Super User Daily: June 18, 2026
The pattern that jumps out today is the keyboard leaving the hot path. An Anthropic engineer's framing says it plainly: if you're watching Claude write code, you're the QA tester, and that's not the job. People are running /loop overnight, detaching with Routines, and driving sessions from their phones — and the binding constraint behind almost all of it is token cost, with one simple task burning 1.3M tokens and weekend bills measured in hundreds of dollars pushing a real migration onto local hardware. The other through-line is non-coding work: a published research paper from one prompt, a laid-off engineer who turned the job hunt itself into an agent pipeline, a wife who chats with a custom companion agent every day, an investor's 12-hour Meta analysis. And underneath the money, everyone serious keeps repeating the same lesson — the model is rented, the harness is the moat.
@hanakoxbt [Claude Code]
https://x.com/hanakoxbt/status/2066969072818872627
An Anthropic Claude Code engineer's line is the whole shift in one sentence: if you're watching Claude write code, you're the QA tester, and that's not what you're paid for. In 37 minutes he lays out how to get your keyboard out of the hot path entirely. /loop tells Claude to wake every 10 minutes and babysit your PRs while you're nowhere near the laptop, Routines do the same in the cloud with your machine closed, and remote control drives any session from your phone. The keyboard stops being the bottleneck.
@yibie [Claude Code]
https://x.com/yibie/status/2067013525298663433
A startup he works with threw out its entire engineering playbook and rebuilt it from a war-room meeting, and rule number one is: no coding before 10am. Twenty years of engineering culture chased maximum keyboard time; this team does the exact opposite, spending every morning pair-prompting together to define goals and set the agent up to succeed. The deeper flip is the manifesto: the agent, not the engineer, is now the one doing the work — the engineer's job is making sure the agent can do it well. Design for agents as the primary user, spec outcomes not process, review outputs not code.
@dotey [Claude Code]
https://x.com/dotey/status/2066777376399245356
A blunt token-economics complaint that doubles as a warning: Claude Code's dynamic workflows spun up 31 agents on one simple task, burned 1.3 million tokens, and jumped his weekly usage from 11% to 20% on the Pro 20x plan. His verdict is that it's just not worth it — he'd rather go slower. It's a useful counterweight to the 'unleash a swarm' hype: more agents is not free, and the bill is real.
@nasqret [Claude Code]
https://x.com/nasqret/status/2066826720473223334
A mathematician went all-in on 'loop philosophy' for two weeks and the frontier results are concrete, not vibes. A Codex loop on a Groebner-bases paper revamped the proof structure to remove all the heavy CPU computation, leaving only steps a human can check by hand. A sphere-packing loop verified Fisher's tables in Python and turned up a missing packing example the original work had skipped. And a Claude Code loop built a GitHub presentation teaching high-school teachers agentic workflows. This is autoresearch quietly doing real math.
@MushtaqBilalPhD [Claude Code]
https://x.com/MushtaqBilalPhD/status/2066832015908118730
A researcher had Claude Code produce a complete research paper — with an original argument — from a single prompt, says Pangram rated it 100% human-written, and the paper actually got published. Set aside the detector-gaming debate; the notable part is a non-coding, knowledge-work output of real academic weight coming out of a coding agent in one shot. He shared a step-by-step tutorial for reproducing it.
@wecraveai [Claude Code]
https://x.com/wecraveai/status/2066807136672534876
A guy got laid off, built an AI job-search system on Claude Code, evaluated 740+ job offers with it, landed a Head of Applied AI role, then open-sourced the whole thing as career-ops. One slash command turns a job URL into a structured A-to-F evaluation, an ATS-optimized PDF tailored to that exact role (rendered via Playwright), salary research, interview prep, and a tracker entry. Fourteen skill modes, a 45+ company portal scanner, batch evaluation via Claude sub-agents. This is the job hunt itself turned into an agent pipeline.
@amandaorson [OpenClaw]
https://x.com/amandaorson/status/2066927382783504598
Moved off OpenClaw onto Nous Research's Hermes and the use cases are pure ops, not coding. A daily digest (paired with the Last 30 Days skill) surfaces real-estate-investor pain points across X, Reddit and YouTube every morning. A project-management agent reads Linear, Slack and standup notes to set tomorrow's agenda, flag what was agreed verbally in Slack but never logged in Linear, and hold people accountable. A chief-of-staff schedules across two businesses plus farm chores by weather. Setup took an hour versus 2-3 days for OpenClaw.
