Super User Daily: 2026-05-26
The loudest story today was the cost panic. Microsoft pulling Claude Code licenses, Uber burning its whole 2026 AI budget by April, Nvidia's VP admitting compute now costs more than headcount. It got reposted in a dozen languages and most of it was half-true product substitution dressed up as an AI-is-dead headline. But scroll past the screaming and the actual signal is the opposite: people are quietly putting these agents into jobs that have nothing to do with writing code. A hedge fund running 750 people on it. A recruiter shortlisting CVs. A tax accountant killing spreadsheet work. The interesting part of the day isn't the bill, it's where the work is going.
@GoshawkTrades [Claude Code]
https://x.com/GoshawkTrades/status/2058562091733458955
Man Group, a 200 billion dollar hedge fund with around 1,700 employees, now has 750 of them using Claude Code every day, across developers, quants, finance, and HR. That's 44 percent of the entire firm on one tool. And it's that broad base that let them push AI-generated trading signals into production managing real capital, not a demo. This is the clearest enterprise data point of the day: the value isn't a flashy autonomous agent, it's getting nearly half a company fluent in the same tool.
@showheyohtaki [Claude Code]
https://x.com/showheyohtaki/status/2058509055862456545
He runs his entire sales process through Claude Code, and the workflow is genuinely worth copying. He builds a persona agent fed with his product docs and target customer profiles so it acts as that company's sales director. It generates the full call script from discovery to close, then he records the meeting (Notta plus Notion, triggered from his Apple Watch when he's out), and afterward feeds the transcript back so Claude scores the meeting, names the top three lines that landed and the top three to fix, and rewrites the script for next time. Sales as a closed feedback loop.
@teovito [Claude Code]
https://x.com/teovito/status/2058428480162054148
Helping a friend hire a software engineer in Dubai, he dumped 150-plus CVs into a folder and asked Claude Code to evaluate every profile and shortlist the ten best, a task he calls miserable to do by hand on Indeed. He then ran the exact same job through Codex as a sanity check and it picked the identical ten people. Then he had Claude draft personalized outreach to each one. Recruiting screening, done in a terminal session, cross-validated by two agents.
@sanpouyoshikici [Claude Code]
https://x.com/sanpouyoshikici/status/2058668554950721674
A concrete equity-screening case: he wired Claude Code to Japan's EDINET filing database and ran his fifteen-point checklist across the market, narrowing 4,000 listed companies down to 55. His point is sharp and underrated: tiny differences in the screening conditions swing how many names survive, which is exactly the kind of judgment work that used to take an analyst days and now runs as a query.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2058647095779926391
Claude Code connects directly to Shopify's official AI Toolkit, and most DTC brands have no idea it exists. One prompt and Claude reads your entire product catalog, rewrites every product description, and pushes the updates live to the store, all from the terminal. For an e-commerce operator sitting on hundreds of stale product pages, that's a week of copywriting collapsed into a single command.
@ZEIRISHI_Ichibe [Claude Code]
https://x.com/ZEIRISHI_Ichibe/status/2058514854106050918
A tax accountant watching his own profession change in real time. He notes that the hours-long Excel grind his firm lives in, summarizing securities holdings, doing the cash-flow ledgers for religious-corporation closings, hand-building real-estate income tables, would likely drop to under a third of the time with Claude Code. It's the honest insider read: not hype, just a working professional doing the math on his own billable hours.
@kawai_design [Claude Code]
https://x.com/kawai_design/status/2058472605431255376
A social-media marketer's full Claude Code playbook, nine concrete tasks: turn 100 post ideas into a CSV tagged by theme, hook, CTA and audience; classify past posts into what worked and what to reuse; break a long note into X-sized posts; mass-produce webinar announcements by pain point and job title; draft comment replies; build the posting calendar across the day's peak hours; structure competitor posts into a table; rewrite weak thumbnail copy; and summarize reports down to numbers and next actions. He starts with one folder and only lets it touch Markdown and CSV.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2058494831647719718
Proof that the faceless-YouTube thing isn't just coding-bro fantasy: he runs math channels in the Numberphile lane, built on Claude Code plus Obsidian plus Python with Manim for the animations. The numbers he posts are 25,418 dollars over 90 days and a 311 percent sales lift across two channels, on roughly 20 dollars a month in API costs, no face, no voice. The second channel, he says, cost one afternoon to spin up.
