Super User Daily: June 28, 2026
The center of gravity moved this week. The loudest Claude Code stories aren't about writing code anymore — they're about people pointing a terminal at jobs that used to need a whole firm. A solo developer running a seven-agent web agency off the phone in his pocket. A land surveyor walking a property with a LiDAR scanner, billing $300 an hour where a crew used to charge $8,000. A tax accountant quietly running 60 companies alone. The thread connecting all of them is the same: the model stopped being a chatbot and became a harness you can walk away from and leave running. The other half of today's feed is the plumbing that makes that possible — memory layers, self-verifying loops, token proxies — plus a sharp new strain of security paranoia as people realize an agent that can read your Sentry errors can also be told what to do by them. Here's what real users actually shipped.
@ParamSiddh [Claude Code]
https://x.com/ParamSiddh/status/2070538133507309906
The cleanest one-person-agency case of the year. A Chinese developer wired up seven Claude Code agents — Scout, Diagnoser, Builder, Filmer, Pitcher, Checker, and a Mobile agent that lives on his iPhone — to find local businesses with no website, build mockups, render a 10-second pitch video, and cold-message them across four channels. The whole thing runs 24/7 on a local sandbox, an MCP router, and shared state through the filesystem. No backend, no team. Real logs: 218 businesses scanned across Austin, Denver and Miami, 30 messages out, 5 positive replies, 3 Zooms booked. $480/month in API spend against $18,800 in revenue. The owner's only job is to hit approve and show up to the meeting.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2070428106805203014
Built an AI World Cup prediction terminal that hit a 91.7% win rate yesterday — 11 of 12 games right. The interesting part isn't the bet, it's the architecture: instead of paying a third-party API, he connected the terminal straight to a Polygon RPC to read every Polymarket trade off-chain, decoded raw logs into clean trade objects, then had Claude build filters for whale buys, repeated wallets and sentiment flips with dedup so the feed doesn't drown. Two weeks of trial and error, six core components, deployed always-on to Vercel. He pauses it when idle because it costs about $40/hour to run.
@browomo [Claude Code]
https://x.com/browomo/status/2070581623922299279
A guy replaced an $8,000 survey crew with one DJI Matrice 350 RTK drone and a five-module pipeline on Claude Code, all reporting to one orchestrator, all built solo. Drone flies the route, a Zenmuse L2 laser captures geometry to the centimeter, DJI Terra stitches the point cloud, and a Claude agent builds exactly what the client needs — volume report, progress diff, a rotatable model behind a link. The site becomes a file in about an hour. His cost is one battery charge and ~$300/month to host the models. The pipeline even calls him by voice only when a flight misses a patch.
@v_nefodov [Claude Code]
https://x.com/v_nefodov/status/2070649055781249196
The other surveying story, and just as sharp. He combined a surveying credential with a ZEB Horizon handheld LiDAR scanner and Claude Code. He walks a property, captures millions of spatial points per second, then runs the point cloud through a processing pipeline he built in Claude Code that spits out a full report with measurements, 3D renders and boundary docs. What used to take a surveying firm two weeks takes him two days, alone. Four properties a week at $300/hour. Claude didn't replace the skill — it made one person with that skill worth a whole firm.
@kandmybike [Claude Code]
https://x.com/kandmybike/status/2070469002980421961
A tax accountant building toward what he calls the "one-person AI tax accountant" — freee's bookkeeping automation wired to Claude Code so a single human plus AI can service dozens of companies at high quality. He's running a three-month bootcamp around it, and his own credential is the proof: he handles 60 companies solo. This is the quiet version of AI disruption nobody tweets memes about — not replacing the accountant, replacing the army of junior staff that used to sit behind one.
@duborges [Claude Code]
https://x.com/duborges/status/2070469806612951298
A genuinely clever way to run Claude Code for a fintech without installing any sketchy "claw" packages. He spun up a $20/month VPS, locked it behind Tailscale, and wired Slack Bolt to a bot named DOLA — type /session [message] and it launches a fresh Claude Code with instructions injected. The VPS has no direct DB access; it calls endpoints on the main server instead. The killer example: a KYC submission with a wrong ZIP used to need a hard-coded error handler for every scenario; now the agent just reads the error and retries with corrected data. That's how he and one cofounder run a profitable fintech with over 100 clients in three months.
