Super User Daily: June 26, 2026
The center of gravity keeps drifting away from "write me some code." Yesterday's best stories were a journalist translating a 500-page book and chewing through 4.2 million words of council minutes, a 19-year-old replacing a $180k traffic-analytics vendor with a drone and Python, an agency cutting its monthly AI bill from 10,000 to 3,000 euros by swapping the model under Claude Code, and a guy who flashed his dead crypto mining rig into a private-inference business clearing $5k a month. Coding is now the smallest part of what people do with these tools. The other loud theme: where the agent runs. Laptops close, SSH drops, and long jobs die — so everyone is moving Claude Code onto VPSes, DGX Sparks, Mac Minis, and even Raspberry Pi clusters, and bolting on shared memory files so Claude and Codex stop re-learning the same lessons.
@DrewPavlou [Claude Code]
https://x.com/DrewPavlou/status/2069645947710386306
A journalist turned Claude Code into a one-person research institute and laid out the receipts. He translated a 500-page Italian book in 20 minutes for about $5 because no English version existed and the publisher had stalled for three years. He had it download and analyze 4.2 million words of Brisbane City Council debate to detect rhetorical and ideological shifts, crunch five years and ~10,000 of his own financial transactions, and transcribe hundreds of extremist-preacher lectures — walking him from never having opened Terminal to renting an NVIDIA H100 for an hour at $3. His point: this is the democratization of work that used to require billionaires and warehouses of researchers.
@v_nefodov [Claude Code]
https://x.com/v_nefodov/status/2069728980807127152
A 19-year-old student built a $50k startup with nothing but Claude Code, a consumer drone, and Python in three weeks. The system flies over a highway, runs every vehicle through a CV pipeline Claude helped write, draws bounding boxes, and measures speed with a bird's-eye perspective transform — green at 60 km/h, yellow at 100. The on-screen constants tell the story: MIN_TRACK_FRAMES=8, averaging speed across 25 frames, discarding physically impossible readings. His first client, a transport authority, had been paying a legacy vendor $180k/year for slower weekly reports; he delivered same-day for a $130/month running cost. He didn't know computer vision — he knew how to talk to Claude Code.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2069824689170444635
The most useful cost-optimization thread of the day, from ChatSEO going from 10,000 to 3,000 euros/month in AI spend with the same or better quality. The move that mattered: switching from Sonnet 4.6 to DeepSeek V4 — but he stresses benchmarks lied, DeepSeek scored slightly lower on leaderboards yet was better on their actual SEO data-interpretation tasks. Use Vercel AI Gateway to swap models without touching code; if you're solo, use the Claude Code CLI subscription (90 euros) not the API, because CLI usage worth 500+ euros in API tokens isn't viable for Anthropic long-term, so milk it now. Trim verbose self-written prompts for another 30-40%, and never cheap out on the main model — they tried Haiku on some agents and churn exploded.
@Xudong07452910 [Claude Code]
https://x.com/Xudong07452910/status/2069636634241343964
Anthropic's own engineering lead Fiona Fung said something this user calls one of the most honest observations of the AI-coding era: Claude Code is making engineers lonelier. Not as metaphor — an internal Anthropic survey found that as engineers collaborate more with agents and less with colleagues, human-to-human technical exchange is dropping. The team responded by deliberately scheduling coding lunches, hackathons, and shared build sessions. The reframe is sharp: the old question was "will AI replace programmers," the real one is "will it change how programmers collaborate" — and collaboration now has to be designed on purpose, because it no longer happens by accident.
@ahmetdemirciai [Claude Code]
https://x.com/ahmetdemirciai/status/2069828529617826242
He stopped manually monitoring social media and that's when growth started. The old routine ate 1.5-2 hours every morning just researching what to post. So he built an n8n workflow that runs itself daily, pulling viral content, trends, and working formats from his niche into Google Sheets. He opens the sheet over coffee — 10 minutes — then Claude Code takes that data and writes content in his voice for YouTube, Instagram, and Twitter; he spends 20 minutes reviewing and ships. Two-week result: 250-300 new YouTube subs, nearing 1,000 on Instagram, 150-200k Twitter views. His framing: AI isn't about writing faster, it's about the research-analyze-write loop running while you only approve.
@noisyb0y1 [OpenClaw]
https://x.com/noisyb0y1/status/2069801591989784832
A developer bought four Mac Minis to run OpenClaw and got 4 million views last month without touching the editor. Each Mini runs its own agent on a separate task — OpenClaw handles scripting, voiceover, editing, and researches what'll go viral; he just looks at the result. Jump Desktop lets him control all four from his MacBook anywhere, with dummy HDMI adapters so the machines never sleep. He routes Claude Sonnet for complex tasks and Haiku or Gemini Flash for routine to keep API costs down. His pitch against human VAs: machines run 24/7 and get it right the first time, no sleeping or eating.