@growing_daniel [OpenClaw]
https://x.com/growing_daniel/status/2058365643687395717
A snake showed up in his backyard and instead of panicking he sent a photo to his OpenClaw. It identified it as a harmless gopher snake, gave him a local number to call to confirm, and even suggested ways to catch it to keep as a pet for the kids. He frames it as the kind of small everyday moment, usually less dramatic, that quietly shows how an always-on personal agent changes daily life. Non-coding, ambient, and genuinely useful.
@Vivek4real_ [OpenClaw]
https://x.com/Vivek4real_/status/2058628348083892471
Someone built an OpenClaw agent that sells pool installations on autopilot. It finds homes in the 500K to 1.2M range that don't have a pool, renders a pool into a photo of their actual backyard, and mails them a before-and-after postcard. It's a tight, fully-automated outbound sales loop aimed at a specific physical-world niche, the kind of thing that used to need a marketing agency and a designer.
@petergyang [OpenClaw]
https://x.com/petergyang/status/2058555226479866312
From an episode with Ryan Carson on how he runs his startup solo with AI agents. OpenClaw is his AI chief of staff, triaging emails, booking meetings, and doing sales outreach, while Codex and Devin act as his engineering team shipping features while he sleeps. The framing that stuck: you used to build the MVP, now you build the system that builds the MVP. A one-person company with a full org chart made of agents.
@holzzak777 [OpenClaw]
https://x.com/holzzak777/status/2058402980840538276
He runs a paid Discord that he wanted to turn into a kind of shared office for stock traders, where everyone trades and he just provides the service. The catch: he has zero development ability and had been carrying the idea around in his head for years, unable to build it. A single 500-dollar OpenClaw setup turned all of it into reality. The clearest no-code-person-ships-real-thing story of the day.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2058661838431441174
A clean, replicable lead-gen recipe in one terminal session: open Claude Code, make a project folder, drop in a CLAUDE.md with your business context, have it define your ideal customer profile from that doc, plug in an Exa API key stored in a .env so it never leaks, then ask it to pull 100 companies matching the ICP and enrich them. No scraping tool, no spreadsheet wrangling, just context plus an API key.
@morpphhhaw [Claude Code]
https://x.com/morpphhhaw/status/2058454273785016344
A weekend project earning 285 dollars a month that runs while he sleeps: a tool that monitors competitor pricing pages and emails weekly reports to e-commerce store owners. Total build was 36 hours and 12 dollars in API credits on one Claude Code subscription. Nothing revolutionary, just a painful three-hours-a-week manual chore automated, and crucially he validated demand before writing a single line.
@Raytar [Claude Code]
https://x.com/Raytar/status/2058688745801662545
Provocative but real: the best-performing OnlyFans page he saw this year isn't a person, it's four files and a 20-dollar Claude subscription. Some kid built the persona in Claude Code, face, content, posting schedule, DMs, all automated, and reportedly pulled 43,000 dollars in 30 days running it from his bedroom. Whatever you think of the use case, it's a stark example of a fully synthetic media business run end to end by one agent setup.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2058571995944517710
His site got hacked, so he handed the trojan file to Claude Code and got a full incident report back in seconds: the attack chain and severity all laid out. The root cause was a single word, nopriv, a WordPress hook that lets unauthenticated visitors call an endpoint with no permission check, which is how a lot of free themes get malware uploaded. The fix was simple (an Nginx rule blocking PHP execution in the uploads folder) and he open-sourced it. Security triage that used to mean an engineer reading PHP line by line.
@Tur24Tur [Claude Code]
https://x.com/Tur24Tur/status/2058577787691127017
Authorized testing on a production API endpoint. Opus 4.7 confirmed a SQL injection was real but couldn't pull database names, and sqlmap called it a false positive. So he switched to DeepSeek V4 Pro inside Claude Code and it figured out a blind technique: make the database answer yes/no questions by crashing on purpose, wrapping a CASE WHEN around two XML casts. A nice example of swapping the model under the same harness to crack a specific problem.
