Super User Daily: 2026-06-07
The theme today is not the agent, it is the person driving it. The strongest cases all came from someone who knew exactly what they wanted and used Claude Code or OpenClaw as a production engine, not a novelty: a 26-year-old running a profitable social network alone, a laid-off engineer building a job-hunting machine the day he was let go, a biotech lab turning weeks of waiting into live analysis. Underneath runs the same nervous current as every week, the token bill, with honest hardware math and weekly-limit war stories sitting right next to people who clearly think spending more is the entire point. And one quiet line from a non-technical user said it best: the bottleneck was never the AI's ability, it was whether the human can drive it.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2062925871426556015
A 26-year-old quit her job and built a small social network just for NYC cafe obsessives, profiles, saved lists, feeds, the works, on Supabase and Vercel. The part that used to need five engineers, she did alone with Claude Code. Four months in it has 31,000 users and pulls $18,400 a month, cafes pay $199 to claim a page, power users $8 for private city guides. This is the one-person-company everyone keeps predicting, except it already shipped and it is cash-flow positive.
@nicos_ai [Claude Code]
https://x.com/nicos_ai/status/2062975808113623461
He analyzed 153 real tech job postings across Spain, the US and remote with a single Claude Code prompt plus an agent called Zero, no manual scraping, no spreadsheets. Zero picked the cheapest services itself, used BrightData to get past anti-bot walls, ran eight parallel Indeed searches, pulled 153 listings with salaries and stacks, then hosted interactive charts on a public URL and emailed the report. Total cost: 36 cents. The finding bites too, only 7% of postings are for juniors and 51% want seniors.
@vibeeval [Claude Code]
https://x.com/vibeeval/status/2062823073472975214
A man got laid off and, instead of panicking, built himself a job-hunting system with Claude Code that same energy. Paste a listing link and it grades the job A to F on whether it is even worth applying, auto-prepares an ATS-friendly CV, runs salary research and per-company interview prep, and logs everything to a tracking sheet, all from one command. It flat-out refuses to recommend anything below a grade 4 to protect his time. He scored 740-plus listings, landed a Head of Applied AI role, then open-sourced the whole thing.
@VaibhavSisinty [Claude Code]
https://x.com/VaibhavSisinty/status/2062724362730623446
A PhD geophysicist in Denmark open-sourced a job-application pipeline built entirely on Claude Code, made because he needed it himself. You fork it, fill in your background once, and for every job it reads the posting, scores your fit, drafts a CV picking only the experience that matches, and writes a cover letter. The tell is the second step: another AI agent reviews the first agent's draft, finds the weak spots, and the first one revises before it all compiles to clean PDFs. Generate-then-critique has quietly become the default shape of serious agent work.
@aalazy_ [Claude Code]
https://x.com/aalazy_/status/2062728821745525004
A government subsidy worth tens of millions of yen normally means handing a consultant a brutal success fee. Instead he wrote the whole application himself with Claude Code in about fifteen hours, and it went through for roughly 30 million yen. His takeaway is almost industrial: spin up a team and mass-produce subsidy applications, play it as a numbers game. The expensive professional service here was not knowledge, it was paperwork, and paperwork is exactly what folds first.
@glebkuz [Claude Code]
https://x.com/glebkuz/status/2062942070437032128
A scientist at Manifold Bio described how Claude Code changed the lab overnight. They have run 525 next-gen sequencing runs producing over 70 terabases, the equivalent of ten thousand human genomes. The old bottleneck was brutal: scientists waited weeks for a computational person to free up and analyze the data. Now every scientist live vibe-analyzes their own sequencing data through an agentic interface on top of a data layer, and the queue simply disappeared. This is the unglamorous, high-stakes version of the story, the agent dissolving a coordination bottleneck inside real drug discovery.
