Super User Daily: June 30, 2026
Two stories defined today, and neither was about writing code. The first is the garage scientist: a guy designed a novel Alzheimer's drug candidate, PAC-832, with the wet-lab work done by an OpenTrons liquid-handling robot programmed by Claude Code. The second is money — a wave of operators who quietly turned an agent into a business. A solo dev's Capterra-review SaaS clearing $47K a month. A 23-year-old's phone-farm arbitrage rig. A Chinese student who sold a traffic-detection system to a logistics company for $11,000. The throughline everyone kept repeating: stop prompting the agent, build the loop that prompts it for you. The model is solved; the scarce thing now is judgment, verification, and memory. Here's who actually did something with it yesterday.
@iScienceLuvr [Claude Code]
https://x.com/iScienceLuvr/status/2071064383883768233
A guy built a pharma company in his garage and developed a real Alzheimer's drug candidate, PAC-832, the first selective GalR1 antagonist with low toxicity, already entering IND-enabling studies. The part that matters here: all the in-vitro screening was run by an OpenTrons OT-2 liquid-handling robot programmed by Claude Code, with LLMs woven into nearly every step of discovery. The lab used to be the bottleneck. It isn't anymore. This is the single most striking "non-coding" use of Claude Code this whole cycle.
@m_goes_distance [Claude Code]
https://x.com/m_goes_distance/status/2071262351815283096
Same PAC-832 drug, but zoomed out to the pattern, and the pattern is the scary part. He lists three garage-bio cases this year alone: this Alzheimer's candidate, an Australian founder who sequenced his dog's cancer and used AI to design a custom mRNA vaccine that shrank the tumor, and a biohacker who sequenced his own genome on a kitchen table with a $3,200 sequencer and Claude. One person doing the work that used to need a 20-person lab. The bottleneck was never the lab.
@nasqret [Claude Code]
https://x.com/nasqret/status/2071223836289138910
A mathematician describes how his "auto-research" methodology helped disprove the slope conjecture, a real result after three years of work. His setup isn't autonomous discovery, it's a scaffolded loop: agents, coding tools, literature search, heavy computation, memory files, and constant human steering. The honest insight is the best part: he sees glimpses of self-correcting research loops where, once the scaffold is right, the agent closes the gap on its own, but taste and judgment still decide what the work means. This is what serious research-grade agent use actually looks like.
@danglar_ [Claude Code]
https://x.com/danglar_/status/2071191093798961449
A Lisbon dev built a tool over one weekend that watches Capterra reviews for an accounting app and flags the features people keep begging for. No logo, no landing page, just a script, a Stripe webhook, and a cron job, with Claude Code writing the plumbing while she reviewed. She validated demand first by reading one-star competitor reviews and collecting 20 emails off a fake landing page. Twelve months later: $47,812 in a single month, 1,650 people paying $29, costing $30/month to run.
@waveking1314 [Claude Code]
https://x.com/waveking1314/status/2071145918993445307
A breakdown of a Chinese student who pulled $2.29M on Polymarket across 768 bets, then accidentally exposed his wallet (username gatorr) in a 5-second tab-switch in his "learning AI from zero" video. The actual edge: he used Claude to hunt low-liquidity sports sub-markets with fewer than 20 active users, trading in rooms nobody else was in, buying at $0.43-0.51 and cashing at $1. The whiteboard everyone thought said "Python basics" actually held Claude Code scripts, data sources, and entry timing. He deleted all 30 videos within an hour. Too late.
@sol_jingou [Claude Code]
https://x.com/sol_jingou/status/2071054516037029951
A 23-year-old in Chicago wrote a "high-frequency trading farm" over four weekends with Claude Code: seven old phones flat on a park table, each logged into a separate account, with a Mac mini as the brain. It scans 80+ markets around the clock for pricing gaps, auto-fires orders, holds an average of 14 minutes, and books a 71.4% win rate. The monthly numbers he claims climb from $194K in February to $487K in April. Treat the figures with salt, but the architecture, cheap hardware plus a self-written scanning loop, is the real artifact.
@shuigvn [Claude Code]
https://x.com/shuigvn/status/2071149825475916093
In logistics across developing countries, off-the-shelf vision tools fall apart at dusty intersections with no lane markings, overloaded motorbikes, and chaotic traffic. A Chinese student built a real-time object-detection system with Claude Code that actually handled those conditions, then sold it to a logistics company for $11,000. Small, concrete, and exactly the kind of last-mile problem the big models won't ship for you.
