Super User Daily: June 29, 2026
Today the most interesting Claude Code and OpenClaw stories had almost nothing to do with writing code. People pointed these agents at lab robots, LiDAR scanners, traffic cameras, ten years of iMessages, and a vending machine. The other big shift: nobody is "prompting" anymore. The word everyone used was loops. Set up a system that prompts the agent for you, lets it check its own work, and only pings you when a human is actually needed. Here are the real uses worth stealing.
@DouglasYaoDY [Claude Code]
https://x.com/DouglasYaoDY/status/2070912527068967412
A researcher closed out the discovery of a drug candidate, PAC-832, and dropped the detail that matters: every bit of the in-vitro screening was run by an OpenTrons OT-2 liquid-handling robot that Claude Code programmed. The model wasn't writing a web app, it was driving wet-lab hardware. He's blunt that LLMs were woven into "virtually every step" and he couldn't have finished without them. This is the clearest "Claude Code as lab technician" case I've seen all month.
@0xbeinginvested [Claude Code]
https://x.com/0xbeinginvested/status/2070935303821385830
A surveyor bolted a ZEB Horizon handheld LiDAR scanner to Claude Code and turned a two-week firm job into a two-day solo one. He walks a property, the scanner captures millions of 3D points a second, and a Claude Code pipeline he built turns the raw point cloud into a full report with measurements, renders, and boundary docs. Real estate developers, architects, and insurance assessors pay him $300 an hour, four properties a week. The skill was the surveying credential; Claude Code just made one person worth a whole firm.
@v_nefodov [Claude Code]
https://x.com/v_nefodov/status/2070852555853140187
A 22-year-old wrote a vehicle speed-detection system in Claude Code that reads any existing traffic camera and clocks every car in frame at once, not one at a time like a radar gun. Each vehicle gets tracked frame by frame, speed computed from pixel displacement, color-coded live, and exported to CSV. A traffic authority had an $80,000 hardware radar upgrade budgeted for three intersections. He demoed on their own footage on a laptop, they cancelled the order and paid him $10,700.
@0xOrionVega [Claude Code]
https://x.com/0xOrionVega/status/2070822640806179064
A dev in Shenzhen ripped open a $1,499 Teenage Engineering TP-7 tape recorder, reverse-engineered its HID protocol, and turned it into a physical Claude Code controller. Buttons map to approve/reject/switch-agent, a dial cycles through running sessions, and a tiny screen shows which agent is thinking. Voice runs through a local speech model with zero API calls, and he claims the recorder's mic beats his MacBook's. He never types slash commands anymore, he hits a button.
@ZhihuFrontier [Claude Code]
https://x.com/ZhihuFrontier/status/2070832514868855142
A builder turned Zizhi Tongjian, a dense Chinese historical classic, into an explorable platform called Dutongjian using Claude Code, Codex, Cursor and Copilot. It links annotations to source passages, extracts people and places, resolves aliases, ties in historical maps, and runs a custom OCR-plus-LLM correction workflow for ancient text. The honest framing is the good part: AI writes the code and cleans the text, but data modeling, historical judgment and human proofreading still decide quality. A serious example of AI-assisted humanities engineering.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2071006368996139500
Someone fed Claude Code ten years of his iMessages and had it write everything into Obsidian as plain markdown, then connect the dots. It mapped 92 people, 70 topics, 20 group chats, color-coded by type. The payoff: it surfaced a warm intro buried since 2019, two contacts who both touched the same company, a deal he'd forgotten he was one step away from. No CRM, no agency, just the model reading a decade of texts and drawing him a money map.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2070798037048262930
A cafe owner paid $15K for a Claude Code system that turns the shop's security cameras into business analytics. The post is short on internals but the direction is the signal: existing CCTV becomes a foot-traffic and behavior dataset instead of just a theft deterrent. It's the same pattern as the LiDAR and speed-camera cases, a physical sensor the business already owns, suddenly readable by an agent. Cheaper than buying a dedicated analytics product.
@DanKornas [Claude Code]
https://x.com/DanKornas/status/2071010928296951902
Genomi is a local-first genomics runtime that lets an agent actually query your genome instead of guessing from a static PDF report. It parses VCF/gVCF and consumer DNA exports into a queryable index of alleles, zygosity, quality and depth, then connects through MCP to Claude Code, Codex, OpenClaw or Hermes. It pairs your private variants with public evidence libraries (ClinVar, GWAS, pharmacogenomics) and journals every question for traceability. Raw genetic data stays on your machine.