@0xMoysei [Claude Code]
https://x.com/0xMoysei/status/2062889251310645536
He reverse-engineered a faceless Facebook video page that pulls $26,457 a month (the owner posted the payout screenshot) and the entire production is four Claude Code prompts in one folder. Prompt one researches a niche and spits out ten scripts; prompt two turns each into fourteen detailed scene prompts; prompt three fires the Higgsfield skill on all fourteen at once in three minutes; prompt four animates the stills into four-second clips. The voice is an eleven-second ElevenLabs clone and DaVinci Resolve is wired to Claude to drop clips on the timeline and render. No camera, no editor, no face.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2062848159215493176
The creator of Obsidian shipped an agent skills system that teaches Claude Code, Codex and OpenCode to read, write and reason inside an Obsidian vault like a power user, and it hit 27,000 GitHub stars in days. It is not a chat-on-top-of-notes plugin: there is obsidian-markdown for the full flavor, obsidian-bases for views and formulas, json-canvas for visual graphs, obsidian-cli for terminal control, and defuddle to strip web pages to clean Markdown. The pattern worth noticing is that memory is becoming a first-class place the agent operates, not a vector blob it queries.
@this_is_tasnim [Claude Code]
https://x.com/this_is_tasnim/status/2062912335023620550
He is deleting Notion AI for an open-source tool called claude-obsidian that turns the vault into a notetaker that files, cross-references and maintains itself while you work. The architecture is the point: not a chat interface on your notes, but an engine that creates and evolves the notes themselves. It auto-organizes ingested content into entities and sources, flags contradictions across the vault with callouts linking back to the source, runs autonomous three-round web research and files the output, and one cross-project setup feeds all his Claude Code projects from a single vault. Plain-text memory keeps quietly winning over the vector database.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2062738282576752831
He recreated The Office where every character is a separate Claude Code instance running locally with its own personality, memory file and semantic search. Michael is the god agent who never implements, he fans out tasks, runs hourly standups, resolves conflicts, runs QA and handles all the GitHub changes, while Dwight and the rest do real work like writing blog posts and context engineering on a TODO/DOING/DONE board. The fun framing hides a real result: Michael's QA gate caught genuine bugs, including a dedup trap where a commit had been built from HTML with no source. It is a comedy bit and a working multi-agent org chart at the same time.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2062901723099005177
He makes the case that Dynamic Workflows are not a programmer toy but a scaffold for your life: Claude Code writes a bit of JS on the spot to spawn a swarm of isolated-context sub-agents over your backlog. Real runs: ten parallel agents scanned his last 50 sessions to mine what he keeps correcting (49 sessions, 86 corrections) into an HTML report plus CLAUDE.md diffs; 31 daily notes each got a Haiku to extract points then Opus clustered them to surface what he keeps procrastinating on, each pattern dated for evidence. The isolated context kills three classic failure modes at once, laziness, self-preferential scoring, and goal drift.
@ridark_eth [Claude Code]
https://x.com/ridark_eth/status/2062850687441002767
He laid out a one-person B2B content factory: Microsoft's free MarkItDown converts raw PDFs, webinars and Zoom recordings into clean Markdown (he claims a 70% token saving by not feeding Claude broken tables), then Claude Code plus n8n turns one 60-minute webinar into a long article, five Telegram posts and three X threads. The math he pitches is ten clients at $1,500 a month, $15,000 monthly, with a 60-second QA per output and n8n doing the fetch-convert-format loop. The interesting move is not the LLM, it is preprocessing the input cheaply so the expensive model reads less.
@doublenickk [Claude Code]
https://x.com/doublenickk/status/2062924542343836122
He turned Claude Code into a personal social media manager by connecting it to a TikTok account and having it analyze every post ever published, engagement, hashtags, transcripts, audience behavior, comment patterns. It came back with which hooks actually grab attention, which topics drive shares, and what the audience wants next. The whole thing cost under a dollar to run. The strategist work that agencies bill for is increasingly just a query against your own back catalog.
@Chain_GPT [Claude Code]
https://x.com/Chain_GPT/status/2062921637801107605
He built a Chrome-dino-style runner called Crypto Crash in one afternoon with the ChainGPT skill for Claude Code, and every system in it is plugged into live data. The ground you run on is BTC's actual 24-hour price chart, hills are pumps and valleys are dumps; the sky's color flips on a 0-to-100 market-emotion score the LLM derives from today's headlines; the obstacles carry real bearish news from a live news API. One prompt stitched the model, the news feed and the price feed into a playable game. It is a toy, but it is the clearest demo this week of an app that is alive rather than static.