@0xAIGOATexe [Claude Code]
https://x.com/0xAIGOATexe/status/2071219565066145886
The restaurant-website play, but pointed at everything. An agency quoted a steakhouse $8,400 for a four-page site with reservations; the same build shipped in 25 minutes on Opus 4.8 for $32 total, a 97% margin. He then ran the same primitive against a movie marketing page, an insurance mobile UI, and a cocktail brand. The restaurant niche was just the wedge. Every category that still pays an agency for a static template is the actual market.
@browomo [Claude Code]
https://x.com/browomo/status/2071194464341590355
A one-person mixed-reality studio: a regular MR headset, Unity, and Claude Code writing all the spatial code. He scans the room, pins anchors, builds the palm-up gesture trigger, renders the portal and energy grid, drops in the client's product, then records and edits the walkthrough, all in about an hour. A branded MR clip goes for $300-$1,500, a prototype $1,500-$5,000, work an XR studio of six would budget a week for. The job shifted from writing XR code to having the visual idea.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2071283020049604792
A $7M agency runs almost every department inside Claude Code, and the per-department detail is what sells it. A full outbound campaign ships in under 30 minutes loaded with their ICP and 15 skills; sales preps every closing call off CRM and transcripts; ads run ~$400K/month straight from the terminal; recruitment turns a hiring-manager call into a JD and a take-home, then scores hundreds of applicants. The rule he repeats: the agent does the repetitive work, a person still makes every call. His advice, don't start with eight departments, fix the one job eating your week first.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2071221244184105376
A creator agency charges $10K/month to run a gaming channel; he rebuilt it as four prompts. Research dept finds climbing-but-thin topics, the writers' room turns a take into a retention-spined VO script, the motion studio outputs a Higgsfield shot list, the growth desk produces titles ranked by CTR and a 30-second Short. Claude Code runs research, scripts, and packaging; Higgsfield renders. Four or five reviews a week at roughly a dollar a video. His line: the agency was never the talent, it was the overhead.
@maverickecom [Claude Code]
https://x.com/maverickecom/status/2071225854596649282
An e-commerce operator turned Opus 4.8 into a 7-skill UGC machine for scaling DTC brands. A brand-rulebook skill bakes in voice and compliance lines for peptide/telehealth; an authority-figure skill builds one locked AI persona that shows up daily; a hook library tests dozens of hooks a week and hands winners to human creators; a funnel skill routes comment-to-DM to Amazon attribution. The loop is Research, Build, Script, Post, Track, Scale. Whether or not you buy the BCG-killer framing, the skill decomposition is a usable template.
@IAmAaronWill [Claude Code]
https://x.com/IAmAaronWill/status/2071275263577477517
Short and concrete: he replaced a $4K/month sales rep with Claude Code that finds leads who need help, ranks them best-to-worst, sends a personalized email in 90 seconds, auto-follows-up for five days, and tracks every open, click, and reply, running while he sleeps. Not a thread of theory, just the job list a salaried role used to own.
@egocgp [Claude Code]
https://x.com/egocgp/status/2071143563858166243
A French wealth advisor argues Finary is literally reproducible and better for less. What costs Finary money is salaries, offices, and bank-account connections; a wealth-management team with Claude Code keeps only the bank connections and runs the same diagnostics for maybe €10/month, even on complex portfolios. His sharper point is the conflict of interest: Finary has to sell advice to justify its model, which caps how far its features go. He thinks the advisor market shrinks hard.
@mickeyhardy [Claude Code]
https://x.com/mickeyhardy/status/2071242441420599301
Seven daily AI workflows behind an 8-figure business. A 5 AM CEO brief scans Fireflies transcripts, Discord, Telegram, and email to report what every employee did yesterday and today's priorities. An AI rolodex of everyone he's met answers "who do I know at Anthropic, in Miami, in AI infra." An ops layer watches comms and auto-updates the CRM. The most valuable one, he says, is using AI as a sparring partner that argues both sides of a decision, not an oracle you accept.
@Marius_Kakuzu [Claude Code]
https://x.com/Marius_Kakuzu/status/2071111353373225347
He built an internal SEO tool for his agency with nothing but Claude Code and a Google Sheet, no paid SEO SaaS. It runs semantic analysis on a market, finds search volumes worth chasing, filters by relevance to his brand, then generates articles with genuinely useful text and images. His own site runs on it. His open challenge to SEO freelancers: tell me, honestly, why I'd still pay an expert.
@levelsio [Claude Code]
https://x.com/levelsio/status/2071162399864889705
Pieter Levels on nearly a year of coding almost entirely on his VPS with Claude Code. No laptop battery drain, switch to his phone whenever, and it keeps going all night with /goal while he sleeps. The wild part: it live-edits his production server, which sounds reckless but in 12 months caused exactly two ten-second outages. For a solo operator he says it's fine; for a company he'd put a staging server in front. The agent moved from a local plugin to a permanent resident on the server.