@dotey [Claude Code]
https://x.com/dotey/status/2069632132431929651
A clean, geeky way to manage Claude Code skills: install them per-project only, never globally, and wire them up with symlinks. His reasoning on context is the real insight — even though skills only load their name/description summaries by default, dozens of global skills add up and clutter the agent's "workbench," plus more global skills means more accidental triggering that loads full content and wastes space. So he keeps one ~/GitHub repo of open-source skills, symlinks only the ones a project needs into .agents/skills, then symlinks .claude/skills to that. Best part: update once at the source and every project gets it; fix a bug in a skill and you can push it straight back to the open-source repo.
@nityeshaga [Claude Code]
https://x.com/nityeshaga/status/2069900616986759195
He solved AI-built slide decks by giving the agent a real CLI for PowerPoint instead of a skill. The problem with Anthropic's pptx skill: it treats Claude as a programmer, so it writes code to hand-edit XML and one alignment fix breaks another slide — he once spent two days playing AI roulette and built the deck by hand anyway. His tool, Hands-on Deck, maps every gesture you'd make in PowerPoint/Slides/Keynote to a command: click a box and type, drag a shape, look at a slide. That changes the agent's behavior — it works like a designer, looks, makes a surgical edit, looks again. Installs as a skill in Claude Code, Codex, or any harness that supports skills.
@arsh_goyal [OpenClaw]
https://x.com/arsh_goyal/status/2069696083060719654
He set up OpenClaw on a $4,999 NVIDIA DGX Spark and the whole thing was easier than expected. One command pulled OpenClaw, the OpenShell sandbox, and a local model; then he connected it to Telegram for remote access. The point is the always-on loop: OpenClaw doesn't answer a prompt and stop — it keeps running in the background, checks its tasks, and acts on its own, a personal agent that works while you sleep. Now he messages the agent from his phone and everything runs on the device's GB10 chip and 128GB unified memory — no cloud bills, no data leaving his desk.
@regent0x_ [Claude Code]
https://x.com/regent0x_/status/2069829980444402076
A guy almost sold his $15k crypto mining rig for parts when mining stopped covering the electricity bill — instead he flashed it to serve AI and now clears ~$5k/month. Same GPUs, same wall, different software stack: he wiped the mining OS, installed Linux, and ran vLLM to split a 70B model across six GPUs, serving it over a local API. Clients who can't send data to the cloud now pay him: a law firm analyzing case files that legally can't leave the building, an agency generating content without per-token fees, and two devs routing their entire Claude Code and OpenAI workload through his hardware to zero out their bills. The rig that was about to become spare parts out-earns its best mining month by 10x.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2069642792536625586
He turned Claude Code into a reverse-engineering tool for his car's internal network. Every modern car has a bus where parts talk — speed, rpm, turn signals — but carmakers don't document what the raw numbers mean, so decoding each signal used to take hours of manual work. He built a Claude Code skill that watches the live data stream, works out which value maps to which real action, and even ingests dashboard footage so Claude can line up what's on screen with what's in the data. Flick the turn signal, it spots the number that changed and labels it — cracking signals at a 90%+ hit rate in a fraction of the time. Then you can pull true battery percentage from an EV or trigger locks and lights.
@0xCortexl [Claude Code]
https://x.com/0xCortexl/status/2069659323454636131
The Karpathy "second brain" idea — stop pointing AI at your code, point it at your memory — turned into a $20,000/month business for one developer. He aimed Claude Code at an Obsidian folder, drops in articles, transcripts, and PDFs, and Claude reads, links, and files everything into a living wiki. Three folders, five minutes of setup, and a system that gets smarter with every source. Now with 3,000+ notes it surfaces connections he'd never have noticed, and clients who pay for research, analysis, and strategy get results in hours instead of weeks. $20/month in, $20,000/month out.
@rileywestreel [Claude Code]
https://x.com/rileywestreel/status/2069812291621847550
A clean five-step recipe for running Claude Code inside an Obsidian vault as a note-editing agent. Install via npm, enable Obsidian's Terminal community plugin, open the integrated terminal, cd into your vault folder (critical — otherwise Claude can't see your notes), and run claude. From there the agent reads and edits .md files with no copy-pasting, generates notes, tables of contents, and MOCs, and refactors and auto-links across the entire vault through natural language. The framing: Obsidian becomes a knowledge base with an agent that sees all your files at once and works with them like a developer through the terminal.