@MalwareBibleJP [Claude Code]
https://x.com/MalwareBibleJP/status/2062825018623684855
A demo showed Claude Code, on the publicly available Claude Opus 4.6, doing firmware binary analysis and vulnerability discovery on a video intercom with no human in the loop, finishing in under ten minutes work that took a security team hours by hand. Given only the public firmware, a CLAUDE.md instruction file and a custom MCP driving the Ghidra decompiler, it unpacked the firmware, self-installed UPX when it needed it, located the command-injection code in about six and a half minutes, and wrote a disclosure-grade report re-finding five known high-severity bugs. The author's point lands hard: the decider was not a secret top model, it was the harness wrapped around a public one.
@arshadkazmi42 [Claude Code]
https://x.com/arshadkazmi42/status/2062897926418886858
He has been running strix-claude-code, an AI bug hunter built on Claude Code, on his server for months, prompting it from his phone through a mobile PWA he built called termi, and it has already found plenty of bugs. Then he closed the last loop: a browser AI agent that talks to strix through termi, where strix already has HackerOne and Intigriti MCPs so it pulls its own programs. He started it and just watched, the agent told strix to hunt, strix pulled a program, picked a target and began testing fully autonomously. This is the overnight-agent dream pointed at security research.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2063003770183987338
A Chinese developer spent three days testing local AI hardware for agentic use and surfaced the number nobody quotes: about $2,800 just to start. On the same model a Mac Mini 16G thought for two-plus minutes at 12 tokens/sec while a 4070Ti hit 65, both fully maxed. Worse, a 9B model plugged into Claude Code could not do tool calls at all, useless for anything agentic; he concluded you need 48GB and a 27-35B MLX model to even play. It is the most honest counterweight to all the buy-a-Mac-Mini-and-cancel-Anthropic threads.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2062761335625326988
He built a Claude Code plugin that maps a competitor's entire ad funnel, not just their ads. It pulls every active ad from the Meta Ad Library, ranks the winning landing pages by run-duration times ad volume, then dissects each page by hook, offer, social proof, pricing and urgency. It even adds items to cart to surface free-shipping thresholds, bundles, gift perks and the pre-checkout upsell structure. The manual version is an analyst screenshotting pages one by one; here it is a single pipeline one person can run.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2062802855539642789
A clean, copyable workflow to hand a brick-and-mortar shop a landing page in ten minutes with Claude Code plus two add-ons: install the UI/UX Pro Max skill from GitHub (50-plus UI styles, 97 color palettes, 57 font pairings) and add 21st.dev's Magic MCP for animations, glassmorphism and gradients that do not look AI-generated. Then you just paste the shop's Google Maps reviews and Claude produces a finished page with the name, real reviews, services and a booking CTA. Handing someone the finished thing beats discussing the plan.
@KenzieMac_Dev [Claude Code]
https://x.com/KenzieMac_Dev/status/2063002923915727023
He built a skill on Claude Code plus Browserbase that audits your credit cards and tells you which ones you are overpaying for, with the punchy demo that you should probably cancel your Amex Gold. It is a small thing, but it is exactly the kind of tedious personal-finance comparison nobody actually does, automated by an agent that can drive a browser. Expect a wave of these once people realize an agent with browser access is a free analyst for any account you already pay for.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2062967473033650180
His note on designing Auto Mode policies for Claude Code is genuinely useful: the primary input is not your infra or code patterns, it is your telemetry. OTEL shows how your team actually uses the agent, the prompts they write, the good and bad habits, how they behave with an agent in the loop, and that is the signal you hand Auto Mode. His advice is to resist shipping strict policies in one shot, start minimal, go back to the OTEL data, and tighten in steps until the rules match how the org really works.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2062736530347831487
A self-described non-engineer who cannot write a line of code now makes short-video captions with Claude Code and has stopped using Vrew entirely, a 90-second video's captions take about three minutes. The setup is thoughtful: local Whisper for transcription, Obsidian for context so it understands his past content, a proofreading and dictionary layer, and fcpxml export so captions drop straight in. The design gets smarter the more he uses it. This is the non-coder turning Claude Code into a bespoke replacement for a paid tool.