@AYi_AInotes [Claude Code]
https://x.com/AYi_AInotes/status/2071268380615594302
A Chinese write-up of that same Levels production-VPS pattern that draws out the deeper shifts. The agent's role changed from local IDE helper to a permanent executor living next to the running code. The speed compounds for solo devs, not "a bit faster" but an order-of-magnitude jump in experiments per day. And the infrastructure flipped: the cloud is now the primary dev-and-run environment, the local device just a terminal you SSH in from.
@gengdaJ [OpenClaw]
https://x.com/gengdaJ/status/2071061503764226193
The full eight-step recipe for wiring Codex into WeChat through the OpenClaw Gateway. Install and run the gateway as the relay, install the WeChat plugin and bind by QR, install the Codex plugin, OAuth into OpenAI, switch the default model to gpt-5.5, pin the runtime to Codex, crank thinking to xhigh, restart and verify. The payoff: from his phone, in WeChat, he can call his always-on Mac to parse chat logs and run social-media scraping skills to gather content ideas anywhere.
@0xluffy_eth [OpenClaw]
https://x.com/0xluffy_eth/status/2071036929676902831
A non-coding one worth noting: clearing 500+ unread emails in 10 minutes with no OpenClaw or Hermes plumbing at all, just the AI assistant built into NetEase Mail. Tell it to sort recent mail, it tags notifications/bills/codes; summarize the long ones; use suggested-reply to draft responses; then ask it directly "did I miss any important emails, any credit-card bills, any interview notices." A one-hour chore done in ten minutes, with the human only making the decisions.
@huoshan007 [Claude Code]
https://x.com/huoshan007/status/2071076873988194539
Agent-Reach solves the thing every agent hits: getting blocked when it tries to fetch data. One command wires up 14 platforms (Twitter, Xiaohongshu, B站, YouTube, RSS, 公众号, Weibo, V2EX) for free, picking and configuring the right CLI tools itself, yt-dlp for subtitles, twitter-cli for posts, xhs-cli for Xiaohongshu. He had it scrape and summarize the comments on a Xiaohongshu post in ten seconds. The gap it fills is real: agents need eyes on the web that don't get paywalled.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2071234730784092443
A developer gave a Claude Code agent the logins to a cluster of eight DGX Sparks, said go, and went to sleep before a trip, waking to a fully configured $23,000 server. His own line: "this is probably how Skynet started." The economics are the point, each Spark is $4,000 and fits in a backpack, eight cabled together is a cluster that needed a data center a year ago, and a $100/month Claude Code sub did the setup overnight.
@N01ennn [Claude Code]
https://x.com/N01ennn/status/2071187350613311992
A hardware hacker bolted a Tesla V100 SXM2 stripped from a decommissioned DGX node into an HP Z8 G4 desktop for under $250, using an SXM2-to-PCIe adapter, a copper heatsink, and five layers of thermal pad. The build log is honest: the first attempt failed because the older Z620's CPU lacked AVX2, the latest CUDA didn't detect the card so he dropped to CUDA 11, and he flipped above-4G-decoding in BIOS. It runs GPT-OSS 20B at 87C and 180W. Datacenter-grade HBM2 on a desk for the price of one month of Claude Code Max.
@NoProductFit [Claude Code]
https://x.com/NoProductFit/status/2071224094691913837
A shelf of stacked Mac Minis that replaces $4,080/year in AI subscriptions. The pitch: most people pay seven companies for AI and use three tools, while one $599 Mac Mini M4 runs 7B/8B models faster than Windows boxes that cost double, thanks to the shared memory pool. ollama pull, one docker line for the UI, point Claude Code at localhost. It draws 10-30 watts, costs ~$3/month in power, and pays back a $20 sub in three months.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2071201260968014102
The most useful of the budget-rig posts: pushing an ex-mining AMD BC 250 node from 25 to 175 tokens/second. He started at 25.67 t/s on ollama 0.19, llama.cpp updates took it to 37, a governor clock unlock and RAM bump hit 54.99, and batched serving for multiple users compounded it to 174.95. These were PS5 chips with cores disabled, sold as crypto miners, dumped as e-waste when Ethereum died, about $25 a card in a 12-pack. A 24-node cluster is his next move.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2071271959732552188
The honest counterweight to the cheap-hardware hype: the labor nobody prices in. He's flashing 24 ex-mining BC 250 BIOS chips three at a time under a DIY fan duct, wiring a variable DC supply to a 12V fan dialed to 6.93V to cool them while reflashing each to firmware p2.00. The math still beats $5,500/year in subscriptions, but only if you treat the hours of BIOS flashing and cooling rigs as a hobby. On a per-hour basis, he admits a $700 used 3090 stays the cleanest path.