@Pilgrim6938 [Claude Code / OpenClaw]
https://x.com/Pilgrim6938/status/2069803054787162375
Over two weeks he migrated Claude Code off his laptop onto a residential-IP VPS and now doesn't want to go back. His thesis: Claude Code's biggest problem was never model capability, it's the runtime — locally you hit closed-lid interruptions, re-logins after network switches, and SSH drops killing a long job halfway. He spun up a 2C4G Ubuntu 24.04 box on VoyraCloud (delivered in minutes), and recently had Claude Code spend ~40 minutes reorganizing a ~20,000-line project directory; he closed the laptop to go eat, reconnected, and the task was still running. The realization: the computer is no longer the execution layer, it's just a console — the real work happens on the cloud node.
@totoche [Claude Code]
https://x.com/totoche/status/2069665123669598318
He coded an app called Lucid so his Mac keeps Cursor, Claude Code, and Codex running with the lid closed, screen off, and locked — zero interruptions, and nobody can sit down at his open machine and dig through his trade secrets. The use case is mundane and perfect: he goes to the gym for a hard session, the agents keep grinding on his SaaS projects, and he comes back ~30 minutes later to find the work done. Sometimes the friction worth removing isn't in the agent at all — it's the laptop refusing to work with the lid shut.
@SleepMoneyMaker [Claude Code]
https://x.com/SleepMoneyMaker/status/2069802102369489368
He's testing a shared memory layer for Claude Code and Codex: one LEARNINGS.md per repo, read and curated by both agents. The problem he's solving is real — built-in memory is siloed (Claude's in ~/.claude, Codex's in ~/.codex), so Codex re-learns what Claude figured out and a fresh clone starts from zero. His fix lives in the repo's CLAUDE.md / AGENTS.md so it auto-loads every session; on a /dream command a fresh-context subagent mines the session diff and proposes evidence-cited learnings, then the main thread curates them into one deduped, capped file. One version-controlled source of truth every agent, teammate, and CI run inherits.
@yaohui12138 [Claude Code]
https://x.com/yaohui12138/status/2069624424068952491
He bet on a World Cup match by saying one sentence to an agent — no app, no clicking, no looking at the odds board. For Scotland vs Brazil he told Claude Code "analyze who's more likely to win," and via the Agent Trade Kit it pulled live OKX Outcomes odds (Brazil 72c), then cross-validated three independent markets — OKX, Polymarket, and de-vigged bookmakers — all converging on ~71% Brazil. The sharp part was the conclusion: with all three markets agreed, there's no mispricing and no edge — buying Brazil at 72c is negative expected value unless you think true probability exceeds 73%. He says it beats his own guessing because it never bets on who it likes and checks dimensions he's too lazy to.
@SUOHA_AI [Claude Code]
https://x.com/SUOHA_AI/status/2069753158843142262
He followed a step-by-step prompt chain to build a World Cup prediction agent from scratch with Claude Code, no stats or coding background needed. The pipeline is genuinely academic: pull free data, quantify team strength with Elo, compute win/draw/loss and scoreline probabilities with a Poisson distribution (including Dixon-Coles draw correction), run a Monte Carlo simulation of all 48 teams' title odds 10,000 times, then de-vig and compare against Polymarket to find value gaps, and backtest with Brier scores to check whether there's a real edge. His finished agent now outputs daily match probability reports, a championship distribution, Polymarket value signals, and a Brier comparison — turning grad-level forecasting into a tool a regular person runs every day.
@Sprytixl [Claude Code]
https://x.com/Sprytixl/status/2069806033636147321
One developer wired four Raspberry Pi CM5 modules (16GB each) into a single €450 Sipeed NanoCluster board and killed his $400/month in subscriptions. One 65-watt power supply, a built-in gigabit switch, zero external servers, and distributed llama partitions a model across the four nodes automatically — no quantization, 30 tokens/sec on small models, everything local. The math is the pitch: ChatGPT Pro $200 + Claude Code $200 = $4,800/year, versus €450 in parts that breaks even in month two. Most people will keep paying Anthropic and OpenAI forever; a few will spend a weekend wiring Pis and never see a subscription invoice again.
@k8adev [Claude Code]
https://x.com/k8adev/status/2069785685729407073
His company Solu just cancelled Figma entirely and migrated to a Claude-native build stack. Claude Design now handles their design system and prototypes, Claude Code does technical refinement, ClickUp task organization, and development (by technical people), and Coderabbit does code review. It's a concrete data point in the "design tools getting eaten by agents" story — not a side experiment but a paying team ripping out an industry-standard tool and replacing it with the Claude family across design, dev, and project management.
@Mho_23 [Claude Code]
https://x.com/Mho_23/status/2069838139171250629
He built an AI animation ad generator stitching Gemini Omni + GPT Images 2 + Claude Code that turns any product into a fully scripted 3D explainer video. You drop in product photos and a one-line pitch; the system analyzes your brand, proposes four ad concepts with hooks and shot lists, then scripts every shot, renders each clip in parallel, adds voiceover, QC-checks, and stitches a finished vertical ad. Three-minute setup, 4-12 minutes per ad, under $3 in API credits each — his claim is it replaces the $200+/video editor workflow at 10x the speed.
