus-JSON fortune-telling thing that pulled a million page views. Given all that, doing her own bookkeeping no longer makes sense, so she begged her accountant to take her back on a fully-outsourced plan where she just mails in the receipts. She had taken accounting in-house during the COVID revenue crash to cut costs; now she is buying her time back to focus on earning.
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2056571055074951250
He is proud of a feature in his latest build, and he should be: he added guidance to VoiceClaw so that people with limited control of their hands can use it through iOS Accessibility Voice Control. Custom commands let them open and quit the app, turn the in-app mic on either by a voice command that taps the screen or an optional autostart-on-open, and more. This is the use case that never trends but matters most, an agent built to give someone back their independence.
@design_proteins [Claude Code]
https://x.com/design_proteins/status/2056753397077102724
Short and sharp: a researcher made a PD1-PDL1 binding visualization with Claude Code. No engagement-bait thread, no course pitch, just a working scientific visualization of a protein-protein interaction that sits at the center of modern cancer immunotherapy. This is the quiet end of the spectrum where Claude Code becomes a lab instrument, not a code editor.
@fba [Claude Code]
https://x.com/fba/status/2056787773013537164
He is done with WordPress and not looking back. No more paying for plugins he does not need, optimizing Redis, hopping through server caches and who-knows-what else to keep performance barely decent. His current content tech stack is Claude Code plus GitHub plus Railway plus JavaScript, and if he needs a CMS, Strapi. This is the small but telling pattern: people are not just building apps with Claude Code, they are dismantling the SaaS subscriptions they were tired of paying.
@tdinh_me [Claude Code]
https://x.com/tdinh_me/status/2056645921023099236
He built a mobile app to check his Paddle revenue, because Paddle does not ship one. You plug in a read-only scoped Paddle API key and get live data with native SwiftUI graphs, multi-account support and unified revenue metrics, plus home-screen widgets. Crucially the data stays on device with no server, since API requests go straight from the phone. He made it free. This is the new default move: vendor missing a feature, build the app yourself over a weekend.
@walls_jason1 [Claude Code]
https://x.com/walls_jason1/status/2056753809024504130
He adopted Karpathy's LLM Wiki schema on May 9 as the backbone for his Claude Code memory system, three layers: raw, wiki, schema. Then this week the person who taught that exact pattern joined the team that ships the product he builds on. Beyond the nice coincidence, the substance is real: a structured, layered memory design is what turns Claude Code from an amnesiac into something that compounds knowledge across sessions.
@_avichawla [Claude Code]
https://x.com/_avichawla/status/2056657776924078555
Claude Code used 3x fewer tokens after one change, and it was not a better CLAUDE.md or a better model. The fix: before writing a single line, the agent needs to know the existing tables and schemas, configured auth providers, storage buckets, deployed edge functions, available AI models and active RLS policies. Most backends expose this via MCP, so feeding that context up front stops the agent from blindly rediscovering the backend every time. Context engineering, not prompt magic, is where the savings live.
@commte [Claude Code]
https://x.com/commte/status/2056800888514138493
Small but genuinely useful: Claude Code's Learning mode keeps his thinking moving. You switch it on via /config, Output style, Learning. The payoff is the star Insight notes, where it briefly explains why it chose that implementation. For anyone trying to actually get better instead of just shipping faster, having the agent narrate its reasoning in-line is an underrated way to learn from your own tool.
🗣 User Voice
User Voice
Token cost is the number-one nerve this week, and people are saying it out loud. @Anaya_sharma876 says the $20 plan has felt like a free plan for weeks because of usage-limit bugs, @agusbuilds warns Claude Code can burn hundreds of tokens in minutes when your context is wrong, and @hqmank points at the extreme: an OpenClaw founder torching $1.3M a month. The takeaway is consistent, the bottleneck is no longer the model, it is whether you can engineer your context to stop the meter running.
Memory and self-improvement are the second drumbeat. @saidstetic and @mnmn94253156337 keep hitting the same wall, a new session forgets everything, and @alexaiworks puts the sharper point on it, remembering context is not the same as actually improving future behavior. People want a structured learning loop, not just a bigger CLAUDE.md.
Run it in the cloud, not on an always-on PC. @udiWertheimer keeps hammering that Codex and Claude Code defaulting to local makes no sense in 2026 when most of the world lives on a phone, and the rush toward Google's cloud-based Spark this week proves the appetite is real.