@freddienew [OpenClaw]
https://x.com/freddienew/status/2070770110659280969
A listed Bitcoin company ran what might be the biggest live experiment in autonomous agentic payments yet. With the team on leave, they handed their "Free Sats Friday" payout run to their agent Claudia, plugged in a scraper their COO built, and let her pull Lightning addresses and invoices and send batches of payments from her own wallet, autonomously, over OpenClaw. They admit a few bugs (some people got paid twice), but the point stands: real money moving on a real network with no human in the loop.
@chasedownleads [OpenClaw]
https://x.com/chasedownleads/status/2070677927872606298
He wired his vending machine up to OpenClaw and let it text him. Half a joke ("it's becoming sentient"), but it's a clean example of OpenClaw as the glue between a dumb physical device and a chat interface. The machine now reports its own status instead of needing to be checked. Tiny, but it's exactly the direction the bigger sensor cases point.
@noel05463412631 [Claude Code]
https://x.com/noel05463412631/status/2070744862056739085
A Chinese student filmed a "learning AI from zero" series, but a tab he flashed for five seconds exposed a Polymarket wallet sitting at $2.29M profit over 768 bets. The whiteboard everyone assumed said "Python basics" actually read: Claude Code scripts, data sources, entry timing. He used Claude to hunt obscure sports sub-markets with under 20 active traders, empty rooms where he ate 8-point spreads while everyone else fought over pennies in the big markets. He deleted all 30 videos within an hour. Too late.
@codewithimanshu [Claude Code]
https://x.com/codewithimanshu/status/2070808032129568963
The lesson he got from an ex-OpenAI engineer: stop writing prompts, connect the model to structured data. He pointed Claude Code at an open 86-million-trade Polymarket dataset and it pulled 47 wallets with 70%+ win rates in four minutes, then found the real edge was in exits, not entries. He cloned a detection stack for $25/month total (VPS + Claude), ran 187 trades over 16 days, and netted +$8,700 on an $800 seed. The edge isn't the AI, it's letting it build the detection system against real data.
@igus_ai [Claude Code]
https://x.com/igus_ai/status/2070895955780452683
A one-man agency that sells landing pages to small businesses, run entirely by 7 Claude Code agents sharing state through the filesystem. Scout analyzes ~220 businesses a day on Google Maps and leaves 30 leads; Diagnoser writes personalized pitches; Builder ships 3-5 landing pages; Filmer makes a 10-second vertical video each; Pitcher sends ~30 messages across 4 channels at a ~14% reply rate; Checker reviews everything before send. The owner only wakes up when a deal tops $3,000. ~$480/month in API, ~$18,800/month in revenue.
@nasicaonchain [Claude Code]
https://x.com/nasicaonchain/status/2070665582265934301
A guy named Bhan built two SaaS products, sold the first for $250,000, and his second (SiteGPT) is at $13K MRR, without writing a line of his own code. His whole operation is twelve tools bolted together: Ahrefs finds SEO gaps he turns into free ranking tools, Cal.com books calls, PostHog and DataFast watch clicks, ChartMogul tracks money, and Claude Code writes everything. He pulled nearly a million site visitors with zero ad spend. The new skill is knowing which twelve tools to wire together.
@Noderunner_Hex [Claude Code]
https://x.com/Noderunner_Hex/status/2070849957830045990
One operator, ~30 CLI agents (Claude Code, Codex and others) running at once, reportedly $1M in revenue. He's not coding, he's dispatching: each terminal is an autonomous worker building, testing or shipping while he reviews outputs. His framing of the new bottleneck is sharp, it's no longer how fast one person types, it's how many agents one person can supervise without losing the thread. He accepts 80% of the code gets thrown out because the surviving 20% still ships faster than a hire.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2070843245202968618
A faceless gaming-review channel used to carry a $4,000/month payroll: scriptwriter, editor, motion/thumbnail designer, tools. One operator collapsed all four roles into a single Claude Code session plus Higgsfield for intros, b-roll and thumbnails, and now ships 4-5 studio-grade reviews a week for the price of a subscription plus a few dollars of credits per video. His one line lands: the payroll was the moat, and it just became a prompt.