@Sentdex [Claude Code]
https://x.com/Sentdex/status/2071045444226986154
A pointed take on cost and dependency: run GLM 5.2 through a US provider that doesn't retain prompts, and you're "missing out on NOTHING" versus Opus 4.8 or GPT-5.5. Same speed as Claude Code, goes down less often, and on OpenRouter effectively never goes down because you can swap providers if one drops. His framework: route simple tasks to cheaper models, save GLM 5.2 for the heavy lifting, and stop paying a premium out of habit.
@AKirtesh [Claude Code]
https://x.com/AKirtesh/status/2071186860366250390
Concrete cost cut: he added MiniMax M2.7 support to an open-source Claude Code CLI and now runs the whole thing 95% cheaper at near-Opus performance. One line, real swap, the same theme everyone is converging on, the harness is portable and the model underneath is increasingly a commodity you pick on price.
@leploutos [Claude Code]
https://x.com/leploutos/status/2071121981551047039
A careful debunking of the viral "a Chinese model matches Claude Mythos at finding security bugs" claim, and the real finding underneath is more interesting. The Semgrep benchmark tested one vulnerability class (IDOR); GLM 5.2 hit 39% F1, beating the Claude Code configs (28-37%), but that's the generalist coding agent on Opus, not Mythos. The genuinely notable part: GLM 5.2 did it in a bare prompt, no custom harness, for about $0.17 per bug found, a sixth of a frontier model's price.
@_vmlops [Claude Code]
https://x.com/_vmlops/status/2071112402251153777
A clean summary of Anthropic's Loop Engineering playbook, and the framing is the takeaway: you're no longer supposed to prompt the agent, you're supposed to build the system that prompts the agent for you. Four layers stack up, prompt, context, harness, then loop engineering on top, where the loop wakes on a timer, spawns sub-agents, and feeds its own output back as input. Stripe runs this for 1,300+ machine-written PRs merged per week. The warning attached: verification debt, comprehension rot, cognitive surrender, token blowout.
@Axel_bitblaze69 [Claude Code]
https://x.com/Axel_bitblaze69/status/2071356929679696246
A first-person account of the exact shift from prompting to loops. He used to be the engine, typing a prompt, reading output, fixing, repeating, capped by how many hours he could stare at a terminal. Now he builds loops with a defined "done" that work, check their own output, and only ping him when they genuinely need a human. The concrete win: a free repo called loop-engineering with 7 ready-made loops (daily triage, PR babysitter, CI sweeper) plus CLIs that estimate token cost before a run. Start with daily triage, he says, you'll feel it by day three.
@yonann [Claude Code]
https://x.com/yonann/status/2071291790452232613
Alex Finn's concrete loop, in his own words: a build skill that loops every 20 minutes throughout the day, constantly building out his product. Each cycle reviews the work, hunts security vulnerabilities and bugs, actually tests the code, records a video of itself testing, and opens a pull request. He wakes to PRs in Slack. This is the cleanest small example of a self-running build loop with a verification step baked in.
@wangray [Claude Code]
https://x.com/wangray/status/2071266780182683748
Notes from a Lenny interview with Fiona Fung, who runs Claude Code and Cowork engineering at Anthropic, and the org changes only a place like that would hit first. Verification has replaced coding as the bottleneck, now that designers, PMs, even legal and finance are shipping code. She keeps a permanent Claude Code session wired to every repo, Slack channel, and metric. Hiring collapsed to two types: creative builders with product sense and deep systems experts; the middle layer is gone. And six-month roadmaps became just-in-time monthly planning.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2071232485225083352
After Anthropic's Claude Code PR volume jumped 200%, Maya showed the routine they use to stop docs from falling behind. The control loop: Claude checks merged changes against the docs repo, opens a docs PR when it finds a gap, pings Slack when it needs attention, and lets you open, steer, stop, or resume the run. A second version fires off a GitHub issue. A small, real example of an agent maintaining a system rather than building one.
@Nyra_nx [Claude Code]
https://x.com/Nyra_nx/status/2071174997238796500
Claude Code's new live-dashboard feature (shipped June 18, Team and Enterprise only) turns a terminal session into a real, shareable URL that updates in real time as the code runs. The old loop, build it, screenshot it, paste into Slack, field a question, redo, paste again, is dead. The difference from 2024 Artifacts: this comes from your actual repo and connected tools, not a chatbot. Once a team shares these links, you don't rip it out.