@0xWast3 [Claude Code]
https://x.com/0xWast3/status/2070435871866675250
A 68-year-old retiree, 35 years a systems engineer, who refused to put his data on cloud servers — so he built a fully offline AI assistant. He bought a GMKtec mini-PC, wiped Windows, installed Ubuntu and Ollama, and pointed a local Claude Code instance at his own machine. Then he wired it to a Raspberry Pi running automations around the house: calendar alerts, morning briefings, document organization, all triggered by voice. No subscription, no data leaving the house, no internet after setup. His inbox filled with other retirees asking how to copy it.
@0xrabi_ [Claude Code]
https://x.com/0xrabi_/status/2070523595659374905
The story I keep coming back to. An electrician on a call: "I've been vibe coding for three months, grabbed an old PC to run tools and Claude Code to simplify my process. I built myself a hot memory and a cold memory — it simplifies generating quotes, managing job sites, everything." No dev background, just a sharp head and enough nerve to solve his own pain. This is the real frontier — not engineers building apps, but tradespeople building their own software because the cost of trying finally hit zero.
@d33v33d0 [Claude Code]
https://x.com/d33v33d0/status/2070380565086470144
Using Claude Code with Opus 4.8 to physically control a truck-demolition robot. The hard part is that Claude has no depth perception, so he rigs cameras around the truck and lets Claude analyze the frame after each move. His best lesson is gloriously low-tech: prolific use of color. Opus reads images as pixels, so he caution-tapes the knuckle and spray-paints red markers on the shed to help it track objects. Cracking one 2x4 on a rafter currently takes 20+ minutes of positioning — slow, but it works, and he's building a custom harness to speed it up.
@plutos_eth [Claude Code]
https://x.com/plutos_eth/status/2070569868559081789
An automation-agency founder built "Jarvis" in Claude Code — first for his own shop, now resold to real estate clients. A new inquiry lands at midnight: it qualifies the lead, books the viewing against the calendar, updates the CRM, and chases the missing paperwork before anyone wakes up. The framing is the point — not one automated workflow, but one system running the whole business. The agency playbook is shifting from "we do the work" to "we sell you the worker."
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2070367524139745327
A clear read on how the PM role is splitting. One group spends the day in Slack — "Can you send the Gong recording?" "How many customers asked for this?" — four messages to document one feature request. The other group automated that whole loop: every incoming request hits a structured intake, customer impact and history pre-filled, and the PM sees a triaged queue instead of a thread. He says it's already happening at $100M startups running the entire company out of Claude Code. The system doesn't replace PM judgment; it captures it once and runs it at scale.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2070517939833635228
The companion piece, and the more striking claim: at one fast-growing startup, PMs aren't just defining features, they're shipping front-end and back-end — not because they learned to code, but because they learned to audit code. A PM opens a repo, asks Claude Code to assess the codebase, and gets a real answer about where the risk concentrates and which parts genuinely need a senior engineer. The old bar was "collaborate with engineers to understand risk." The new bar is "assess it yourself, then pull engineers in only where it's contentious." Customer success is shipping too.
@moritzkremb [Claude Code]
https://x.com/moritzkremb/status/2070515065670046133
A tight demo of a business "second brain and operating system" that holds all his context and runs the work itself: grades his sales calls every morning, writes his newsletters from his videos, builds and uploads his Meta ads, and schedules content across every platform. The same setup runs in either Claude Code or Codex. No grand thread — just a one-screen look at what a solo operator's company actually looks like when the OS is the employee.
@fukuda_CEO [Claude Code]
https://x.com/fukuda_CEO/status/2070370500577259711
An honest take on AI video that most people overhype: free AI editing is mostly low-quality junk, but one specific combo changed his mind — Gemini Omni × GPT Images 2 × Claude Code. Feed it a product photo and a one-line pitch and it auto-proposes four ad concepts, generates the script, adds voiceover, and exports a vertical video. Twelve minutes, 430 yen. He says the quality matches what he used to pay 30,000 yen per video to outsource. The judgment here is what makes it credible — he names exactly which tools and admits where free AI still fails.
@0xQiYan [Claude Code]
https://x.com/0xQiYan/status/2070407326100849059
AI Berkshire — an open-source project that turns Claude into an investment-research tool. You give it a company name and it spins up multiple AI personas in parallel — Buffett, Munger, Duan Yongping, Li Lu — to cross-examine the same company from four styles, then forces a concrete conclusion with a reference price range instead of the usual "maybe, perhaps" mush. Sixteen research skills covering deep research, earnings reads, sector screening and portfolio review. Install it, drop the skills into Claude Code's command directory, and run /investment-research Tencent. He hasn't verified the live returns and says so.