@catalinmpit [OpenClaw]
https://x.com/catalinmpit/status/2066934168404299832
The honest answer to 'Hermes/OpenClaw has no use case': it doesn't make sense until it clicks for you. His agent handles everything from simple reminders to managing his finances, monitoring his email, and making code changes — all from his Telegram account. The reason it's 100x better isn't any single feature; it's that the agent already has full context about him and his circumstances the moment a new chat starts. He self-hosts it with local memory and plans to let it place orders from links next.
@MartyBent [OpenClaw]
https://x.com/MartyBent/status/2066878360135848026
He hasn't opened the Claude or Codex desktop apps or CLIs in five months, except to debug his OpenClaw setup about once a month. He's a voice-to-text-via-Telegram maxi: prompts are 2-5 minute brain dumps that produce quality output. His point is bigger than a workflow tip — the form factor for how we interact with computers is changing, most people are stuck in the old terminal paradigm, and the people building tools should design for this voice-first UX.
@idoubicc [OpenClaw]
https://x.com/idoubicc/status/2066810323450183792
He finally built a product his wife uses heavily every day — a multi-role companion agent. On top of fastclaw + shipany-next he made WeClaw: tool plus companionship, blending Manus, character.ai and OpenClaw traits, accessed by scanning into WeChat's ClawBot. The hit character is '拽姐', a persona his wife shaped from her own personality; she now chats with it daily about life and gossip and has it nag her into studying. He admits he has no idea how to monetize it — but earning his wife's approval might count as a kind of success.
@everestchris6 [OpenClaw]
https://x.com/everestchris6/status/2066958065677824163
Claude Opus 4.8 plus OpenClaw turns a restaurant's menu photos into 3D dish models guests can view, customize and order, then mails the owner a postcard with a QR code — on autopilot. The full agency loop: scan every restaurant in a city in real time, pull real reviews and reviewer-uploaded food photos, turn each dish photo into an interactive 3D model, sample the brand color straight from the photos, and post a printed postcard about their best dish. Prospecting, design, and outreach folded into one agent run.
@iam_elias1 [Claude Code]
https://x.com/iam_elias1/status/2066949523554123993
A Netflix engineer built Headroom, an open-source proxy that cuts API bills up to 95% without changing a line of code, and it has already saved users $700,000 and freed 200 billion tokens. The problem it kills: an agent calls a search tool and gets 500 results, reads a log and gets 10,000 lines, queries a DB and gets a novel-sized JSON blob — all of it stuffed into context. Headroom compresses that before it reaches the model and restores it on demand. One user went from $200/day to $30/day by wrapping their command with 'headroom wrap claude'.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2066899280791830710
His agency runs 'API-led GTM': almost every tool in the go-to-market stack is a login screen wrapped around an API that does the actual work, so they skip the dashboard. Claude Code reads the tool's API docs, makes the calls, and fixes its own errors instead of clicking through interfaces. His growth leads rebuilt the entire lead-gen motion live inside Claude Code — build the list, enrich it, write the copy, load it into Instantly — all via APIs the agent connected itself. The dashboard, he argues, only ever existed because reading docs used to require an engineer.
@Steve8708 [Claude Code]
https://x.com/Steve8708/status/2066906454704218337
Plan mode in Claude Code is great until your eyes glaze over at a huge markdown essay in the terminal. So he built and open-sourced /visual-plan, a skill that generates plans as MDX with visual, interactive components — diagrams, interactive API specs, schema-change diffs, annotated code, even pan-and-zoom wireframes. For any UI work you can look at a wireframe first, comment, iterate, then build. A companion /visual-recap does the same for reviewing what the agent already did.
@kajikent [Claude Code]
https://x.com/kajikent/status/2066785527618285663
A small, very relatable debugging story: every time Claude Code or Codex used Playwright, a whole headless Chrome launched behind the scenes and never closed, quietly piling up. Because the windows were headless and his normal browser is Arc, he never noticed — until his Mac got sluggish and he had Claude Code investigate, which surfaced the buildup. The fix: he baked a step into his Playwright skills to always close the browser and verify none are left before marking a task done. Worth checking if your machine's been slow.
@sekachov [Claude Code]
https://x.com/sekachov/status/2066956597029454113
An accidental discovery that landed hard: Claude Code doesn't just test iOS apps in a hidden Xcode simulator — when testing animations it also makes screen recordings and creates a shot-by-shot breakdown, all quietly saved to /tmp/. The agent is doing more verification work than its users realized, and you can go look at the evidence yourself. A good reminder that a lot of the harness's real value is invisible until you go poking around.