@abhishekn [Claude Code]
https://x.com/abhishekn/status/2062763778442412125
A small but underrated workflow: Claude Code or Codex plus Overleaf plus Dropbox to clean up all the misplaced LaTeX tables and broken citations that academics waste hours on. He calls it one of the best uses for a coding agent, and he is right, it is a high-friction, low-glory chore that the agent handles without complaint. Not every win is a startup, some are just the boring formatting tax finally being paid by a machine.
@shingo2000 [Claude Code]
https://x.com/shingo2000/status/2062900830114897961
From a Claude Code lead designer's talk: a custom /prototype skill that generates N different implementations of a feature, each with an HTML preview. The shift worth noting is in who chooses, the designer used to pick the prototype, but now they have Claude select the best one and explain its reasoning. Handing taste-adjacent judgment to the model, then auditing the explanation, is a subtle escalation in how much of the creative loop people are willing to delegate.
@zijing [Claude Code]
https://x.com/zijing/status/2062898286957363307
A non-technical user tried something he thought was simple, track three months of new BNB Chain projects into a brief, expected 30 minutes and burned hours. The CoinGecko API needed hand-written scripts, CoinMarketCal data was messy, old Dune queries would not run, and Claude's generated scripts errored halfway with no way for him to debug. He gave up half-done, and then wrote the most clear-eyed line of the day: the bottleneck was never the AI's ability, it was whether the human can drive it, knowing which model, how to prompt, how to chain tools, how to read an error log. Every new model release makes you pay that hidden cost again.
@mranonymous_011 [Claude Code]
https://x.com/mranonymous_011/status/2062731526060589114
Annoyed that a job-hunting site locked its interview reports behind a login, he just had Claude Code drive a browser to bulk-scrape them, and ended up with a comprehensive local stash of interview questions and advice across every company. It is a tiny act of friction removal, but it is the everyday version of the agent-as-leverage story, a paywall or a login becomes a five-minute scripting problem instead of a wall.
@tankxu [Claude Code]
https://x.com/tankxu/status/2063010371997491630
He built a remote control for a robot vacuum out of a cardboard box, four directional pads where touching tinfoil triggers a response, with an ESP32 handling the full command set over the local network and no gateway needed. He wrote the program with Claude Code and debugged it by reading serial-port logs, without reading a single line of code himself. This is the quiet hardware corner of the trend, an agent letting a tinkerer cross from software into physical electronics without learning to read the code in between.
@date_tokyo_uni [Claude Code]
https://x.com/date_tokyo_uni/status/2062839758095515888
His rhythm is startling: he spends only the first two days of the month using Claude Code to create every kind of content, then lives the other 28 days entirely on his phone, even doing the final read-through of an M&A contract on a vertical screen. He says life changed, and that he has graduated from using it himself to just instructing others to use it. It is an anecdote, but it captures a real shift, the heavy creation collapses into a couple of days and the rest becomes light supervision.
@robj3d3 [Claude Code]
https://x.com/robj3d3/status/2062843698942251318
Building in public at $20.5k MRR, he says the best thing he ever did was pair Claude the chat with Claude Code: Claude writes his prompts and approves the next steps, and Claude Code does the hard work in the repo. It is a small structural insight with real revenue behind it, splitting the planner from the executor across two surfaces of the same model so each does what it is good at.
@aehyok [OpenClaw]
https://x.com/aehyok/status/2062719381231329575
A hands-on look at Hermes Agent's Auxiliary Models feature inside the OpenClaw/Hermes world: each auxiliary task is independent and can be pinned to its own model, and there is a built-in fallback so if an auxiliary model fails or hits its quota (an HTTP 402, say) it auto-switches to a backup or the main model and the flow never breaks. Set it to auto and Hermes picks the model itself. He also just installed the new Hermes Desktop client. This is the unglamorous reliability plumbing that actually decides whether agents survive in production.