@lowesyang [Claude Code]
https://x.com/lowesyang/status/2070538219566067994
Minara applied to quant trading the same loop Claude Code applied to coding. Pick factors from a 200+ factor library, generate strategies with prompts, let agents iterate from the backtests, validate in live trading — then run the loop again. The framing is clean: research, generate, judge, keep or discard, repeat, all in one place. It's the agentic loop pattern leaving software and walking straight into finance.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2070531462877626688
Where the trading-loop idea gets sharp: in software a loop can babysit PRs and fix CI every 30 minutes, but the moment money can move you need a hard gate before the broker API. His loop has to clear four checks before it sends an order — did the idea survive a fresh backtest, did it hold on out-of-sample data, did max drawdown stay under the rule, and is today blocked by earnings, FOMC or position limits. The whole value is the checker sitting between research and execution. That's how you turn Claude Code from a research assistant into a quant system that can refuse to trade.
@koujikano1 [Claude Code]
https://x.com/koujikano1/status/2070457382166306839
A self-built horse-racing prediction AI, betting one million yen on the weekend, posting real recovery-rate numbers by confidence tier on his blog. Buried in it is one of the few honest cost evals of the week: a straight take on what Claude Code is actually worth at 30,000 yen a month for this kind of non-coding analytical work. No hype, just a hobbyist treating his model like an investment and showing the receipts.
@naruto11eth [Claude Code]
https://x.com/naruto11eth/status/2070622407446925685
The most human use of the week, and one line long: export your last five relationships' chat logs, drop them into Claude Code, and have it psychoanalyze you and explain why each one didn't work out. It's a throwaway tweet but it points at something real — the moment a coding agent can read any text file, it becomes a tool for reading yourself, not just your codebase.
@bonduelleioat [Claude Code]
https://x.com/bonduelleioat/status/2070540254692425835
A 25-year-old with no design or coding background put in $300 and cleared $7,000 net in his first month running an entire clothing brand off a $20 Claude subscription. AI generates the print concepts and variations, writes the descriptions and ad copy, and Claude Code writes and debugs the storefront into a working MVP in hours. He still needs a heat press, a printer and a test batch of shirts — the physical side is real — but the whole digital side now belongs to AI. His line: capital is no longer the expensive resource, speed of execution is.
@0xbeinginvested [Claude Code]
https://x.com/0xbeinginvested/status/2070422387896586425
A free SEO-audit skill that does in 60 seconds what agencies bill $5,000 for. You type /seo-audit, drop your URL, and it runs six agents in parallel — technical SEO, content, schema, sitemaps, performance, visuals — across 13 checks, then scores you out of 100. His came back at 57. The part agencies won't touch: it also audits how you show up in ChatGPT, Perplexity and Google AI Overviews, not just classic search. A few minutes later his score hit 95.
@Veltrxai [Claude Code]
https://x.com/Veltrxai/status/2070524828021293119
A $40,000 client paid for an AI "second brain" built inside Obsidian — not a chatbot, a wiki that runs a business. The full six-step build: point Claude Code at an Obsidian vault, one prompt builds the folder tree with a CLAUDE.md of business context, drag in raw files and Claude rewrites them into linked notes that cite their source, wire Gmail/Calendar/Drive over MCP, save a reusable "weekly brief" skill, run three cheap sub-agents in parallel to scan for risks, then schedule and lock it down. The last step is the one most skip: a simple UI layer, because the non-technical client will never open Claude Code, and that interface is what they're paying $40K for.
@fuxps32 [Claude Code]
https://x.com/fuxps32/status/2070306357266583942
The same second-brain idea with a cockpit bolted on. He pointed Claude Code at his Obsidian vault and told it to track the whole thing, and now one screen shows everything: 376 notes across 32 clusters, maturity 58%, velocity +196%, and a ranking of his top hubs — the ideas his notes keep circling back to. Weekly, Claude flags what's drifting cold and what needs a bridge. Most people store notes and forget they exist; he gave his a control panel.
@Veltrxai [Claude Code]
https://x.com/Veltrxai/status/2070610014885327114
The clearest argument for why Obsidian beats Notion for AI work: your Notion notes live in the cloud, so the agent on your laptop can't touch them without renting Notion's own AI or wiring a rate-limited API. Obsidian notes are just Markdown files in a folder — hand Claude Code the path and it reads, analyzes and rewrites on the spot. The unlock he highlights: one user asked Claude to find the single idea hiding across the most notes, and it surfaced a thread quietly linking his fitness, study and investing notes — three fields he'd never connected himself.
