Super User Daily: 2026-06-09
The center of gravity moved off the keyboard today. The loudest stories were not about shipping faster code, they were about people with zero CS background pointing Claude Code at a problem and walking away with a business. A factory engineer wired it into his line and tripled his pay. A 70-year-old manufacturer built his own quality-control software instead of paying 150k for it. A humanities researcher rebuilt a months-long Excel macro in half an hour. The pattern underneath all of it is the same: the bottleneck is no longer writing code, it is knowing exactly what you want and being able to check the output. Here are the most concrete uses people shared.
@sspriint [Claude Code]
https://x.com/sspriint/status/2063609455602810902
A factory process engineer wired Claude Code straight into his task queue to read specs, write the control logic, and catch errors before defects reached the line. Output went up 30 percent with near-zero defects, which got him a tripled salary and a promotion. He then hired two teenagers at 200 dollars a month just to do the prompting, and is now aiming to automate the entire floor. This is the clearest 'AI as leverage on a real job' story of the day.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2063645440982110211
A 23-year-old non-CS gamer vibe-coded a mobile app for finding class-action payouts in 14 days, feeding Claude Code a JSON data structure and competitor screenshots, shipping cross-platform on Expo, Next.js and Supabase. It hit 12,000 downloads in 50 days and now makes 20,000 dollars a month through RevenueCat subscriptions behind a paywall, and won a hackathon against 55,000 entrants. No prior development background.
@dotey [Claude Code]
https://x.com/dotey/status/2063674134903603302
A detailed, reproducible workflow: use Claude Design to spit out HTML+CSS+React+data.js design specs (more AI-friendly than Figma and git-versionable), hand them to Opus 4.8 to build a macOS MVP incrementally while explicitly forcing AppKit over SwiftUI, then iterate by feeding back screenshots and the git diff of design changes. He even had Opus do a performance pass that flagged an NSScrollView memory blowup and recommended NSTableView, then reverse-engineered Claude Design's front-end to run it locally as a Skill inside Cursor.
@dali51334388 [Claude Code]
https://x.com/dali51334388/status/2063467613947363692
A breakdown of Matt Van Horn's no-IDE workflow: everything is driven by a plan.md and voice input, using Every's Compound Engineering plugin where one plan command dispatches parallel agents to read the codebase, review past mistakes, and fetch best practices into a plan.md with an acceptance checklist that doubles as a resumable save point. He has 70 plans and 263 commits in 30 days, runs 4-6 terminals at once, and once turned a recorded lunch transcript into a full proposal that landed a hire.
@DeRonin_ [Claude Code]
https://x.com/DeRonin_/status/2063573847220052325
Runs a one-person automated marketing agency on Garry Tan's open-source gstack repo (23 skills) inside Claude Code. /office-hours scopes a client launch in five minutes, /design-shotgun ships six landing-page variants per session, /qa catches broken automation before the client sees it, and a GBrain layer persists each client's voice and history across sessions. Work that used to eat his whole morning now finishes in an hour.
@rkwy [Claude Code]
https://x.com/rkwy/status/2063415631471653117
A humanities co-researcher had outsourced an Excel macro that took months and still didn't work. The author took the source material and a few correct input/output examples and rebuilt it with Claude Code in under 30 minutes as one self-contained HTML+JS file: drag in a CSV, get instant calculation and visualization. Then added image/data/report export, color and axis controls, and a tutorial video, all in about two hours.
@seylorra [Claude Code]
https://x.com/seylorra/status/2063668142904365377
A solo creative agency where Claude Code does the thinking, Obsidian holds the memory, and an MCP server links them so the agent wakes up each morning already knowing every client, brief, and brand color. He used it to produce a streetwear brand ad campaign overnight, and runs eight retainer clients at 9,400 dollars each, alone. The memory-plus-MCP setup is what turns a coding agent into a colleague that doesn't reset every session.
@OwenGregorian [Claude Code]
https://x.com/OwenGregorian/status/2063605807514526047
The Salesforce case making the rounds: a migration scoped at 231 days shipped in 13 with Claude Code, an 18x speedup with fewer incidents. The trick wasn't raw model power, it was migrating 33 API endpoints through a rule-based framework with reference implementations and autonomous build-fix-validate loops, while engineers scoped the tasks, wrote the rules, and reviewed the output. The humans set the boundaries, the agent ran the repetitive cycle.
@razib_ul47671 [Claude Code]
https://x.com/razib_ul47671/status/2063616443896459517
A Polymarket trading and intelligence system built with Claude Code, open-source tools, and custom workflows for under 30 dollars a month. One workflow mines millions of trades to surface consistently profitable wallets; another monitors live markets, filtering by liquidity, volume spikes, and whale activity; and three AI agents (arbitrage, probability convergence, trader-following) vote on entries and study how winners exit. A full quant desk assembled by a non-institution for pocket change.