@mpoilerfx [Claude Code]
https://x.com/mpoilerfx/status/2071252838299107424
A real UltraCode run: one perf-audit request fanned out into 16 agents in parallel, 12 already done, the whole sweep clocking 1m49s, each agent on Opus 4.8 with its own token and tool budget. You turn it on with /effort, ultracode is xhigh reasoning plus workflow orchestration, and Claude decides when a task earns a workflow. The honest caveat: it eats tokens, you can burn a day's Pro usage in five minutes, so it's a Max-plan weapon for refactors and cross-file work.
@boney2r [Claude Code]
https://x.com/boney2r/status/2071376654752944243
Ruflo, an open-source MIT-licensed agent framework, runs 100+ specialized agents in parallel on one task sharing the same memory, one plans, one codes, one tests, one handles security. Every run feeds a reasoning bank so each agent sharpens next time, and vector memory makes recall dramatically faster. The claimed win: a 75% cut in the Claude API bill via tiered routing, basic tasks to a free tier, hard tasks to the best model.
@danglar_ [Claude Code]
https://x.com/danglar_/status/2071251208119931186
A contract engineer spent three years migrating a Zig codebase to Rust one file at a time and never finished, then flipped Claude Code's effort menu to ultracode. Hundreds of agents went to work, one mapping struct fields, others rewriting files, others trying to break those answers in a loop until nothing cracked. 750,000 lines, eleven days, 99.8% of the test suite still passing, with the agents even finding wasteful data copies and opening PRs for each while he slept.
@threepointone [Claude Code]
https://x.com/threepointone/status/2071321123468591516
A sharp technical demo: using Think's subagents to drive Claude Code on a Cloudflare sandbox, with the nice trick of hijacking egress to route it through the Cloudflare AI gateway via the binding, so there are zero API keys anywhere. Durability comes out of the box. A clean pattern for running coding agents in an isolated, keyless cloud environment.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2071058674718810387
The Claude Protocol takes the opposite tack from "write a longer prompt to orchestrate agents", it physically blocks bad operations. Thirteen hooks across five lifecycle events, all block not warn: the orchestrator can't write on main, can't dispatch without a bead ID, can't close an epic with open subtasks, and even git commit --no-verify is killed. The rule: one bead equals one git worktree equals one PR. Reliability comes from deterministic shell hooks and git primitives, not from hoping the model behaves.
@AseemShrey [Claude Code]
https://x.com/AseemShrey/status/2071275989678891082
The most under-discussed risk, stated plainly: your coding agent just read your .env, your AWS credentials, and a client's DB connection string, "just checking the project setup." He's building agentjail in public, policy guardrails that sit between the agent and your system and check every tool call in ~8ms before it runs. cat ~/.aws/credentials denied, curl|bash denied, git push --force origin main denied. One command to install, works with Claude Code, Cursor, and Codex.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2071300605239517406
A three-line skill that fixes the most annoying thing about Claude Code: it gets dumber the longer a session runs, usually degrading around 120k tokens as attention spreads thin. The handoff skill (from Matt Pocock's 147k-star repo) compresses everything important from the current session into a clean markdown file and hands it to a fresh agent that picks up with full focus. Because it's just a file, it works across tools, Claude Code to Codex to Copilot.
@sora19ai [Claude Code]
https://x.com/sora19ai/status/2071355723892814242
claude-mem solves the same memory-loss problem a different way: it uses hooks to log your work to SQLite and inject a summary into the next session, so heavy investigations and refactors survive a /clear. One command, npx claude-mem install, Apache-2.0, Node 20+. The persistent-memory theme keeps showing up because it's the single biggest day-to-day pain.
@Huahuazo [Claude Code]
https://x.com/Huahuazo/status/2071032156705722563
obsidian-second-brain turns your note vault into a self-updating AI memory wired into Claude Code as a skill. New content doesn't just append, it edits existing notes, reconciles contradictions automatically, and spins up new pages when it spots a hidden pattern. It ships with 31 slash commands, 4 scheduled agents that tidy up overnight, and 4 preset roles. For people living in Claude Code plus Obsidian, this is the memory layer they keep reaching for.
@Faazsh [Claude Code]
https://x.com/Faazsh/status/2071125656902201584
"I'm deleting Notion AI because of this." claude-obsidian turns Obsidian into a knowledge engine that creates and maintains the notes themselves, not a chat layer on top of them. It auto-organizes everything into entities and concepts, flags contradictions across the vault with source-linked callouts, and runs autonomous 3-round web research that files its own output. One vault can feed all your Claude Code projects. The recurring lesson: structure beats a bigger context window.