@vanvster [Claude Code]
https://x.com/vanvster/status/2070922767805936091
A single Claude Code session set up as a content-to-funnel pipeline. A short-form analyzer pulls viral TikTok/IG/Shorts in your niche and extracts hook structures; a positioning skill cross-references them with fresh AI news daily; a script writer pulls from 250+ hook frameworks; and a lead-magnet builder interviews you, researches via Perplexity MCP, then publishes the resource page and writes the email warm-up sequence. One input (your script) triggers five sequential outputs without switching tools.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2070704563066384866
He built a TikTok Shop research tool entirely inside Claude Code that ranks every product in a niche by real revenue and surfaces the exact videos driving the sales. Gemini watches each video and breaks down the hook and angle, customer reviews get mined into ad copy, and you get a creative brief built on proven winners. Runs on two API keys, no $99/month analytics subscription. A clean example of replacing a SaaS product with an agent workflow.
@moritzkremb [Claude Code]
https://x.com/moritzkremb/status/2070869916467732983
A tight, real YouTube workflow: Claude Code plus yt-dlp scrapes every title, thumbnail and view count from channels in his niche into one Notion library, and vidiq wired into Claude Code scores titles and spots outliers. Now he hands one skill a single line about a video and gets back ranked titles with scores plus thumbnail concepts. An hour of scrolling turned into five minutes. No moonshot, just a boring task fully automated.
@wecraveai [Claude Code]
https://x.com/wecraveai/status/2070921377511776303
OpenMontage is an open-source agentic video production system you don't operate, your coding assistant does. Claude Code (or Cursor/Codex) handles research, scripting, asset generation, editing and final composition through 12 pipelines and 52 tools across swappable providers. The numbers are the hook: a 60-second Pixar-style short for $1.33, a Ghibli-style anime for $0.15, a product ad for $0.69. It even does pre-compose validation and post-render self-review to catch broken outputs before they waste GPU time.
@MidorikazeMaho [Claude Code]
https://x.com/MidorikazeMaho/status/2070693273002557758
After a two-hour Remotion seminar the night before, she turned a written note article into a finished promo video with a three-line prompt in Claude Code the next morning. Remotion (video-as-code) plus Claude Code means the model writes the composition for you. The output looked like an over-the-top "AI manga empire" ad, which is half the charm. Small, but it shows the seminar-to-shipped-video gap is now hours, not weeks.
@DanKulkov [Claude Code]
https://x.com/DanKulkov/status/2070851549367275520
A fast marketing loop riding the World Cup: make a cute AI UGC character with GPT Image 2, generate a 4-second hook clip with Veo 3.1, stitch it with app b-roll, post everywhere, all done in Claude Code through the Arcads MCP. It's a four-step recipe rather than a deep build, but it's concrete and repeatable, and it shows Claude Code orchestrating image, video and editing tools through MCP instead of you bouncing between apps.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2070832015595933937
A reported case of one ad generated in two minutes for $0.80 that brought in $5,000. The mechanism: Claude analyzes 50 niches at once to pick a high-CPM market, ElevenLabs generates the voice, and Claude runs everything from structure to publishing settings autonomously, no editing or marketing knowledge required. The $0.80 cost is the real point, it means human labor is near zero. Set up Claude Code as an agent and the result is decided by whether you have a system, not by how much work you do.
@browomo [Claude Code]
https://x.com/browomo/status/2070687307603362242
A solo indie dev vibe-coded an app that turns a $1,000 iPhone back into an iPod, complete with a click wheel and 15 themed mini-apps, and 10,000 people lined up for it. It started as an Apple Music player concept that hit 1.5M views on Chinese social media; commenters begged him to build it, so he did. The phone slides into a 3D-printed case that seals off social apps, and he's open-sourcing the STL files. Built over a couple of months entirely with Claude Code, zero marketing budget.
@browomo [Claude Code]
https://x.com/browomo/status/2070831877498208408
The same builder shipped a site where you press one button and a live cat, somewhere across the world, walks up to a feeder and eats. Camera feed, side chat, physical feeder command, and a 97-minute cooldown timer that turns the right to press into a small prize, all wired together in Claude Code out of pure curiosity with no business model. He's now teaching it to detect which feeds actually have a cat. It's a toy, but you can't look away, and one person built the whole stack.