@lliu54827 [Claude Code]
https://x.com/lliu54827/status/2058643916468150641
A neat split on how to use two agents for trading. He runs his US-stock dashboard through Claude Code, auto-refreshed every 30 minutes via a launchd job that re-triggers the map. His take: Codex's 8am scheduled daily report is for passively reading the news, while Claude Code is for actively changing positions in real time, and the two don't conflict. A real recurring, scheduled financial workflow, not a one-shot prompt.
@girlinAI [Claude Code]
https://x.com/girlinAI/status/2058371201299591344
A guy wanted a reason to go outside, so over a weekend he vibe-coded a running app in Claude Code that turns your city into a battlefield: you run, you claim real streets on a map, and every road you touch becomes your territory. It's a small, complete consumer product built from a personal itch, the kind of thing that two years ago would have been a funded startup's MVP.
@eternityspring [Claude Code]
https://x.com/eternityspring/status/2058515934864908407
For the price of skipping one Starbucks, he built a physical hardware status light for Claude Code: green steady means idle waiting for input, yellow blinking means thinking, red blinking means running tools. Two cheap modules, an Arduino board (he had Codex adjust the code for esp32), and a 3D-printed case. A small, joyful maker hack that points at a bigger idea, wiring agents into the physical devices around you.
🗣 User Voice
User Voice
The cost question dominated everything, and it's no longer a footnote. @arthurkatcher reported burning over 1.2 billion tokens in cache reads and writes in a single day inside one session, and wants Claude Code to surface total cached tokens so people understand what a self-hosted harness without proper caching would actually bill. The recurring plea: give us real-time, total cost and cache visibility, not just a weekly stat.
The other half of the cost story is that most people blame the tool when the problem is their workflow. As @suni_code put it, you burn 75k tokens fixing a four-line bug, blame Anthropic, and repeat, while someone else runs seven agents in parallel and ships 12 PRs overnight for a tenth of the tokens. The difference is a CLAUDE.md and a plan, not the model. @Sean_CP_Founder made the same point more soberly: uncontrolled usage gets expensive fast when nobody defines the model tier, the review loop, or what good enough means per task.
Speed and reliability are the loudest product complaints. Multiple users (@sudoingX, @JapanTank, @oikon48) report Claude Code feeling sluggish, with 120-to-240-second hangs before tool calls, and ask for a Fast mode like Codex's even at the cost of tighter rate limits. There's a real perception gap right now between Claude Code and Codex on raw iteration speed.
Memory and context persistence is the feature everyone is hacking around. A whole cottage industry (@obsidianstudio9 and others on Obsidian vaults, the Agentmemory crowd) exists purely to give agents cross-session memory, and Anthropic's just-announced Memory Files feature is the official answer. The shared frustration: re-explaining your context every session is the single biggest tax on daily use.
The deeper note, voiced by @ZEIRISHI_Ichibe and @jinchenma_ai: the gap between the X bubble and reality is enormous. Inside the bubble everyone runs Claude Code; walk into a traditional manufacturer or a local firm and most people still treat a chatbot as a flaky novelty. The opportunity, and the knowledge-curse blind spot, both live in that gap.
The cost question dominated everything, and it's no longer a footnote. @arthurkatcher reported burning over 1.2 billion tokens in cache reads and writes in a single day inside one session, and wants Claude Code to surface total cached tokens so people understand what a self-hosted harness without proper caching would actually bill. The recurring plea: give us real-time, total cost and cache visibility, not just a weekly stat.
The other half of the cost story is that most people blame the tool when the problem is their workflow. As @suni_code put it, you burn 75k tokens fixing a four-line bug, blame Anthropic, and repeat, while someone else runs seven agents in parallel and ships 12 PRs overnight for a tenth of the tokens. The difference is a CLAUDE.md and a plan, not the model. @Sean_CP_Founder made the same point more soberly: uncontrolled usage gets expensive fast when nobody defines the model tier, the review loop, or what good enough means per task.
Speed and reliability are the loudest product complaints. Multiple users (@sudoingX, @JapanTank, @oikon48) report Claude Code feeling sluggish, with 120-to-240-second hangs before tool calls, and ask for a Fast mode like Codex's even at the cost of tighter rate limits. There's a real perception gap right now between Claude Code and Codex on raw iteration speed.