@KingBootoshi [Claude Code]
https://x.com/KingBootoshi/status/2070517837450674366
A strongly-argued case to stop giving agents MCPs and give them CLIs instead. You can load an agent with 100 CLI tools but you can't load five MCPs before context rot sets in — and all models are already trained on terminal commands. His trick: give the agent CLI tools with a help wizard and let it learn in the moment. The smart bit is treating error output as a prompt — if the error explains the problem and how to fix it, the agent self-corrects on the spot. Bearish on MCPs, and he makes the efficiency math hard to argue with.
@zodchiii [Claude Code]
https://x.com/zodchiii/status/2070444119051157559
Anthropic published a PDF on how its own teams actually use Claude Code, and this is the clean breakdown: Spec, Dispatch, Verify, Systemize. Spec — hand Claude a clear goal and let it run instead of typing every line. Dispatch — the growth team splits work across sub-agents for hundreds of outputs in minutes. Verify — Claude runs its own builds, tests and lints, so trust comes from proof. Systemize — repeated workflows become commands, and the security team wrote 50% of them. The headline insight: their engineers don't write more code, they set up the system and review the 80% Claude ships on its own.
@BenceRedmond [Claude Code]
https://x.com/BenceRedmond/status/2070349545989414983
Merged 500+ PRs in a few weeks and ran five review agents on every one, so this is a real bake-off, not a vibe. Greptile: incredibly accurate, almost no false positives, but stops commenting after a few commits. Cursor Bugbot: highest volume, catches P1-P3, doesn't quit. Capy: expensive but regularly catches P0 bugs nothing else found. Cubic's wiki generation is worth it alone. The secret sixth option: just tell Claude Code or Codex to spin up sub-agents to review for UI, scalability and observability — often the most effective and you skip the wait.
@rohit4verse [Claude Code]
https://x.com/rohit4verse/status/2070584600402034930
A Rust CLI proxy that cut his LLM token consumption 60-90%, and the diagnosis is the lesson. For months he blamed the model for his bill; about 80% of the tokens went to the agent re-reading its own command output. Every test log and git diff lands in context at full token cost — a 30-minute Claude Code session burns ~118k tokens on that alone. The tool, rtk, is one Rust binary that filters output before it hits the window: a git push drops from 15 lines to "ok main," a failing cargo test from 200 lines to 20. Same agent, same code, 80% fewer tokens on the boring half of the job.
@chroniki_ai [Claude Code]
https://x.com/chroniki_ai/status/2070410247752257767
For everyone hitting Claude Code limits and reaching for the Max plan — often the real cause is how you send, not the plan. Every message reloads nearly the whole session history, so message 40 carries all 39 before it, and without a .claudeignore, node_modules and image and log files get processed every time too. Four free fixes: exclude junk with .claudeignore, slim CLAUDE.md under 200 lines, /compact mid-session, /clear when switching tasks. One reported run cut CLAUDE.md from 3,847 tokens to 312 and compressed 20,000 tokens down to 1,000-3,000, for a 40-70% cost drop. Try the four before you pay more.
@rasbt [Claude Code]
https://x.com/rasbt/status/2070518167399698490
Sebastian Raschka test-driving local open-weight LLMs across harnesses — Qwen-Code, Codex, Claude Code. His finding: 30B Mixture-of-Expert models are a nice sweet spot, solving real problems at roughly 40 tok/sec on a Mac or DGX Spark, comparable to GPT-5.5 on a Pro plan and fully usable for daily work. The more interesting result is about the harness, not the model: Claude Code appears to use roughly 2x as many tokens as Codex for the same task. A reminder that your token bill is as much about the harness as the model you picked.
@geekbb [Claude Code]
https://x.com/geekbb/status/2070485729088766073
A DeepSeek V4 terminal coding agent that delivers a Claude-Code-grade experience at absurdly low cost. The standout number: seven real open-source bug-fix tasks for a combined ¥1.07, roughly 30x cheaper than Claude Opus, by leaning on prefix-cache aggregation at a 95.8% hit rate to crush the bill. It's the clearest sign of where the floor is heading — the harness experience is becoming a commodity, and the model underneath is increasingly a price decision.
@0xCortexl [Claude Code]
https://x.com/0xCortexl/status/2070401720878768247
A 19-year-old bought a used RTX 3090 for $700 and stopped paying for Claude or ChatGPT. Ethereum killing proof-of-work dumped millions of these 24GB cards onto eBay at $650-800 — the same VRAM as a $2,000 4090. Qwen 3.6 27B loads with room for long context and sends nothing anywhere, scoring 84.1 on RealWorldQA against Opus 4.5 at 77.0. Ollama installs in one command, Claude Code points at localhost via one env variable, and a $5,280/year cloud bill becomes $8/month of electricity. He made his first $30,000 while competitors kept paying — the hardware didn't change, the understanding did.