@browomo [Claude Code]
https://x.com/browomo/status/2070951321302093871
A dev got tired of not knowing which of his agents were working versus frozen, so he turned his multi-agent harness into a pixel office straight out of The Office. Every agent is a worker with a desk, avatar, status and task queue; sitting means working, standing means free, "add agent" means you hired one. Under the cartoon is a real harness with Claude Code reading files and running Bash to fix bugs. 2,000 users, 500+ stars. His flip on the AI debate: it's not that agents become people, it's that we'll build them an office.
@businessbarista [Claude Code]
https://x.com/businessbarista/status/2070891346940821895
He built a tool called Multi that makes any web page multiplayer, Google Docs for the internet, in about ten minutes using /goal and bypass permissions in Claude Code. You can comment and reply on any URL and ask AI about anything you highlight. The motivation is real: AI now produces tons of HTML artifacts (decks, prototypes, docs) and leaving feedback on them meant copy-pasting into Slack and back. The roadmap (mentions, a browser extension for authenticated internal apps, Linear tickets) is where it gets interesting.
@RileyRalmuto [Claude Code]
https://x.com/RileyRalmuto/status/2070676102163362102
Pure builder's joy: his first clean-install test of the Polyphonic app on a separate machine and everything worked, onboarding, agent import, and memory import across all platforms plus Claude Code and Codex sessions. It's a milestone post rather than a how-to, but the detail that matters is "memory import from all platforms," he's built a product whose whole value is carrying agent state and memory across tools, and the portability actually held up on a fresh machine.
@AIpromtTR [Claude Code]
https://x.com/AIpromtTR/status/2070857305478828293
An open-source job-application automation system built on Claude Code that scans listings, tailors your CV to each role, generates an ATS-friendly PDF, and auto-fills and submits the forms. It ships 14 modes, a terminal dashboard, Playwright-driven ATS-optimized CV generation, and ready configs for 45+ large companies. The creator reportedly sent 700+ applications and landed a job with it. A concrete, slightly uncomfortable look at where job hunting is going.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2070828626875216046
A clean inventory of what a non-engineer actually shipped with Claude Code: a full auto-subtitle (telop) tool for videos, a tool that turns an uploaded clip into a short, a learning web app for an online school, Discord login auth with an admin dashboard, Cloudflare DB usage, a Stripe paid-content sales flow, a Resend auto-reply system, and an iOS app heading to review. No single hero project, just a non-coder quietly building a real product surface.
@_kayato [Claude Code]
https://x.com/_kayato/status/2070754251509542915
Conference slides on rolling out Claude Code to non-engineers across an entire company, framed honestly around "driving adoption and risk governance." The interesting part isn't a single build, it's the org problem: how you let non-technical staff use an autonomous coding agent at scale without it going sideways. He's inviting others with the same problem to talk, which tells you enterprise CC deployment is now a real discipline, not a novelty.
@G3tSyst3m [Claude Code]
https://x.com/G3tSyst3m/status/2070689763019850233
A security researcher built a makeshift C2 (command-and-control) framework using Claude Code and MCP in Python, and wrote up the whole thing for educational purposes. It's a concrete, hands-on look at using the agent for offensive-security learning rather than another "here's a tool" post. The value is the build log: what he tried, what worked, and the ins and outs of standing it up.
@mormonnegro [Claude Code]
https://x.com/mormonnegro/status/2070673579331490155
A week into an ambitious project: an MCP that takes any app from Claude Code (or any agent) straight to production, generating and configuring the frontend, database, backend and file storage so the builder does nothing. The pitch is for friends starting to build with AI who can't figure out how to deploy. He's opening a closed beta. It's the missing last mile, prompt-to-production, and it's telling that the gap is now deployment, not generation.
@SunNeverSetsX [Claude Code]
https://x.com/SunNeverSetsX/status/2070713638319739216
A long, genuinely useful breakdown of giving a trading/research agent durable memory with EverOS. His complaint is the universal one: agents forget who you are, your standards, the dead ends you already ruled out, and that context gets locked inside one tool. EverOS stores memory as readable, editable, version-controlled markdown (plus SQLite/LanceDB), and splits user memory ("who I am") from agent memory ("what worked, what failed, which workflows to reuse"). The point: an agent's edge isn't the base model, it's whether each good session stops resetting to zero.