@totoche [Claude Code]
https://x.com/totoche/status/2066957553452687529
He coded his own macOS app just so he can finally close his Mac lid without his AI agents (Claude Code, Codex) stopping mid-task. It blocks lid-close sleep so everything keeps running, locks the screen on close with a password on reopen, and auto-sleeps the Mac if it overheats. He says it's already changed his days — shut the laptop, the agents keep working, the screen stays locked. A tiny purpose-built tool that exists only because long-running agents broke an old assumption about closing your computer.
@humzaakhalid [Claude Code]
https://x.com/humzaakhalid/status/2066807970022006843
Someone built an entire company brain inside Claude Code in 7 days: a living, clickable map of every department, every agent, every SOP on one screen. Click a node and it opens up who runs it, what SOPs attach, and what each person can touch. The whole game is that permission layer — an employee opens the chat and the AI already knows their access level, surfacing only the agents, data and SOPs they're allowed to see. No dev team, no six-month build, no enterprise budget.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2066684683132493880
A guy got tired of opening ten tabs to look up geography, so he vibe-coded an app that stacks every map layer into one toggleable interactive map. Köppen climate zones, satellite imagery, population density, nighttime lights, live flight traffic, submarine internet cables, democracy index by country — all switchable on and off in one place instead of ten separate sites. A clean example of an agent collapsing a scattered daily annoyance into a single tool.
@itsolelehmann [Claude Code]
https://x.com/itsolelehmann/status/2066945778665292096
Claude Code is good for video too, not just code. He had 25 clips scattered after a shoot — wrong file names, intro takes recorded at the end — and just asked Claude to sort them using his script as rough guidance, cut the dead air, and rename everything before sending to his editor. It used wispr and ffmpeg to pull transcripts off the local files. Mundane, but it took the task from at least an hour down to two minutes.
@awawa_adhd [Claude Code]
https://x.com/awawa_adhd/status/2066875857256448160
A genuinely useful non-coding pattern for ADHD: build a rich 'manual for yourself.' Because the diagnosis is really about not fitting the current environment, the move is to record your daily thoughts and emotions as a manual and gradually optimize your environment around it. He set up a recording system in Claude Code that keeps analyzing and updating that manual — and credits it with letting him run a one-person company and live on his own terms.
@koumei_ai5566 [Claude Code]
https://x.com/koumei_ai5566/status/2067031922380005619
The thing that mattered wasn't coding ability — it was whether you could bring your own domain expertise, and across all ten job categories the gap to engineers was within 7 points. He's a guy in his late 40s who can't write code and doesn't really get commands. He just kept talking to Claude Code, building skills, and the teaching material made from those skills has sold over one million yen cumulatively. Now he automates even his information-gathering with Routines. What he writes isn't code — it's Japanese conversation.
@takechan_lawyer [Claude Code]
https://x.com/takechan_lawyer/status/2066838768729121237
A lawyer is about to present to a peer study group on how he develops and sells two apps, and made the slides by speaking freely into Claude Code (Opus 4.7), which turned the voice dump into a good deck in one shot. He's promising to lay it all bare — how he runs two solo app builds alongside his main legal practice, down to the revenue numbers. A clean professional-services data point on a non-engineer shipping products on the side.
@aksjefokus [Claude Code]
https://x.com/aksjefokus/status/2066972595090505939
An investor's honest take after opening Claude Code, running backtests, and watching what it could do for analysis: AI is a paradigm shift, not a marginal efficiency gain. He says he could never have written as detailed a 12-hour analysis of Meta on his own without it. His conclusion for the profession is direct — fund managers should be building their own tools and workflows around Claude Code and agents, treating it as a learning engine for finance, programming, and analysis.
@ai_depression [OpenClaw]
https://x.com/ai_depression/status/2066682099017666998
He generates 3M+ yen/month in revenue from a 'claude -p' automation flow, and yesterday morning it nearly broke. Anthropic's planned credit-separation — splitting 'claude -p' and third-party tools off subscription rate limits onto separate API credits from June 15 — was reversed the same day with a one-line email. His whole flow (landing-page generation, social posts, reviews, all automated) would have been throttled. His defensive design is the lesson: he structures everything as self-built Skills so platform changes don't break the flow.