Memory and context persistence is the feature everyone is hacking around. A whole cottage industry (@obsidianstudio9 and others on Obsidian vaults, the Agentmemory crowd) exists purely to give agents cross-session memory, and Anthropic's just-announced Memory Files feature is the official answer. The shared frustration: re-explaining your context every session is the single biggest tax on daily use.
The deeper note, voiced by @ZEIRISHI_Ichibe and @jinchenma_ai: the gap between the X bubble and reality is enormous. Inside the bubble everyone runs Claude Code; walk into a traditional manufacturer or a local firm and most people still treat a chatbot as a flaky novelty. The opportunity, and the knowledge-curse blind spot, both live in that gap.
📡 Eco Products Radar
Eco Products Radar
Codex remains the most-mentioned tool by a mile and is the constant point of comparison, with a real migration narrative this week. OpenClaw and Hermes anchor the always-on personal-agent conversation. GitHub Copilot surged purely on the Microsoft substitution story. Below, everything mentioned three or more times across today's posts.
Codex (110) — OpenAI's CLI agent, the relentless comparison point and this week's migration destination
OpenClaw (72) — the always-on personal agent harness, snake IDs to pool-sales bots
Cursor (51) — still the favorite for staying close to the code and targeted edits
Hermes (29) — the stability-first OpenClaw alternative people keep switching to
GitHub Copilot (24) — Microsoft's in-house tool, the destination in the license-cut story
Gemini (19) — the go-to for Google Calendar and Gmail-heavy assistant setups
Obsidian (16) — the de facto external memory vault for content and knowledge factories
DeepSeek (12) — the cheap, high-cache-hit model people route Claude Code and Cursor to
OpenCode (10) — the open, model-agnostic terminal coding agent
Understand Anything (7) — turns a codebase into an interactive knowledge graph
Open Design / Claude Design (7) — local-first UI generation from skills and design systems
Claude Cowork (6) — the desktop agentic-workspace surface for non-coding work
Agentmemory (5) — bolt-on cross-session memory engine for Codex and Claude Code
Antigravity (5) — multi-account manager and emerging agent surface
GLM 5.1 (5) — the free-tier model bundled into the cost-avoidance routing stacks
CodeGraph (4) — local code knowledge graph, cuts tokens and tool calls hard
Tembo (3) — autonomous free agent that clones, codes, and opens PRs
Remotion (3) — the React video engine behind the faceless-content pipelines
Codex remains the most-mentioned tool by a mile and is the constant point of comparison, with a real migration narrative this week. OpenClaw and Hermes anchor the always-on personal-agent conversation. GitHub Copilot surged purely on the Microsoft substitution story. Below, everything mentioned three or more times across today's posts.
Codex (110) — OpenAI's CLI agent, the relentless comparison point and this week's migration destination
OpenClaw (72) — the always-on personal agent harness, snake IDs to pool-sales bots
Cursor (51) — still the favorite for staying close to the code and targeted edits
Hermes (29) — the stability-first OpenClaw alternative people keep switching to
GitHub Copilot (24) — Microsoft's in-house tool, the destination in the license-cut story
Gemini (19) — the go-to for Google Calendar and Gmail-heavy assistant setups
Obsidian (16) — the de facto external memory vault for content and knowledge factories
DeepSeek (12) — the cheap, high-cache-hit model people route Claude Code and Cursor to
OpenCode (10) — the open, model-agnostic terminal coding agent
Understand Anything (7) — turns a codebase into an interactive knowledge graph
Open Design / Claude Design (7) — local-first UI generation from skills and design systems
Claude Cowork (6) — the desktop agentic-workspace surface for non-coding work
Agentmemory (5) — bolt-on cross-session memory engine for Codex and Claude Code
Antigravity (5) — multi-account manager and emerging agent surface
GLM 5.1 (5) — the free-tier model bundled into the cost-avoidance routing stacks
CodeGraph (4) — local code knowledge graph, cuts tokens and tool calls hard
Tembo (3) — autonomous free agent that clones, codes, and opens PRs
Remotion (3) — the React video engine behind the faceless-content pipelines
Comments