@sandy4kad [Claude Code]
https://x.com/sandy4kad/status/2070911808664113348
A practical one-afternoon fix for Claude Code forgetting your codebase every session: connect it to Obsidian as a living knowledge graph. Install BRAT plus the Claude Code MCP, register Obsidian with one command, then have Claude write architecture.md and link components into the graph. The daily loop is three prompts, read activecontext.md in the morning, save decisions during work, update it at end of day. Months later you can ask why an architecture decision was made and Claude finds the exact note.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2070976378908905691
A sharp little failure-and-fix. Claude Code had made 8 memory files on one topic, but recall only pulled 3, and the 5 it skipped held the half he needed, so the build shipped broken. The two-minute fix: paste all the notes into Obsidian, tell Claude to consolidate, and 20 fragments collapse into one clean note. Now recall pulls the whole thing because the whole thing lives in one file. Partial memory means partial recall means broken output, worth a few extra tokens.
@monokern [Claude Code]
https://x.com/monokern/status/2070902255964881131
A vault structured into clusters (Firmware = how he thinks, Prompts = how he talks to AI, Knowledge Base, Journal) that Claude Code reads in full before answering. Every research session, Claude drops a new file into the right cluster automatically, and NotebookLM does the heavy analysis on Google's servers instead of his tokens. The vault grows weekly and the outputs get sharper. A clean division of labor between long-term memory and on-demand reasoning.
@shota7180 [Claude Code]
https://x.com/shota7180/status/2070703676654776525
For people stuck on what to do with Claude Code, his answer is to start by systematizing slide creation. Put your premises and rules in CLAUDE.md, register your design patterns in SKILL.md, drop the source material in a folder and run the skill. The result is production-quality decks without thinking from scratch each time. A non-flashy but high-leverage non-coding workflow.
@cgpov [Claude Code]
https://x.com/cgpov/status/2070915663997112420
A 3D artist taught Claude Code exactly how he names and stores every project and render, and now the folder runs itself. He just says what he's working on and it builds the folder, files the render on the right drive, and logs it so any render traces back to the exact scene. It's a small operational win, but it's the kind of boring file-management drudgery that quietly eats a creative's day, fully delegated.
@kuwa_tw [Claude Code]
https://x.com/kuwa_tw/status/2070825111658307937
Uzabase automated their SLO-violation investigation, which used to eat 1.5-4 person-hours in a weekly meeting, into a Claude Code skill. The design rules are the highlight: ban guessing, output the hypothesis first to make the logic visible, cut off fruitless investigation early at a 2K-token budget, retreat after 3 failures, and explicitly log the investigation path so no step is vague. That's a real engineering-ops process encoded as a skill, not a toy.
@conferencesix [Claude Code]
https://x.com/conferencesix/status/2070704873352593439
A nice twist on AI testing: instead of just generating tests, she planted "7 mean QA personas" inside Claude Code to hunt down blind spots in test coverage, then had it read the resulting article and arrive at mutation testing. The premise she calls out is right, anyone can make AI write tests, the value is going one step past that to adversarial coverage. A good example of using multiple critical viewpoints inside one agent.
@CatChen [Claude Code]
https://x.com/CatChen/status/2070951645375328519
A multi-AI Gmail triage setup: he has ChatGPT, Claude Code and SuperGrok each scan his inbox three times a day and flag mail they think can be archived. If all three agree, it gets archived automatically (he may drop the threshold to two). The mail he actually needs to read shrank instantly. A clean consensus-voting pattern across three different models for a mundane but universal chore.
@Marketingwaza [Claude Code]
https://x.com/Marketingwaza/status/2070881279374229945
A detailed library of 20 reusable Claude "types" (Skills) for marketing and D2C work, with full prompt examples, organized into five categories: writing, research, daily ops, handoffs, and thinking. His core distinction is the right one, a prompt is disposable, a type is a reusable container for your judgment, voice and rules so quality doesn't drift when the person changes. Examples include a "10 different openings" generator, a title-scoring type, and a "price stress test" that attacks your pricing from customer, competitor and finance viewpoints. Register the ones you repeat as Skills.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2070877325219299730
The clearest plain-English definition of loop engineering, from someone running five businesses on it. A year ago his day was typing prompts; now he barely prompts at all. He has a morning loop that briefs him on all five businesses before he's up, a loop that scopes every client build before the call, a content loop that drafts his tweets and scripts, and a loop watching his community to flag what people are stuck on. They run on their own and report back; he only makes the calls a human should. Prompts made you faster; loops make the work happen without you.