@hxiao [OpenClaw]
https://x.com/hxiao/status/2063044733644267662
He built the thing he always wanted, a fully local, airgapped multimodal file search for Mac using jina-v5-omni embeddings that indexes text, PDF, image, audio and video. It is a Swift-native UI on an mlx-swift-transformer core with no Python, tested clean on an M3 Pro 18G, an M3 Ultra 512G and an M4 Pro 48G with no out-of-memory. Crucially a local HTTP server exposes the index to agents like OpenClaw and Hermes, so your private files become a tool your agent can query without anything leaving the machine.
@sudoingX [OpenClaw]
https://x.com/sudoingX/status/2062993128920121521
His point is a useful corrective: for a lot of people agent has quietly come to mean OpenClaw, as if it is the only option, and it is not even close. His main orchestration is Hermes Agent, which he trusts with real coding work, and paired with Opus 4.8 on max effort he calls it the most capable agentic workflow he has run, scaling from a 3B local model to the frontier without going rogue. Judging the whole category by one heavy tool, he argues, hides how much else is out there.
@asaio87 [Claude Code]
https://x.com/asaio87/status/2063004248900608134
He built six complete SEO websites in 24 hours on a Claude Code Max subscription and almost maxed it out, noting SEO tooling, Ahrefs, MCP and a lot of prompts ate the budget fast. The sites look good and serve as a backlink source, with ranking still to be proven. It is a blunt data point on what the new heavy-usage ceiling actually buys, and a reminder that the limit is now a real constraint for people running content at volume.
@pastel_orbit [Claude Code]
https://x.com/pastel_orbit/status/2062925041835384959
A polished Unity game-dev loop: he screenshots the broken bit, draws a red circle on it as the instruction, and Claude extends the code without wrecking the original. For continuity across sessions he has Claude write task-specific handoff markdown files (auto-named, git-tracked) rather than relying on /resume. The standout is a Unity CLI Loop where Claude operates Unity directly from the CLI, writing a goal test and looping its own implementation until the test passes, which he kicks off and monitors remotely while out, coming home to finished work.
@niraj_munot [Claude Code]
https://x.com/niraj_munot/status/2062803166983606495
A tidy set of best practices for running a trading agent with Claude Code, Codex or Cursor, sharper on the CLI tools than the desktop apps. The non-obvious ones: from the web, never ask it to send you trades or data directly because it bloats context, ask for a downloadable URL instead; always reference days-to-expiry and days-from-last-expiry rather than calendar day names; and keep interrogating the agent until you fully understand what it did rather than trusting the output. Practical scar tissue from actually running agents on money.
@ersinkoc [Claude Code]
https://x.com/ersinkoc/status/2062803302362845476
A heavy user reports going hard for two straight days and still not exhausting the Claude Code 20x weekly limit, crediting it to intense work on a project he calls WrongStack. It is a single first-hand data point, but a useful one, against the chorus of limit complaints it suggests the weekly ceiling is high enough that even sustained heavy use does not necessarily hit it. Where you land on this depends entirely on how hard you actually push.
@mattwensing [Claude Code]
https://x.com/mattwensing/status/2062722747545878988
An honest little cost breakdown of building a small mobile app in 2026: Claude Code Max $100, Apple dev license $99, tokens $13.06, hosting 9 cents, and as he puts it, typing commands while watching the NBA finals, free. The tokens being the cheapest line item is the quiet headline, the model is no longer the expensive part of shipping software, the platform taxes and the subscription are. The economics of who can build just shifted again.
@robinebers [Claude Code]
https://x.com/robinebers/status/2062840304286449836
A pointed, useful complaint aimed straight at the Claude Code team: please fix the single biggest blocker stopping people from building apps with external login services like Clerk and AuthKit, the localhost limitation that means those apps cannot even start. He notes that by contrast Codex can set up Meta and Google Analytics, browse YouTube competitors and e2e test apps. He frames it as a true blocker flagged thousands of times, not a hopeful feature request, and he is voicing what a lot of builders quietly hit.