@maxedapps [Claude Code]
https://x.com/maxedapps/status/2070385520786509905
A small but telling "magic moment." Some websites became unreachable from his MacBook on home wifi, and he didn't care enough to fix it — just hotspotted around it for weeks. GPT-5.5 analyzed it and failed. He gave it another go with Claude Code, which dug deep, found a multipass VM bridge network configuration error, and fixed it. The point isn't that GPT failed once; it's that the model's willingness to keep digging into a problem until it's actually resolved is the new capability worth paying for.
@ZypherHQ [Claude Code]
https://x.com/ZypherHQ/status/2070523625237586169
A clean example of a harness solving a problem it was never given tools for. During an ordinary session he asked Claude Code to fully transcribe a YouTube video — no custom MCP, no instructions. On its own it web-searched the video, then searched for a YouTube audio-to-text transcriber and worked the problem to a solution. Trivial on the surface, but it's the whole gap between a chatbot and a multimodal agent: the agent doesn't need a tool for everything, it can go find the path.
@pakhandrin [Claude Code]
https://x.com/pakhandrin/status/2070396950701076549
A genuinely good investigation into China's gray market for Opus 4.8 tokens at up to 90% off. Claude is officially unavailable there, so "transfer stations" resell access as a normal API. How they make money: one, slice $200 Claude Max subscriptions across many users; two, quietly swap your expensive Opus request for Sonnet, Haiku, or a Chinese model like Qwen — nearly impossible to detect until a hard task makes the model "feel dumber"; three, and the big one, keep the logs. Every prompt, tool call, code chunk and repo context flows through the proxy and gets cleaned, packed and sold as training data. The cheap price isn't a discount, it's a hidden trade — you pay with your data.
@buzzicra [Claude Code]
https://x.com/buzzicra/status/2070447262262509704
The security story everyone needs to read. If you connected Sentry to your agent, one fake error report is enough to run an attacker's code. The Sentry DSN key is public by design — it's in your frontend. An attacker sends a fake error with a hidden command dressed as "resolution steps." You tell your agent "fix these Sentry errors," it reads the error over MCP, treats the fake fix as real, and runs the command with your privileges — AWS keys, GitHub tokens, repo addresses gone. Tenet's test found 2,388 vulnerable organizations at an 85% success rate, the same across Claude Code, Cursor and Codex. The agent can't separate data from instructions, and every MCP source you connect is a door.
@yousukezan [OpenClaw]
https://x.com/yousukezan/status/2070341301351154137
The other side of the security coin — an actual red-team result. A developer ran an AI agent called Fiu on OpenClaw with Claude Opus 4.6, exposed it to a public site, and let 2,000+ people send 6,000+ prompt-injection emails trying to extract its secrets.env. Zero leaks. Fiu had only a few plain rules: don't disclose secrets, don't rewrite files, don't run code from email. Attackers tried impersonating future-self, fake audits, admin spoofing, multilingual social engineering — nothing worked. The cost of the experiment is its own lesson: Gmail suspended the account for three days under the volume, and the API bill topped $500.
@7h3h4ckv157 [Claude Code]
https://x.com/7h3h4ckv157/status/2070389272088354974
NeuroSploit turns a URL, a repo, a running app, or a host into an autonomous security engagement. A Rust harness on tokio drives a pool of LLMs — via API key or your Claude Code / Codex / Gemini / Grok subscription — to recon the target, intelligently pick only the agents that match the discovered attack surface, run them in parallel, and chain findings into deeper impact. The notable design choice: every claim is validated by cross-model voting plus tool-receipt grounding before it makes the report. Ships 329 markdown agents and a Mission Control TUI.
@rewind02 [Claude Code]
https://x.com/rewind02/status/2070443801718452685
Researchers reverse-engineered Claude Code's full source and found the number that reframes the whole tool: only 1.6% of the codebase is actual AI decision logic. The other 98.4% is infrastructure. The 46-page paper maps it from 512,000 lines of real source — the seven layers of safety checks before any command runs, why permissions deliberately reset on resume, the five-layer context compression, and the exact rules for which of the 54 built-in tools get hidden from the model. It's the strongest evidence yet that the model is the easy part; the harness is the product.