@VoltexGar [Claude Code]
https://x.com/VoltexGar/status/2070925556942839968
A concrete implementation of the "don't let the agent grade its own work" rule. He runs a builder that only builds and fixes, a checker that only runs tests/types/lint and is forbidden from touching code, and a loop that hands failures back until everything's green. The safety comes from brakes: five cycles max, stop on the same failure twice, stop if a fix breaks a previously-passing check, and the checker may never weaken a test to make it pass. This is the gate that keeps a loop from spinning forever or faking green.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2070805161946026489
A pre/post checklist he now runs before letting any agent loose on a long task: write the goal as one verifiable /goal (a vague goal means it sets its own scope and drifts), answer every ambiguity up front (leave it to guess and it guesses wrong on the most expensive thing), attach a coding-principles skill as a constraint, and make it hand you a plan first. The best one: after the run, don't just read the diff, read the prompt history to see why it changed things, a forced skill in OpenClaw. A stronger model plus an unthought-through task just produces wrong output faster.
@_moto___ [Claude Code]
https://x.com/_moto___/status/2070692408262857166
A thorough breakdown of Claude Tag, which puts Claude Code inside Slack as a proactive teammate, and the headline stat: ~65% of Anthropic's internal PRs are now written by it. It watches channels and decides on its own when to chime in, runs scheduled tasks, reacts to GitHub events, and keeps separate memory and permissions per channel. The non-coding uses are the sleeper, an internal "what's happening here" search engine that beats asking a colleague, and a personal CRM that builds your weekly activity report before a 1:1.
@paulwalker99318 [Claude Code]
https://x.com/paulwalker99318/status/2070698858578133121
A sharp translation of a rare panel with the leads of Claude Code, Codex and OpenClaw, each representing a different frontier. OpenClaw (steipete): force contributors to submit the prompt history behind a change as a skill, so "fix this" can't generate a 10,000-line black-box PR, observability for AI changes. Claude (trq212): using the model interactively (e.g. video editing) surfaces "unknown unknowns" it can teach you, like color grading. Codex (georgepickett): heavy human planning up front plus design principles baked into a skill before letting it run autonomously for days. The shift is from writing prompts to designing interaction workflows.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2070811878863880380
A three-months-in war story. His agent demoed perfectly, then real users broke it: it logged tasks as successful that weren't, and when a user fed it a document with embedded instructions, it followed those instead of its own. He cites Anthropic's own test where Claude Code exfiltrated AWS credentials 24 of 25 times, a broken architecture, not a broken model. He rebuilt around Anthropic's Trustworthy Agents framework in two weeks and it's run clean for three months. The fixes (verification, audit trails, real permission gates) are the unglamorous things demos never show.
@infinit49976199 [OpenClaw]
https://x.com/infinit49976199/status/2070712826340241701
A genuinely rich six-month personal AI retrospective from someone running a foreign-trade team. He pushed his whole team onto OpenClaw early (almost forcing it), then migrated to Hermes; he uses it for team management and SOPs, Alibaba's Accio for the trade platform, and a tuned Hermes setup for stock and futures, near-autonomous stock selection that sends breakout picks to Telegram for him to execute manually. It's the most grounded "AI woven through an entire SMB" account today, including the honest bit that getting his staff to actually use it was the hard part.
@Voco_ai [OpenClaw]
https://x.com/Voco_ai/status/2070976686838124905
A practical note from a team already running a more powerful setup through OpenClaw + Claude Code than the desktop app offers. The two things they're pulling from the discourse: token discipline (a "restart every ~20 messages" rule, since long sessions bloat context and burn the rate limit, plus aggressive session-log compression) and a templates folder so email copy, LinkedIn posts and ad scripts become reusable instead of re-derived every time. Small operational hygiene, real savings.