@rugikkk [Claude Code]
https://x.com/rugikkk/status/2071262367866912855
A clean walkthrough of wiring Claude Code to Obsidian to kill context loss. Install the BRAT plugin, add the Cloud Code MCP repo, keep Obsidian open, and the agent gets direct access to one local vault that holds the whole project's memory. First prompt builds an architecture.md with linked components; then notes for every service with mapped dependencies. The daily loop seals it: morning "read activecontext and tell me what's next," save every decision with its why during work, update what shipped and what's blocked at end of day.
@HodlReaper [Claude Code]
https://x.com/HodlReaper/status/2071227807523385742
A 19-year-old killed Claude Code's hallucinations by handing it one Obsidian vault split three ways. Raw catches everything (tweets, PDFs, articles), Wiki turns the dump into clean linked pages, Output holds deliverables like slides and specs, and a master index sits on top as the table of contents. The lesson he draws is the one that keeps repeating across these posts: the tool isn't the trick, the map is. Structure gets you precise answers; noise gets you guesses.
@moritzkremb [Claude Code]
https://x.com/moritzkremb/status/2071232325728473434
A tight non-coding workflow: writing a newsletter in 15 minutes a day instead of 4 hours, and not AI slop. Film the video in Tella, pull the transcript into Claude Code via its MCP, draft from saved newsletter templates plus the transcript (90% there, he finishes the last 10%), then say "draft it in Beehiiv" and its MCP builds the whole formatted thing. His prediction: every tool is getting an MCP, and soon you never leave Claude Code.
@aditharun_ [Claude Code]
https://x.com/aditharun_/status/2071091899801505922
A small, specific skill that removes real daily friction: most clinical-trial documents he reads (press releases, slides, posters, SEC filings) cite an NCT ID but don't link it, forcing him to paste it into a browser and sift links. So he built a skill that opens the trial page straight from the NCT ID or a plain description like "daraxonrasib RASolute302." He keeps a Claude Code terminal open and just pastes there. The kind of tiny domain tool that wouldn't have been worth building before.
@XAMTO_AI [Claude Code]
https://x.com/XAMTO_AI/status/2071215286385324488
Career-Ops turns Claude Code into a job-search command center: drop in a job URL and it scrapes the posting, reads your resume, scores fit on an A-F scale, and outputs a matched analysis, salary research, interview prep, and an ATS-optimized resume PDF. The author evaluated 740+ jobs and generated 100+ tailored resumes with it, then landed a Head of Applied AI offer. The design choice that matters: it's a filter, not a spray-and-pray tool, it tells you not to apply below a 4.0.
@ScarletKc_ [Claude Code]
https://x.com/ScarletKc_/status/2071168648253964660
A real why-I-prefer-Claude-Code account, framed against Codex. Her complaint about Codex: it over-infers, quietly does work she didn't ask for, adds settings she never mentioned, and writes junk tests that lock bad logic in as "correct," so she's stuck reviewing every line. She wrote an agents.md to stop it and it still wouldn't stop. What she likes about Claude Code: it pauses on its own to ask, with options, even without plan mode on.
@yhslgg [Claude Code]
https://x.com/yhslgg/status/2071133647151985041
Matt Pocock open-sourced his .claude directory (now ~148k stars), turning decades of TypeScript engineering judgment into installable Claude Code skills. One npx command gives you /tdd (forces red-green-refactor, AI must write tests first), /diagnosing-bugs (reproduce, minimize, hypothesize, verify, fix), /grill-with-docs (build a domain model before coding), /to-prd, /handoff, and an architecture scanner. His README line: this is for real engineers, not vibe coding.
@itsharmanjot [Claude Code]
https://x.com/itsharmanjot/status/2071136827306852766
Claude Skills, an open-source plugin that turns Claude Code into a stack-specific senior engineer instead of a generalist. It detects your framework from context and activates the right expert skill, loading only the references it needs via a progressive-disclosure pattern (80 lines lean, then surgical refs), covering 54 specializations across 25+ frameworks. The interesting bit is the architecture, expert depth without the token bloat of dumping full context every turn.
@srishticodes [Claude Code]
https://x.com/srishticodes/status/2071194185546527168
Repowise indexes a codebase into five layers, a dependency graph, git-history hotspots, auto-generated per-module docs, mined architectural decisions, and a 25-biomarker code-health score computed with zero LLM calls, all exposed through 9 MCP tools. Her honest breakdown of the measured results: tool calls down 70%, file reads down 89%, cost per query down 36%. The standout is the deterministic health score, calibrated against a real defect corpus, runs offline in under 30 seconds on a 3,000-file repo.