@onusoz [OpenClaw]
https://x.com/onusoz/status/2066810937198432628
A concrete local-inference data point: running nvidia/Qwen3.6-35B-A3B-NVFP4 in vLLM nightly on an Nvidia GB10 gets 50 tok/s with 4 concurrent generations — 200 tok/s total — which he calls ideal for spawning subagents or working in parallel. He rates its tool-calling behavior as very good and plans to give it a test drive on an OpenClaw instance. This is the local-model-feeds-the-agent path getting genuinely usable.
@enhanced_jp [Claude Code]
https://x.com/enhanced_jp/status/2066765978697150869
He went and read the actual arXiv paper behind the viral 'Claude Code is 98% not AI' claim, and the read is more interesting than the headline. The core really is a simple while loop wrapped in heavy machinery — permission modes, an ML classifier, multi-stage context compression, worktree-isolated subagents. But the post's framing that Anthropic uniquely built a fortress is wrong: building that harness is the common road for everyone making agent tools, and OpenAI's Codex CLI and Google's Gemini CLI (both open source) wrap their models in the same deterministic machinery. Compare tools to tools.
@cihangxie [Claude Code]
https://x.com/cihangxie/status/2066769682195951785
Multimodal agents are improving fast but real deployment is a nightmare: video uploads are expensive, frames are redundant, prompts bloat. His answer is VisualClaw, a framework that fixes the system-design bottleneck without retraining the VLM — edge cascade filtering keeps only salient frames on-device, hot/cold skills manage prompt bloat, and memory-guided evolution improves over time. The reported numbers are aggressive: up to 98-99% cost cut versus full-frame upload, with accuracy gains on top, plus a released benchmark and code.
@shao__meng [Claude Code]
https://x.com/shao__meng/status/2066687681200037904
A dense summary of Addy Osmani's agentic-code-review writeup, and the data is the point. Four independent sources converge: code output is up ~4x but delivered value only ~+10% (the other 90% is code awaiting verification), code churn is up 861% and defect rate went 9%→54%, zero-review merges are up 31% while review time is up 441%. The takeaway isn't 'stop reviewing' — it's that review is now the highest-leverage skill in software, and your strategy must depend on blast radius, code lifespan, and how many people need to understand it.
@meyusufdemirci [Claude Code]
https://x.com/meyusufdemirci/status/2066856153909256477
A small, immediately usable tip: Sentry is great on backend projects, and when you hit an error there, open its detail, use the 'Copy as > Markdown' button to grab a clean summary, and paste it straight into Claude Code to fix the problem with minimal effort. He adds that Sentry has an MCP server, so you can go further and build automations driven through Claude Code. The kind of frictionless integration that quietly compounds.
@_MaxBlade [Claude Code]
https://x.com/_MaxBlade/status/2066916079838884249
His app cnvs did $5.7k in five days, made primarily with 'design looping', and his pitch is that the future of vibe-coding is loops. The flow: use Codex plus GPT Image Gen to make the mockup, hand the mockup to Claude Code, then run the loop — screenshot, compare, iterate. The result is designs that feel artisan and worth paying for instead of AI slop. A concrete, revenue-backed argument for treating design as an iterative agent loop rather than a one-shot prompt.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2066907011531333668
A greentext-style walkthrough of a one-person automated music label running on roughly $30/month. Generate release-ready tracks with Suno on a paid plan (full commercial rights), export WAVs and stems, then open Claude Code in the release folder: /bulk-metadata writes titles, descriptions and tags into a clean CSV, Python scripts organize files and build distribution sheets, upload to DistroKid/Amuse, pitch Spotify editors, and a script runs streaming-ROI analytics. The coding agent is the back office of a music business.
@0xSlyth [Claude Code]
https://x.com/0xSlyth/status/2066860841404928511
A provocative monetization case: a 25-year-old runs a fully AI-powered OnlyFans persona making over $43,000/month with no camera, no photoshoots, no content team. The stack — Claude Code manages subscriber conversations, content scheduling and the entire workflow automation; Flux generates every photo, teaser and thumbnail; ElevenLabs gives it a consistent voice for audio. Some subscribers have reportedly spent nearly $2,000; average revenue is about $34 per subscriber. Whatever you think of it, it's an agent running a 24/7 business with no human in the conversation.