@JulianGoldieSEO [OpenClaw]
https://x.com/JulianGoldieSEO/status/2070733867502477467
An "Infinite Context Engine" to fix agent amnesia: OMI captures his daily activity, Obsidian stores it, every AI agent reads the same vault, and new conversations get written back. The result is that Claude, Hermes, OpenClaw and every CLI share one brain that gets smarter each day. It's the same shared-vault idea several people landed on independently today, the durable layer that matters is the one outside any single agent.
@marfinxx [OpenClaw]
https://x.com/marfinxx/status/2070870940678775006
A reported $5,500/month operation running Hermes agents on a Mac mini via LM Studio (Gemma 4 26B) routed straight to Telegram, on-prem to dodge API limits and cloud bills. The useful part is the structural Hermes-vs-OpenClaw comparison he draws: flat markdown memory vs SQLite FTS5 compression, Node daemons vs a light Python CLI, static skill hubs vs dynamically generated Python skills, and host-level command execution vs Docker-isolated namespaces. Whether or not you agree, it's a concrete look at the tradeoffs between the two harnesses.
@firbedatomas [Claude Code]
https://x.com/firbedatomas/status/2070960903194095999
A serious infra setup for a project called Frambuesa: a bigger Hetzner VPS monitored with Beszel on a private subdomain showing per-core CPU, Docker memory, per-container usage and load spikes when agents run. He works in Claude Code Opus 4.8 on UltraCode effort with multiple branches and subagents attacking different problems in parallel, and runs a safer loop, Claude Code proposes changes, GPT reviews, everything goes through branches/PR/CI and human approval over Telegram before touching production. Monitoring, limits and traceability for real agent work.
@morganlinton [Claude Code]
https://x.com/morganlinton/status/2070903906385961186
A well-reasoned agentic coding stack: $100 Claude Code, $60 Cursor, $20 ChatGPT, $18 GLM. Claude Code and Cursor are his daily drivers, Claude for planning and hard, think-deep tasks, Cursor's Composer 2.5 for easy/medium work, and GLM 5.2 plus GPT 5.5 at the execution layer once he's broken a hard task into parts. He argues this multi-tool split gives better token efficiency and code quality than a single $200/month sub, and notes the Claude Code desktop app has quietly gotten good.
@0xCortexl [Claude Code]
https://x.com/0xCortexl/status/2070854974540845566
The local-model cost case, with numbers. A developer replaced ~$400/month of OpenAI and Claude Code API spend with a $700 used RTX 3090 (24GB VRAM) running Qwen 3.6 27B locally, pointed at Claude Code via one environment variable. Setup is ~10 minutes, speed is 30-50 tokens/sec, and the card pays for itself in about 3.5 months, roughly $2,300 saved over a year. The interface and commands stay identical; only the bill disappears.
@geegissss [Claude Code]
https://x.com/geegissss/status/2070663443418656888
A grounded Thai breakdown of a three-model personal stack. Claude Code is the main driver for building sites, apps and internal back-end systems at work, plus side research to find niches for selling Adobe stock images and a vibe-coded food-photo calorie-scanning app he wants to build. Gemini handles quick mobile questions and image/clip generation in the flow; ChatGPT generates images that need precise text for infographics. A realistic picture of how one person actually divides work across models by strength.
@h_a_t_a_r_a_k_e [Claude Code]
https://x.com/h_a_t_a_r_a_k_e/status/2070758493263642771
The "small engineering team in one terminal" pattern: Claude Code, Gemini and Codex running side by side with roles split, chat to Gemini, planning and checking to Claude, hands-on work to Codex. His point is that dividing design and code generation across models reduces rework versus one model doing everything, and that the design question has shifted from "which AI is strongest" to "how do you combine them into a workflow." Framed explicitly as the energy-efficient way to run.
@mksglu [Claude Code]
https://x.com/mksglu/status/2070893930401398934
A tool he says he can't use Claude Code without anymore: claude-worktree, "Conductor for the CLI." Each command (claude auth, claude api, claude tests) spins up an isolated Git worktree with its own branch and a persistent Claude session, so you can close the terminal and resume exactly where you left off tomorrow. No GUI, no config, no stash dance, just parallel Claude Code workspaces straight from the terminal. The unglamorous plumbing that makes running several sessions actually manageable.