Observability is the quiet ask. @stackzz nails it, at 1am the scoreboard that matters is what files changed, what assumption the agent made, and which tests actually ran, not which model is hot. People are building token dashboards (@geekbb) and tracing tools just to see what their agent is really doing.
And the sore spot for OpenClaw users specifically: @jamil_zakirov flags that Anthropic is now the one company that bans OpenClaw usage, a recurring frustration about third-party harnesses getting locked out at the terms-of-service level.
Token cost is the number-one nerve this week, and people are saying it out loud. @Anaya_sharma876 says the $20 plan has felt like a free plan for weeks because of usage-limit bugs, @agusbuilds warns Claude Code can burn hundreds of tokens in minutes when your context is wrong, and @hqmank points at the extreme: an OpenClaw founder torching $1.3M a month. The takeaway is consistent, the bottleneck is no longer the model, it is whether you can engineer your context to stop the meter running.
Memory and self-improvement are the second drumbeat. @saidstetic and @mnmn94253156337 keep hitting the same wall, a new session forgets everything, and @alexaiworks puts the sharper point on it, remembering context is not the same as actually improving future behavior. People want a structured learning loop, not just a bigger CLAUDE.md.
Run it in the cloud, not on an always-on PC. @udiWertheimer keeps hammering that Codex and Claude Code defaulting to local makes no sense in 2026 when most of the world lives on a phone, and the rush toward Google's cloud-based Spark this week proves the appetite is real.
Observability is the quiet ask. @stackzz nails it, at 1am the scoreboard that matters is what files changed, what assumption the agent made, and which tests actually ran, not which model is hot. People are building token dashboards (@geekbb) and tracing tools just to see what their agent is really doing.
And the sore spot for OpenClaw users specifically: @jamil_zakirov flags that Anthropic is now the one company that bans OpenClaw usage, a recurring frustration about third-party harnesses getting locked out at the terms-of-service level.
📡 Eco Products Radar
Eco Products Radar
Codex - the constant co-pilot to Claude Code, now runnable inside it; the pairing dominates every workflow this week.
Cursor - still the default comparison point, though several users report drifting away from it.
Hermes - OpenClaw's main local-first rival, paired with it constantly in real setups.
Google Antigravity / Gemini Spark - Google's cloud-based, always-on answer to OpenClaw and Codex, the week's biggest competitive splash.
Lovable - the no-code app builder people pipe into Claude Code for fast MVPs and dashboards.
Higgsfield - the MCP powering the Claude-Code-to-video creative pipelines (storyboards, motion design, commercials).
Obsidian - the markdown vault people wire to Claude Code as a personal memory and knowledge layer.
GBrain - Garry Tan's open-source memory/retrieval base running behind OpenClaw and Hermes brains.
n8n - the automation backbone in the one-person-business stack alongside Claude Code.
MiniMax - models and MiniMax Music showing up in free-model proxies and creative builds.
PolyAI - shipped an ADK that plugs voice agents straight into Claude Code and Cursor.
Warp (Oz) - multi-harness cloud orchestration for Claude Code, Codex and the Warp Agent.
Codex - the constant co-pilot to Claude Code, now runnable inside it; the pairing dominates every workflow this week.
Cursor - still the default comparison point, though several users report drifting away from it.
Hermes - OpenClaw's main local-first rival, paired with it constantly in real setups.
Google Antigravity / Gemini Spark - Google's cloud-based, always-on answer to OpenClaw and Codex, the week's biggest competitive splash.
Lovable - the no-code app builder people pipe into Claude Code for fast MVPs and dashboards.
Higgsfield - the MCP powering the Claude-Code-to-video creative pipelines (storyboards, motion design, commercials).
Obsidian - the markdown vault people wire to Claude Code as a personal memory and knowledge layer.
GBrain - Garry Tan's open-source memory/retrieval base running behind OpenClaw and Hermes brains.
n8n - the automation backbone in the one-person-business stack alongside Claude Code.
MiniMax - models and MiniMax Music showing up in free-model proxies and creative builds.
PolyAI - shipped an ADK that plugs voice agents straight into Claude Code and Cursor.
Warp (Oz) - multi-harness cloud orchestration for Claude Code, Codex and the Warp Agent.
Comments