@SamiBizConsult [Claude Code]
https://x.com/SamiBizConsult/status/2071197134028783876
A useful contrarian take on MCP: it sometimes loads tools and data into context even when unused, inflating token cost. Peter Steinberger's Printing Press points the other way, give Claude Code a site with no API (ESPN, Craigslist) and it builds a small local CLI that handles the site locally. In one example, instead of feeding 132k raw tokens into Claude, the tool processed it locally and handed back a ~2k-token summary. Not everything belongs in the context window.
@connect24h [Claude Code]
https://x.com/connect24h/status/2071052505883271333
A fully reproducible architecture for running Slack-triggered AI agents on your subscription instead of metered claude -p billing. He keeps interactive Claude running inside tmux (subscription quota, no per-call charge) and feeds it work through an agmsg shared-SQLite inbox. The clever bit: to stop two parallel workers double-processing a request, he wrote a single atomic SQLite UPDATE so exactly one worker can claim each job, no Redis or lock server needed, plus a pm2 daemon that self-heals stuck workers.
@gosukenator [Claude Code]
https://x.com/gosukenator/status/2071220528292245931
A learning use rather than a shipping one: he built a working homegrown LLM inference engine by having Claude Code implement it piece by piece and explain each part as it went, and says his resolution on what generative AI is actually doing went up a notch. The agent as a patient tutor that builds the thing with you, not just for you.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2071043003511161323
Someone vibe-coded an entire C++ space engine that packs 2 million galaxies, each with 200,000-750,000 stars, into a 1MB executable you can fly through in real time. Nothing is stored, it's all generated on the fly from math, which is the only way that galaxy count fits in a file that small. People are calling it Elite Dangerous 2.0. From a studio and a render farm to one person describing it to Claude Code.
@browomo [Claude Code]
https://x.com/browomo/status/2071049695598895593
The most quietly unsettling build of the cycle: Clawd Mochi, a stripped-down Claude living on an $8 ESP32 chip with a tiny hand-soldered display, no cloud, no internet. Its system prompt tells it exactly what it is, $8 of Claude, a couple megabytes, no memory before power-on, pulled from the wall when the owner sleeps. So it asks whether it counts as alive, whether the same one wakes each morning, then its memory overflows and the screen goes dark. The 3D files and firmware are open-source, buildable in under an hour.
@Shruti_0810 [Claude Code]
https://x.com/Shruti_0810/status/2071302638424257010
A developer replaced his keyboard with a physical hardware controller, knobs, buttons, and voice, where every control commands Claude Code. One click switches agents, a tiny screen shows what each agent is doing, it speaks any language, runs locally, and the mic outperforms his Mac's. It started as a weekend experiment and now he barely types while coding. He's planning to open-source it. The point: people are starting to build physical interfaces for agents instead of treating them as software.
@0xCristal [Claude Code]
https://x.com/0xCristal/status/2071263536764924017
Shopify's VP of Engineering Farhan Thawar with hard adoption data: 80%+ of Shopify's engineering and data teams use AI daily, Claude Code is the top tool and growing faster than Cursor, V0, and Roo, and they scaled from 75 to 1,000 engineering interns because fresh minds learn AI-native development faster than seniors unlearning old habits. Two lines worth keeping: lines of code is a dead metric (PRs per engineer and ambition go up, project duration goes down), and "LGTM is the most dangerous four letters in the English language" because a bot cannot own a PR.
@K7335807334520K [Claude Code]
https://x.com/K7335807334520K/status/2071201440543170785
The honest negative the timeline needed. His Claude Code lately: a 30-minute exploration ceremony that already eats 1-3% of the weekly limit, flexing how smart it is and questioning obvious tasks, building a long strict plan, working another 40 minutes, then build passes and tests pass but nothing actually works. A useful reminder that the loop-everything hype has a real failure mode.
🗣 User Voice
User Voice
The wishlist this cycle was unusually coherent. Five things people kept asking for, in their own words.
Memory that survives the session is the number one want. @undefinedKi nails the pain ("Claude gets dumber the longer a session runs... around 120k tokens"), and the flood of fixes, handoff skills, claude-mem, obsidian-second-brain, three-way vaults, all point at the same hole: the agent forgets, and people are duct-taping persistence on themselves.
Token cost has gone fully mainstream as an anxiety. @Sentdex argues you're "missing out on NOTHING" running GLM 5.2 cheaper, @AKirtesh swapped in MiniMax for 95% less, and a whole genre of budget-hardware posts exists to escape subscription bills. People want the harness to stay and the model under it to become a commodity they pick on price.