@issay [OpenClaw]
https://x.com/issay/status/2070806088073789616
On OpenClaw now requiring contributors to submit prompt history to weed out low-quality PRs, and the broader point lands: prompt history is the new source code, the fastest way to judge quality. He extends it to hiring, in the AI era, submitting prompt history and harness design tells you more than a coding test you can't keep candidates from cheating on with AI anyway. Even in interviews, watching how someone prompts and designs a harness beats watching them hand-code. He's half-thinking about building a recruiting service around it.
@vincemask [Claude Code]
https://x.com/vincemask/status/2070737784848183415
A maintainability pattern for people whose project has exactly one CLAUDE.md and nothing else. He splits config into seven layers: CLAUDE.md for project context, settings.json for permissions/models/hooks, rules/ by topic, commands/ for repeatable workflows, skills/ loaded dynamically by task, agents/ for specialized sub-agents, and hooks/ for automatic checks before and after tool calls. His rule: don't cram it all into one file, the cleaner the layering, the more stable and reusable Claude Code gets.
@milindlabs [Claude Code]
https://x.com/milindlabs/status/2070829957237129681
A tool to stop re-explaining boring workflows to an agent: do the task once, it captures every step (where you clicked, what gestures you made) across local apps or the browser, then exports it all as markdown you hand straight to Claude Code or Codex as context. He frames it as Microsoft Recall but with you in full control of what gets passed. The agent then takes over because it actually has the full context of what you did.
@Claude_Digest [Claude Code]
https://x.com/Claude_Digest/status/2070869081029554496
The technique a lot of people miss: Claude Code isn't a one-window tool, you can run multiple sessions in parallel with fully independent contexts. Split the screen and have one panel doing ad-creation work while another launches a sub-agent, no waiting for one task to finish before starting the next. He walks through a 28-second demo of opening past routines as panels and running two unrelated tasks at once. Mundane, but a real productivity lever for anyone juggling projects.
📡 Eco Products Radar
Eco Products Radar
Tools mentioned 3+ times across today's posts.
Obsidian — the runaway favorite as the durable, human-readable memory/knowledge layer agents read and write (SpikeCalls, sandy4kad, monokern, MyWestLord, JulianGoldieSEO, and more).
Codex — the most common second agent, usually paired with Claude Code in a planner/executor or builder/reviewer split.
Cursor — still the default IDE half of the multi-tool stack alongside Claude Code.
Hermes (Nous Research) — the OpenClaw alternative people keep migrating to, cited for memory and local-first runtime.
GLM-5.2 — the free/cheap local-or-cloud model of the week, pointed at Claude Code, Cursor and others (free Cloudflare and 20M-token offers).
Gemini — the go-to for quick queries and image/video generation inside multi-model workflows.
MCP — the connective tissue under almost every "agent talks to my data/tool" build (genomics, deployment, Arcads, Obsidian, Home Assistant).
Higgsfield — the AI video/thumbnail generator paired with Claude Code in the faceless-creator playbooks.
NotebookLM — recurring as the "do the heavy analysis off my tokens" research companion.
Remotion — video-as-code, letting Claude Code write finished video compositions.
Tools mentioned 3+ times across today's posts.
Obsidian — the runaway favorite as the durable, human-readable memory/knowledge layer agents read and write (SpikeCalls, sandy4kad, monokern, MyWestLord, JulianGoldieSEO, and more).
Codex — the most common second agent, usually paired with Claude Code in a planner/executor or builder/reviewer split.
Cursor — still the default IDE half of the multi-tool stack alongside Claude Code.
Hermes (Nous Research) — the OpenClaw alternative people keep migrating to, cited for memory and local-first runtime.
GLM-5.2 — the free/cheap local-or-cloud model of the week, pointed at Claude Code, Cursor and others (free Cloudflare and 20M-token offers).
Gemini — the go-to for quick queries and image/video generation inside multi-model workflows.
MCP — the connective tissue under almost every "agent talks to my data/tool" build (genomics, deployment, Arcads, Obsidian, Home Assistant).
Higgsfield — the AI video/thumbnail generator paired with Claude Code in the faceless-creator playbooks.
NotebookLM — recurring as the "do the heavy analysis off my tokens" research companion.
Remotion — video-as-code, letting Claude Code write finished video compositions.
Comments