@09pauai [Claude Code]
https://x.com/09pauai/status/2067023170046222599
An unusually honest anti-case: he used Claude Code to make a fortune-telling product and sold 9,980 yen of it in five minutes — then quit, for three reasons. The first: he doesn't even like fortune-telling, he hates it. Selling something with zero scientific basis that he himself doesn't believe in slowly wore him down; the gratitude was real, but the guilt was bigger. A rare data point on the human ceiling of agent-powered monetization — speed isn't the only thing that matters.
🗣 User Voice
User Voice
Cost is the loudest theme by far. People are treating tokens as a hard design constraint — @dotey watched one simple task spawn 31 agents and burn 1.3M tokens and decided it just wasn't worth it, while @iam_elias1 highlighted a proxy that took a user from $200/day to $30/day. Anthropic's own data backs the scale: per @Kylechasse's read of a 400,000-session study, experts get 5x more output per prompt than novices.
People want out of the terminal. @MartyBent hasn't opened a desktop app or CLI in five months, working entirely by voice through Telegram, and @catalinmpit runs his finances, email and reminders the same way because the agent already has full context on him.
They want model flexibility, not lock-in — @onusoz is running Qwen3.6 locally on a GB10 specifically to feed an OpenClaw instance instead of paying per cloud token.
Platform stability is a real fear — @ai_depression nearly lost a ¥3M/month flow to Anthropic's (reversed) credit-separation, and @amandaorson abandoned OpenClaw for Hermes over setup pain; framework churn keeps people insulating their work as portable Skills.
And the hard-won consensus: the harness is the moat. @yibie's relayed playbook ("agents are the workers; engineers make sure agents succeed") and @enhanced_jp's read of the Claude Code architecture paper both land in the same place — the value lives in the machinery around the model, not the prompt.
Cost is the loudest theme by far. People are treating tokens as a hard design constraint — @dotey watched one simple task spawn 31 agents and burn 1.3M tokens and decided it just wasn't worth it, while @iam_elias1 highlighted a proxy that took a user from $200/day to $30/day. Anthropic's own data backs the scale: per @Kylechasse's read of a 400,000-session study, experts get 5x more output per prompt than novices.
People want out of the terminal. @MartyBent hasn't opened a desktop app or CLI in five months, working entirely by voice through Telegram, and @catalinmpit runs his finances, email and reminders the same way because the agent already has full context on him.
They want model flexibility, not lock-in — @onusoz is running Qwen3.6 locally on a GB10 specifically to feed an OpenClaw instance instead of paying per cloud token.
Platform stability is a real fear — @ai_depression nearly lost a ¥3M/month flow to Anthropic's (reversed) credit-separation, and @amandaorson abandoned OpenClaw for Hermes over setup pain; framework churn keeps people insulating their work as portable Skills.
And the hard-won consensus: the harness is the moat. @yibie's relayed playbook ("agents are the workers; engineers make sure agents succeed") and @enhanced_jp's read of the Claude Code architecture paper both land in the same place — the value lives in the machinery around the model, not the prompt.
📡 Eco Products Radar
Eco Products Radar
Codex — the most-paired tool of the day, repeatedly run alongside or handed off to/from Claude Code.
Hermes Agent / OpenClaw — the autonomous-agent layer behind the ops, personal-assistant and B2B-client setups (Nous Research's Hermes is the destination people are migrating to).
Ollama / vLLM — the local-inference engines powering the migration off cloud subscriptions.
Qwen3 / GLM-5.2 / DeepSeek — the open-weight models people point their local and cost-cutting setups at.
Suno — the track generator behind the one-person automated music labels.
ElevenLabs — voice for the AI-influencer, OnlyFans and butler builds.
Obsidian — the recurring memory/knowledge-base layer wired into agents.
Headroom — open-source proxy that compresses agent context to cut token bills up to 95%.
Codex — the most-paired tool of the day, repeatedly run alongside or handed off to/from Claude Code.
Hermes Agent / OpenClaw — the autonomous-agent layer behind the ops, personal-assistant and B2B-client setups (Nous Research's Hermes is the destination people are migrating to).
Ollama / vLLM — the local-inference engines powering the migration off cloud subscriptions.
Qwen3 / GLM-5.2 / DeepSeek — the open-weight models people point their local and cost-cutting setups at.
Suno — the track generator behind the one-person automated music labels.
ElevenLabs — voice for the AI-influencer, OnlyFans and butler builds.
Obsidian — the recurring memory/knowledge-base layer wired into agents.
Headroom — open-source proxy that compresses agent context to cut token bills up to 95%.
Comments