@sandy4kad [Claude Code]
https://x.com/sandy4kad/status/2063693053982838850
A Texas trading analyst with no dev background spent two days conversationally directing Claude Code to build a content-tracking tool: every YouTube and Instagram video he posts becomes searchable in one click, with auto-generated transcripts, a hook score per video, a built-in script writer and research column, sortable by timeframe, platform, and performance. He runs it for about 17 dollars a month versus a comparable 200-dollar SaaS.
@zaimiri [Claude Code]
https://x.com/zaimiri/status/2063556258255556951
Priscilla Tina, a 28-year-old SF product manager, built Postcard Press in four hours with Claude Code: upload a photo, type a message, pay about two dollars, and Postgrid prints and mails the postcard. She shared the actual unit economics (Postgrid 0.82, Stripe 0.30) and has sent 100-plus postcards since launch. The value here is how cheaply a non-engineer can now stand up a real physical-world product with margins.
@iam_elias1 [Claude Code]
https://x.com/iam_elias1/status/2063503368459346143
A walkthrough of SynthTeam, a free MIT-licensed Claude Code plugin that distills colleague personas from your Slack history (stored locally) so you can pressure-test plans against them. Three skills: distill-slack-persona builds a five-facet persona doc, ask-colleague gives one persona's likely pushback, and ask-team convenes a simulated multi-round panel. It's a clever non-coding use: turning an agent into a stand-in review committee.
@maxwellfinn [Claude Code]
https://x.com/maxwellfinn/status/2063440780077732087
A marketer makes text-heavy ads directly in Claude Code, skipping image models like Nano Banana entirely: Claude designs the ad in HTML/CSS like a landing page, then saves it to PNG via headless Chrome. He lays out a whole taxonomy of ad types this works for, fake-organic native screenshots, receipts and documents, data and proof cards, typographic, analog/handmade, UI mockups. For anything text-driven, a code agent beats an image model on precision.
@narukun_ai [Claude Code]
https://x.com/narukun_ai/status/2063727810418794620
A concrete review-article assembly line in Claude Code: feed in a product name and angle, and a script generates three reader-persona analyses, an article structure, a 3,000-character body with story structure, five titles, three opening hooks, and an SEO meta description, then posts a WordPress draft. He says review articles dropped from three hours each to twenty minutes. This is content ops turned into a button.
@narukun_ai [Claude Code]
https://x.com/narukun_ai/status/2063727841821532430
Two more automations from the same builder: an X-growth script that analyzes 1,000 past viral posts to extract 20 syntax patterns and drafts 10 article-linked posts every morning at 7am into Slack (cutting X ops from three hours to ten minutes a day), and a LINE step-delivery customer-service script that sends day-by-day messages and runs daily A/B tests on click, read, and response rates logged to a spreadsheet. Both treat marketing as a measurable loop, not a vibe.
@libukai [Claude Code]
https://x.com/libukai/status/2063557771765940416
Rebuilt his company's LeapAll PPT website using Codex's new Product Design plugin, which folds 11 skills into one and replaced his previous Pencil + Shadcn Studio + Claude Code stack. He singles out the annotate-on-the-webpage feature for making precise edits in natural language plus CSS, replacing his whole design-to-code pipeline. A useful data point on how fast the design-to-code tooling is consolidating.
@DailyDoseOfDS_ [Claude Code]
https://x.com/DailyDoseOfDS_/status/2063551356863214048
A clean before/after on token cost: adding Insforge Skills plus a CLI as a backend context-engineering layer cut Claude Code's token use 2.5x on the same task, from 15.7M tokens at 12.95 dollars down to 6.3M at 4.87. As more people run agents in long loops, the real skill is shaping context so you pay for intelligence, not for re-reading the same files.
@IsraeliHadash [Claude Code]
https://x.com/IsraeliHadash/status/2063607856834732514
After weeks sketching schemas on paper and a week reading up on Claude Design, this user paid 20 dollars for Claude Pro and produced a clickable two-style design for a 10-screen mobile app in 25 minutes, then iterated four times, exporting to Flutter via Claude Code and fixing self-found errors. Final clickable prototype done in about 50 minutes for 20 dollars, with the actual build still ahead. A grounded look at what the design phase actually costs now.
@jchencxh [Claude Code]
https://x.com/jchencxh/status/2063740904658313463
Introduced Claude Code to a manufacturing owner in his 70s, who then vibe-coded a quality-control software upgrade himself instead of paying the 150,000 dollars he'd been quoted for it. The reason this matters: it's hard evidence the tool is diffusing past the developer bubble into non-technical, non-coding hands where the savings are largest.