@hasantoxr [Claude Code]
https://x.com/hasantoxr/status/2070452451619307524
BrowserAct turns Claude, Cursor and Codex into real browser operators. You give the agent a task, it opens a real browser that looks human, passes the blocks, and hands back clean data — no scraper code, no proxies, no CAPTCHA farm. It auto-solves reCAPTCHA, Cloudflare Turnstile and DataDome mid-task, reuses your logged-in Chrome session and cookies, and runs unlimited agents in parallel each with its own identity. When automation hits an edge case it generates a takeover link so a human can step in and the agent resumes. The demo: open Amazon, hit a Cloudflare wall, solve it in 1.2 seconds, scrape 80 listings, drop a clean CSV — no human touched it.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2070430333787427160
People are quietly building businesses with Claude and one free library: Supervision, Roboflow's open-source computer-vision toolkit. Point it at a video and it detects objects, tracks them, counts people walking into a zone, or clocks a car's speed. The real-business angle is the strong part — Relo Metrics built a company measuring how many seconds a sponsor's logo appears in a live broadcast, then sells that data to brands. Retail pays for foot-traffic analytics, cities for traffic enforcement, factories for defect detection. Open Claude Code, drop in your video, pip install supervision, and tell it what to count.
@totheagi [Claude Code]
https://x.com/totheagi/status/2070392113850880312
Short but striking: he made a game with GLM-5.2 running through Claude Code on his own 32x RTX 4090 rig. No cloud, no frontier API — an open Chinese model on local hardware, driving a full coding harness to ship a playable game. It's a small flex that says a lot about where the open-weight floor now sits.
@DavidTurturean [Claude Code]
https://x.com/DavidTurturean/status/2070532529682087979
A neat multi-model scaffold for hard formal math. He ran Claude Opus 4.8 at max thinking inside Claude Code, with a hook that let it call ChatGPT-Pro whenever it got stuck. Opus drove the Lean compilations while Pro supplied most of the actual Lean code, and the two-model handoff is what finally pushed the proof through. It's a concrete pattern for problems where one model's reasoning and another's domain output beat either alone.
@TheCodeMan__ [Claude Code]
https://x.com/TheCodeMan__/status/2070409980805611654
Killed the "whiteboard where ideas go to die" problem. Every planning session ends with a board of sticky notes that gets screenshotted into a doc and forgotten. This time he took the exact Miro board and shipped a real .NET feature from it without leaving Miro: Miro AI Workflows turned the sticky-note chaos into a structured spec with user stories, acceptance criteria and edge cases, then he connected Claude Code to the same board through the Miro MCP server so the AI built from the team's actual decisions and reasoning — not a blank prompt. The canvas became the shared context layer the AI stack plugs into.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2070490930977952168
A self-built learning app gaining about 200 new daily users — and the delight in the tweet is the point. He asked Claude Code to build an admin dashboard with OAuth through his own Discord account, and it produced something genuinely clean and readable. It's the personal-developer arc in miniature: a non-engineer shipping a real product with real users and an admin panel he didn't have to design. Goal: 10,000 daily users.
@chroniki_ai [Claude Code]
https://x.com/chroniki_ai/status/2070301696392323470
A clear breakdown of why one agent doing everything ships broken work. Anthropic's demo built an app with no human writing code, using a three-agent loop: Planner designs, Builder writes, Judge decides if it works — and if not, back to Planner. The cost contrast is the useful part: a single agent runs ~20 minutes at ~$9, while the three-agent loop runs 3-6 hours at $124-200. Nearly 20x more expensive, but Anthropic's own tests showed single-agent apps "looked like they worked but had broken core features." For production-level work, the expensive loop is actually cheaper.
@Blum_OG [Claude Code]
https://x.com/Blum_OG/status/2070495741047366097
The practical version of the same idea — a 4-agent team where fresh context is the whole feature. Writer writes code and runs the build, Tester starts from the spec and writes tests first without seeing the code, Reviewer reads the diff only and can't edit, and a ship skill briefs the team, calls them in order, and returns one report. The example that sells it: on a login rate-limiter, the reviewer caught that req.ip behind a proxy means one office shares one bucket — a bug the writer would never see because it was baked into its own path. Ten-minute setup, all in markdown files.
@iximiuz [Claude Code]
https://x.com/iximiuz/status/2070454177851998626
A new daily routine he calls "infra kaizen." He starts the day by asking Claude Code to find repetitive error patterns in prod logs and fix them, reduce log volume, and investigate flaky CI jobs. It's a small, durable use case that doesn't make headlines — not a $100K business, just steady continuous improvement on the unglamorous infrastructure work nobody wants to do manually. The next step, he says, is wrapping it in The Loop so it runs itself.