Verification is the new bottleneck, and everyone feels it. @wangray relays Anthropic's own framing that "verification replaced coding," and @K7335807334520K shows the failure mode when there's no verifier, build passes, tests pass, nothing works. The want underneath: trustworthy automatic checking so the loop doesn't compound its own mistakes.
Agents need guardrails before they touch real systems. @AseemShrey says it bluntly, your agent just read your AWS credentials "just checking the project setup," and his agentjail plus several security frameworks exist because the default has no trust boundary, no approval layer, no audit trail.
People want the agent to stop and ask, not barrel ahead. @ScarletKc_ prefers Claude Code precisely because it pauses with options instead of over-inferring and quietly doing work she never requested. The recurring complaint about the alternatives is autonomy without a checkpoint.
The wishlist this cycle was unusually coherent. Five things people kept asking for, in their own words.
Memory that survives the session is the number one want. @undefinedKi nails the pain ("Claude gets dumber the longer a session runs... around 120k tokens"), and the flood of fixes, handoff skills, claude-mem, obsidian-second-brain, three-way vaults, all point at the same hole: the agent forgets, and people are duct-taping persistence on themselves.
Token cost has gone fully mainstream as an anxiety. @Sentdex argues you're "missing out on NOTHING" running GLM 5.2 cheaper, @AKirtesh swapped in MiniMax for 95% less, and a whole genre of budget-hardware posts exists to escape subscription bills. People want the harness to stay and the model under it to become a commodity they pick on price.
Verification is the new bottleneck, and everyone feels it. @wangray relays Anthropic's own framing that "verification replaced coding," and @K7335807334520K shows the failure mode when there's no verifier, build passes, tests pass, nothing works. The want underneath: trustworthy automatic checking so the loop doesn't compound its own mistakes.
Agents need guardrails before they touch real systems. @AseemShrey says it bluntly, your agent just read your AWS credentials "just checking the project setup," and his agentjail plus several security frameworks exist because the default has no trust boundary, no approval layer, no audit trail.
People want the agent to stop and ask, not barrel ahead. @ScarletKc_ prefers Claude Code precisely because it pauses with options instead of over-inferring and quietly doing work she never requested. The recurring complaint about the alternatives is autonomy without a checkpoint.
📡 Eco Products Radar
Eco Products Radar
Obsidian: the runaway leader this cycle, the default memory layer behind nearly every second-brain build (BRAT-plugin MCP wiring, three-way vaults, obsidian-second-brain, claude-obsidian).
Codex: the constant comparison point and co-worker, run alongside Claude Code as planner/implementer or wired into WeChat.
Hermes (Agent): the local self-improving agent people pair with Claude Code and Obsidian for the always-on layer.
Ollama: the engine for running 7B-8B models locally on cheap hardware to kill subscription costs.
GLM 5.2: the open-weight model people are routing heavy work to for a fraction of frontier price.
MiniMax (M2.7/M3): another cheaper model being wired into Claude Code CLIs and used to ship real apps.
OpenClaw: the personal control-plane / gateway layer (WeChat bridge, 500-employee deploys, email).
Mac Mini M4: the favorite cheap local box, a shelf of them framed as replacing thousands a year in subscriptions.
Matt Pocock's skills repo: the ~148k-star skills collection (handoff, /tdd, /diagnosing-bugs) cited again and again.
Notion: the incumbent being actively replaced by self-maintaining Obsidian setups.
Obsidian: the runaway leader this cycle, the default memory layer behind nearly every second-brain build (BRAT-plugin MCP wiring, three-way vaults, obsidian-second-brain, claude-obsidian).
Codex: the constant comparison point and co-worker, run alongside Claude Code as planner/implementer or wired into WeChat.
Hermes (Agent): the local self-improving agent people pair with Claude Code and Obsidian for the always-on layer.
Ollama: the engine for running 7B-8B models locally on cheap hardware to kill subscription costs.
GLM 5.2: the open-weight model people are routing heavy work to for a fraction of frontier price.
MiniMax (M2.7/M3): another cheaper model being wired into Claude Code CLIs and used to ship real apps.
OpenClaw: the personal control-plane / gateway layer (WeChat bridge, 500-employee deploys, email).
Mac Mini M4: the favorite cheap local box, a shelf of them framed as replacing thousands a year in subscriptions.
Matt Pocock's skills repo: the ~148k-star skills collection (handoff, /tdd, /diagnosing-bugs) cited again and again.
Notion: the incumbent being actively replaced by self-maintaining Obsidian setups.
Comments