@seinu_note [Claude Code]
https://x.com/seinu_note/status/2063441022034846000
A working content flywheel as a business: he uses Claude Code to scrape information off YouTube, repackages it into note (a Japanese publishing platform) posts sold on social media for around 200,000 yen a month, then has Claude Code auto-gather fresh material and propose five new product themes, keeping one. Scaled across multiple accounts it runs near a million yen a month. The agent isn't just building, it's sourcing the next product.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2063467264817479775
A real-life Pokedex app called Gotcha, vibe-coded with Claude Code: point your phone at any animal, it identifies and logs it to a personal index, with geo-based rarity (a rabbit is common on a farm, legendary in a city), collection profiles, achievements for rare species, and trading and battles between players. The kind of playful consumer product that used to need a small studio.
@takapon_jp [Claude Code]
https://x.com/takapon_jp/status/2063510849600516517
Vibe-coded an app on an old MacBook with Claude Code while walking on a treadmill watching Netflix, just issuing instructions and reviewing the output instead of writing code. His advice is the interesting part: don't start with a big service, start by updating the small everyday tools that annoy you personally. A year in, he notes how dramatically the agent itself has improved.
@levelsio [Claude Code]
https://x.com/levelsio/status/2063598613008101585
Over a year, with Claude Code's help, levelsio built a working web-based dot-matrix printer that prints from an emulated Windows 3.11 in the browser. LPT1 wouldn't cooperate, so he captured the serial COM2 data over WebSocket by reusing his old dial-up-modem COM1 capture trick, then rendered the raw print stream. A reminder that the agent is also a patient retro-systems debugging partner.
@xMigma [Claude Code]
https://x.com/xMigma/status/2063681879274873325
An honest field report on the limits: Codex and Claude Code are great on small and mid projects (days of work in a few prompts), but on large codebases the agent is too scared to delete code, so it keeps stacking wrappers and redundant 'just in case' functions until the codebase is unmanageable. His advice: don't let it dump 1,000-plus lines in one prompt on anything serious. Worth keeping next to all the success stories.
@0xDeliriumm [Claude Code]
https://x.com/0xDeliriumm/status/2063578696133296451
An 18-year-old Oxford student built an Obsidian knowledge graph with thousands of nodes and wired a free repo of 23 skills (product thinking, architecture, QA, docs) into Claude Code so a single prompt runs a full plan-build-test-review-ship pipeline. He reports earning 14,000 dollars a month shipping products for clients. Same pattern recurring all day: a personal knowledge base plus a skill stack is the real moat.
@milindlabs [Claude Code]
https://x.com/milindlabs/status/2063760274734350731
Built Bluey, a local-first tool that lives under your cursor and lets you draw on screen, speak, or type to turn screenshots, annotations, transcripts, and the active app/window into rich context for Claude Code or Codex. He made it to design websites and debug UI by pointing at the actual screen instead of writing long explanations. A neat example of building the missing context layer around the agent.
@dunik_7 [Claude Code]
https://x.com/dunik_7/status/2063632615727350149
Built a browser game in a weekend using Claude Code plus the Claude-Code-Game-Studios repo (49 agents, 73 skills, 12 hooks), with clearly split roles, Game Director, Design Lead, an Engineering Lead on Phaser/React/Next.js, Art Lead, QA, and Release to Vercel/Stripe. He even lays out the economics: 9.99 a month, infra under 50, 95-percent margin via streak mechanics. A full studio org chart, run by one person.
@ern_exe [Claude Code]
https://x.com/ern_exe/status/2063539405059396031
Shared /last30days, an open-source research agent you install into Claude Code, Codex, or Cursor that simultaneously scans Reddit, X, YouTube, Hacker News, GitHub, TikTok, and Polymarket, scoring results by real engagement rather than SEO. It's positioned as a trend radar you run with your own API keys, a sign people increasingly want their coding agent doing market research too.
@BennyKokMusic [Claude Code]
https://x.com/BennyKokMusic/status/2063635194179961219
A concrete recipe to keep agents running 24/7: a 40-dollar Hetzner VPS, Claude Code on the Max plan running in tmux in the background, Tailscale for secure access, and a custom vibecoded UI to reach the session from anywhere. Nothing exotic, which is the point, the always-on agent setup is now a weekend's worth of off-the-shelf parts.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2063508353792561196
A document-production technique: build only the skeleton in Claude Code, then drive Manus via browser automation to format it into the final deliverable, which he finds more accurate and stable, and cheaper on tokens, than HTML/CSS, direct PowerPoint, or image-model approaches. His broader point: browser-driving the API-less tools is itself a way to level up what Claude Code can do.
@orevbajohn_ [Claude Code]
https://x.com/orevbajohn_/status/2063713579417612715
Started building 3D prototypes in Blender using Claude Code, currently modeling a mascot for a project. Small but notable: it's another non-coding domain (3D content) where people are now driving the agent through a creative tool rather than a codebase.