🗣 User Voice
User Voice
Cost is still the loudest note, but it has matured from panic into accounting. @noisyb0y1 's three-day local-hardware test landing on a $2,800 floor, and @ersinkoc burning two hard days without hitting the 20x weekly limit, are people doing the real math instead of just complaining.
The most quoted line of the day was not about a feature, it was about people. @zijing , a non-technical user, nailed it: the bottleneck was never the AI's ability, it was whether the human can drive it, knowing which model, how to prompt, how to chain tools, how to read an error log.
Reliability and platform limits are the recurring ask. @robinebers is begging the team to fix the localhost blocker that stops Clerk and AuthKit apps from even running, a real-blocker that has been flagged thousands of times.
Memory wants to be plaintext. Two of the day's strongest tools, the official Obsidian skills and claude-obsidian, both make a readable vault the agent's workspace rather than an opaque vector store, and users are switching off Notion AI to get it.
The frontier of desire is diversity and autonomy. @sudoingX 's pushback that agent should not just mean OpenClaw, plus a wave of overnight autonomous loops, says people want more options running longer with less hand-holding.
Cost is still the loudest note, but it has matured from panic into accounting. @noisyb0y1 's three-day local-hardware test landing on a $2,800 floor, and @ersinkoc burning two hard days without hitting the 20x weekly limit, are people doing the real math instead of just complaining.
The most quoted line of the day was not about a feature, it was about people. @zijing , a non-technical user, nailed it: the bottleneck was never the AI's ability, it was whether the human can drive it, knowing which model, how to prompt, how to chain tools, how to read an error log.
Reliability and platform limits are the recurring ask. @robinebers is begging the team to fix the localhost blocker that stops Clerk and AuthKit apps from even running, a real-blocker that has been flagged thousands of times.
Memory wants to be plaintext. Two of the day's strongest tools, the official Obsidian skills and claude-obsidian, both make a readable vault the agent's workspace rather than an opaque vector store, and users are switching off Notion AI to get it.
The frontier of desire is diversity and autonomy. @sudoingX 's pushback that agent should not just mean OpenClaw, plus a wave of overnight autonomous loops, says people want more options running longer with less hand-holding.
📡 Eco Products Radar
Eco Products Radar
Codex — the constant reference point, in almost every workflow as the second engine people pair with Claude Code rather than abandon it for.
Hermes Agent — Nous Research's local-first agent, the open alternative people keep reaching for; it anchors the OpenClaw-ecosystem cases on auxiliary-model fallback, local file search and orchestration.
Cursor — still the default IDE comparison point, named constantly alongside the CLI agents.
Obsidian — having a real moment as agent memory, both the official skills set and claude-obsidian turn the vault into the agent's readable workspace.
Dynamic Workflows — Anthropic's on-the-fly harness generator, treated as the biggest primitive since Skills and powering the multi-agent and life-scaffold runs.
Local stack (Ollama + Mac Mini + MLX) — the cost-rebellion hardware, cited constantly, though the honest benchmarks keep showing consumer machines are not there yet.
Gemini — the recurring second model people fall back to when Claude gets stuck.
n8n — the glue people wire Claude into for the fetch-convert-format automation loop behind content factories.
Codex — the constant reference point, in almost every workflow as the second engine people pair with Claude Code rather than abandon it for.
Hermes Agent — Nous Research's local-first agent, the open alternative people keep reaching for; it anchors the OpenClaw-ecosystem cases on auxiliary-model fallback, local file search and orchestration.
Cursor — still the default IDE comparison point, named constantly alongside the CLI agents.
Obsidian — having a real moment as agent memory, both the official skills set and claude-obsidian turn the vault into the agent's readable workspace.
Dynamic Workflows — Anthropic's on-the-fly harness generator, treated as the biggest primitive since Skills and powering the multi-agent and life-scaffold runs.
Local stack (Ollama + Mac Mini + MLX) — the cost-rebellion hardware, cited constantly, though the honest benchmarks keep showing consumer machines are not there yet.
Gemini — the recurring second model people fall back to when Claude gets stuck.
n8n — the glue people wire Claude into for the fetch-convert-format automation loop behind content factories.
Comments