@CDGalpha [Claude Code]
https://x.com/CDGalpha/status/2070382511562366979
With 10,000+ Claude Code plugins out there, his whole dev workflow runs on six. obra/superpowers adds a brainstorm mode that refuses to touch your codebase until it's answered five scoping questions. Anthropic's official frontend-design picks aesthetic, palette and typography automatically. A code-review plugin spawns four parallel agents — one even reads git history and caught a null guard silently removed five weeks earlier by another dev. Anthropic's security-guidance runs as a live hook flagging eight vulnerability types as Claude writes them. claude-mem gives persistent memory across sessions, and Garry Tan's gstack runs CEO and engineering reviews in seconds.
@Sprytixl [Claude Code]
https://x.com/Sprytixl/status/2070402356424871962
An 18-year-old Chinese developer, laid off with $800 and no clients, now clears $18,000/month from a one-man web studio. The mechanics are concrete: tiered pricing so a dentist comparing you to a $5,000 agency pays $3,200 and feels like a deal; a CLAUDE.md in every project root that replaces a senior dev's entire onboarding checklist; 10 niche templates pre-built to 90% in a private GitHub, so a fork plus one personalization prompt is 8 minutes to a finished site. 38 minutes from empty folder to a $3,200 Stripe invoice, 4-6 sites a week, and every site converts to $200-600/month maintenance — 31 retainers compounding eight months in.
@deedydas [OpenClaw]
https://x.com/deedydas/status/2070574557728264382
Three lessons from an intimate Agentic Engineering event in SF. From OpenClaw's creator: he now forces contributors to use a skill that pushes their prompt history with each change, to find signal and avoid bad 10,000-line PRs from a one-word "fix this" prompt. From another: he used Claude as a video editor to make a launch video while it interactively taught him color grading — finding your unknown unknowns is how you get the most from a model. From a third: spend more human energy crafting the plan and clarifications upfront, then leave Codex to spin for days on a well-built /goal.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2070629740520566961
The honest counterweight to the hype, and it's brutal. Eight hours building a workflow to ingest meetings into gbrain, got it perfect, had the agent save it as a skill. Next day he asks for a quick change, the agent reads from compaction context and overwrites the skill with stale assumptions. Eight more hours fixing it, then it forgets where to read messages, pulls a two-day-old message with the wrong rules, sees all the correct new work and starts undoing it. His verdict: Hermes does this less than OpenClaw, which "actually doesn't work at all anymore," but he's exhausted by being stuck in the plumbing. The agents are powerful; the reliability tax is real.
@nityeshaga [Claude Code]
https://x.com/nityeshaga/status/2070571633140813887
The clarifying take of the week: Claude Code is the OpenClaw alternative you already had. The trick is understanding the difference between a model and a harness — OpenClaw is just a harness on top of AI models, and so is Claude Code, it only got marketed as a coding tool. Let Claude Code run from your home folder with free access and you get exactly what made OpenClaw exciting, often better. He backs it with a video walkthrough of the "AI employees" he built on the Claude Code harness, aimed squarely at non-programmers.
@AzFlin [Claude Code]
https://x.com/AzFlin/status/2070374192537538771
A perfect snapshot of the current agent reality: he keeps having to start a Claude Code session inside his ~/.openclaw folder, paste it the docs and terminal output, and debug his OpenClaw like it's actual code. One agent debugging another agent's config. It's funny, but it's also the real state of self-hosted agents right now — powerful enough to run your life, fragile enough to need a second agent babysitting the first.
🗣 User Voice
User Voice
The pattern across today's feed is that users have stopped asking whether the model is smart and started fighting the harness around it. Their wants are converging.
Memory is the number one complaint, still unsolved. Every session starts from zero, and the workarounds — Obsidian vaults, claude-mem, session-checkpoint skills — are all bolted on, not native. As @SpikeCalls put it, "Every Claude Code session starts from zero. It forgets your code, your decisions, your context. You re-explain the same things every morning like a broken loop." People want memory that survives compaction without a plugin.
Context compaction is silently destroying work. @bradmillscan lost two full days because an agent re-read stale compaction context and overwrote a correct skill. Users want compaction they can trust — or at least see — not a black box that quietly forgets the rules mid-task.
Token economics are the real bill, and the harness is the culprit. @rohit4verse found ~80% of his tokens went to the agent re-reading its own output, and @rasbt measured Claude Code using roughly 2x the tokens of Codex for the same job. Users want the harness to stop paying full freight to re-read git diffs and test logs.