@0xfinkus [Claude Code]
https://x.com/0xfinkus/status/2063580071344955393
Reports Jake Morris building a roughly 200-Mac-Mini AI cluster in his spare room instead of paying 1,400 dollars a month for an H100 server, with Ollama plugging directly into Claude Code for local inference. The cost math driving people off cloud inference is becoming a recurring storyline, and the rigs are getting serious.
@matteocollina [OpenClaw]
https://x.com/matteocollina/status/2063649122163269778
Running OpenClaw locally on top of antirez's DwarfStar, which he says solves his main grievance, data sovereignty, and is a good use of his GB10. He's also using DwarfStar as a model for a local Pi setup. A clean, real OpenClaw use case amid a lot of noise: the appeal is keeping the whole agent and its data on your own hardware.
@koromoops [OpenClaw]
https://x.com/koromoops/status/2063507775020544460
A thoughtful note on actually living with OpenClaw: the hard part isn't the number of automations, it's deciding at what granularity to hand off daily tasks. He finds that slicing work into units you can check and approve from your phone, rather than just piling on more automation, is what keeps the agent embedded in daily life. The keep-a-human-in-the-approve-loop instinct, applied to non-coding chores.
🗣 User Voice
User Voice
A few wishes and frictions came up again and again under today's posts.
Make the loop affordable. The cost of long-running agents is the number one complaint, and people are voting with their wallets, moving to local models, Mac Mini clusters, and free gateways to escape subscription and API bills. @0xfinkus and the Mac-Mini-cluster crowd want the same intelligence without the 1,400-a-month tab.
Stop paying to re-read the same files. Token and context bloat is the silent cost. @DailyDoseOfDS_ showed a 2.5x token cut just from a better context layer, and @xMigma warns that on big codebases the agent burns budget piling on redundant code it's too scared to delete.
Give agents real memory. The thing that turns a coding tool into a colleague is persistent, shared memory across sessions. @seylorra wires Obsidian plus MCP so the agent wakes up knowing every client; @0xDeliriumm leans on a thousand-node knowledge graph. Everyone is rebuilding the same memory layer by hand.
Let me drive it from anywhere, with fewer interruptions. People want mobile and voice access and fewer permission confirmations once a plan is approved, so the agent can actually run unattended.
A few wishes and frictions came up again and again under today's posts.
Make the loop affordable. The cost of long-running agents is the number one complaint, and people are voting with their wallets, moving to local models, Mac Mini clusters, and free gateways to escape subscription and API bills. @0xfinkus and the Mac-Mini-cluster crowd want the same intelligence without the 1,400-a-month tab.
Stop paying to re-read the same files. Token and context bloat is the silent cost. @DailyDoseOfDS_ showed a 2.5x token cut just from a better context layer, and @xMigma warns that on big codebases the agent burns budget piling on redundant code it's too scared to delete.
Give agents real memory. The thing that turns a coding tool into a colleague is persistent, shared memory across sessions. @seylorra wires Obsidian plus MCP so the agent wakes up knowing every client; @0xDeliriumm leans on a thousand-node knowledge graph. Everyone is rebuilding the same memory layer by hand.
Let me drive it from anywhere, with fewer interruptions. People want mobile and voice access and fewer permission confirmations once a plan is approved, so the agent can actually run unattended.
📡 Eco Products Radar
Eco Products Radar
Tools mentioned three or more times across today's posts.
Codex - the other coding agent people pair with Claude Code, often in a build-then-review loop.
Cursor - still the default editor people install skills and agents into.
Obsidian - the recurring choice for the agent's persistent memory and knowledge graph.
MCP - the connective tissue linking agents to tools, files, and that external memory.
Ollama - the local-inference runtime powering the Mac Mini and RTX cost-cutting rigs.
gstack - Garry Tan's open-source skill pack showing up in multiple solo-agency setups.
Claude Design - increasingly the front-end of the design-to-code workflow.
Supabase / Vercel - the default ship-it backend and host for these vibe-coded apps.
Tools mentioned three or more times across today's posts.
Codex - the other coding agent people pair with Claude Code, often in a build-then-review loop.
Cursor - still the default editor people install skills and agents into.
Obsidian - the recurring choice for the agent's persistent memory and knowledge graph.
MCP - the connective tissue linking agents to tools, files, and that external memory.
Ollama - the local-inference runtime powering the Mac Mini and RTX cost-cutting rigs.
gstack - Garry Tan's open-source skill pack showing up in multiple solo-agency setups.
Claude Design - increasingly the front-end of the design-to-code workflow.
Supabase / Vercel - the default ship-it backend and host for these vibe-coded apps.
Comments