Security is the new anxiety, and it's justified. @buzzicra and @yousukezan are two sides of the same realization: an agent that reads your tools can be commanded by them. Users want a real boundary between data and instructions, because right now "don't trust external input" in the system prompt doesn't actually hold.
Reliability over capability. The most-felt frustration isn't "Claude isn't smart enough," it's @james406's: you can build the perfect prompt, the perfect loop, the perfect /goal — and none of it matters if you return to find Claude Code asking you to log in again before any work started. People want the agent to just keep running.
The pattern across today's feed is that users have stopped asking whether the model is smart and started fighting the harness around it. Their wants are converging.
Memory is the number one complaint, still unsolved. Every session starts from zero, and the workarounds — Obsidian vaults, claude-mem, session-checkpoint skills — are all bolted on, not native. As @SpikeCalls put it, "Every Claude Code session starts from zero. It forgets your code, your decisions, your context. You re-explain the same things every morning like a broken loop." People want memory that survives compaction without a plugin.
Context compaction is silently destroying work. @bradmillscan lost two full days because an agent re-read stale compaction context and overwrote a correct skill. Users want compaction they can trust — or at least see — not a black box that quietly forgets the rules mid-task.
Token economics are the real bill, and the harness is the culprit. @rohit4verse found ~80% of his tokens went to the agent re-reading its own output, and @rasbt measured Claude Code using roughly 2x the tokens of Codex for the same job. Users want the harness to stop paying full freight to re-read git diffs and test logs.
Security is the new anxiety, and it's justified. @buzzicra and @yousukezan are two sides of the same realization: an agent that reads your tools can be commanded by them. Users want a real boundary between data and instructions, because right now "don't trust external input" in the system prompt doesn't actually hold.
Reliability over capability. The most-felt frustration isn't "Claude isn't smart enough," it's @james406's: you can build the perfect prompt, the perfect loop, the perfect /goal — and none of it matters if you return to find Claude Code asking you to log in again before any work started. People want the agent to just keep running.
📡 Eco Products Radar
Eco Products Radar
Products mentioned 3+ times across today's posts.
Codex (174) — the constant comparison point; the token-efficiency lead flipped toward Claude Code this week.
MCP (91) — the default integration layer, and increasingly the default attack surface.
OpenClaw (79) — still the reference for "personal agent," but the reliability complaints are mounting.
Cursor (62) — paired with Claude Code as often as it's compared to it.
Hermes (54) — the OpenClaw alternative people are migrating to for stability.
Obsidian (46) — the runaway winner for AI memory: local Markdown files an agent can actually read.
GLM (29) — the open-weight model people run locally and overnight to dodge API bills.
Gemini (25) — showing up in video and multimodal pipelines.
Qwen (24) — the local-model default on used-GPU rigs.
n8n (14) — still quietly running production automations for 12+ months.
Higgsfield (11) — the content-factory connector for video.
Ollama (11) — one command to go fully local.
DeepSeek (11) — the cheap-backend swap, including the ¥1.07 bug-fix run.
Lovable (8), Copilot (8), Perplexity (7), Vercel (7), gstack (6), Framer (6), Calendly (6), claude-mem (5), Supervision/Roboflow (3), Miro (3), Whisper (3).
Products mentioned 3+ times across today's posts.
Codex (174) — the constant comparison point; the token-efficiency lead flipped toward Claude Code this week.
MCP (91) — the default integration layer, and increasingly the default attack surface.
OpenClaw (79) — still the reference for "personal agent," but the reliability complaints are mounting.
Cursor (62) — paired with Claude Code as often as it's compared to it.
Hermes (54) — the OpenClaw alternative people are migrating to for stability.
Obsidian (46) — the runaway winner for AI memory: local Markdown files an agent can actually read.
GLM (29) — the open-weight model people run locally and overnight to dodge API bills.
Gemini (25) — showing up in video and multimodal pipelines.
Qwen (24) — the local-model default on used-GPU rigs.
n8n (14) — still quietly running production automations for 12+ months.
Higgsfield (11) — the content-factory connector for video.
Ollama (11) — one command to go fully local.
DeepSeek (11) — the cheap-backend swap, including the ¥1.07 bug-fix run.
Lovable (8), Copilot (8), Perplexity (7), Vercel (7), gstack (6), Framer (6), Calendly (6), claude-mem (5), Supervision/Roboflow (3), Miro (3), Whisper (3).
Comments