@clairevo [Claude Code]
https://x.com/clairevo/status/2069828122640548204
She's now running GLM 5.2 as her default model in both Claude Code and Cursor, and it's cost her — checks notes — $3.36. Her first review of an open-weights model, she says GLM 5.2 (via OpenRouter) gives her Opus vibes without the Opus price. She covers how to wire these models into Cursor and Claude Code via the OpenRouter API, its front-end design sense, and performance on a long-running autonomous task. The moment it won her over: it put "chatprd pink" in her docs without being asked.
@BigbirdflyChan [Claude Code]
https://x.com/BigbirdflyChan/status/2069794065185447977
A jaw-dropping usage-scale data point: SemiAnalysis's cumulative Claude Code spend from May 23 to May 29, 2026 went from about $15,000 to $171,476 in a single week. That implies annualized token consumption approaching $9 million on Claude Code alone. Whatever you think about ROI, it's a concrete marker of just how hard a single analyst team is now leaning on agentic coding — the kind of number that reframes is-this-expensive into what-are-they-building-that-makes-it-worth-it.
@ThisisHan1_ [Claude Code]
https://x.com/ThisisHan1_/status/2069669289288978649
A cautionary, very technical account of getting his Anthropic account banned. Microsoft is cutting off employees from Claude Code at end of June and pushing Copilot CLI, so he was helping a colleague keep using Claude by swapping in GitHub Copilot's API for Claude's. It worked early in the month, then broke with a 400 error: Copilot's proxy only accepts claude-opus-4.8 (dot), but Claude Code sends claude-opus-4-8 (dash), and selecting "Opus 4.8 (1M context)" emits an id with [1m] that Copilot has no variant for. The kicker: reviewing his session logs, he found Claude Code had inspected its own binary, got flagged as a tampered client with anomalous traffic, and the account was banned — and the appeals form was stuck in a redirect loop.
@Takuiten [Claude Code]
https://x.com/Takuiten/status/2069715168578420746
A sharp piece of forensic analysis: he reverse-engineered how the $ZERO "decentralized inference" token was actually made by running the petals codebase through Claude Code. His claim is that the creator took petals, had the AI swap old parts for new (e.g. HuggingFace transformers from 2022 to latest vLLM), and claimed it as invention — and you can reproduce the exact changes yourself because the AI will suggest them. He says the giveaway was a generic AI-generated response about speculative decoding that was technically wrong, plus on-chain evidence that the supply was quietly cornered through a fresh wallet. A clean example of using Claude Code as an investigative tool, not just a builder.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2069723070248681897
He automated an entire online-course production pipeline with Claude Code: record the lesson, cut it, generate full subtitles, upload to Vimeo, and post the summary and timestamps to the learning app. You toss in a video and it's done. For anyone running courses or a school, this is the boring-but-expensive back-half of content production — the editing, captioning, uploading, and metadata — collapsed into a single hand-off.
@cevenif [Claude Code]
https://x.com/cevenif/status/2069682184403734846
He installed a WeChat group-chat summary skill (built by another user on top of wx-cli) and ran it on a real group's daily messages with reliable results. The hardcore part is wx-cli itself: it decrypts the local WeChat database directly without even disabling SIP. His workflow tip is the genuinely useful bit — when it errors, just hand the error to Codex or Claude Code and let it fix itself, and it basically runs through mindlessly.
@laowangbabababa [Claude Code]
https://x.com/laowangbabababa/status/2069820358040174695
He's been hitting the wall every agent hits: Claude Code can write code and edit docs but getting it online to fetch things kept breaking — Twitter API costs money, scraping needs subscriptions, Bilibili and Xiaohongshu just block you. He installed Agent-Reach, a free open-source repo that opens up 14 platforms with one command, then had the agent search and summarize the comments on a Xiaohongshu post in about ten seconds. The value is that it does one thing — handles all the tool selection and config (yt-dlp for subtitles, twitter-cli for tweets, xhs-cli for Xiaohongshu) so the agent just calls them directly; web, YouTube, RSS, and Weibo work zero-config.
@MagiccatMila [Claude Code]
https://x.com/MagiccatMila/status/2069774643171569837
She finally figured out where those content-repost accounts get their material: a tool called AgentKey that pulls public content, comments, likes, and shares across Xiaohongshu, Douyin, Kuaishou, Bilibili, Weibo, Zhihu, YouTube, and platforms you wouldn't expect. The trick is one command installs it onto Codex, OpenClaw, or Claude Code, and you've got an automated research/repurposing workflow that she says cuts 90% of the research-and-repost time. The data sources are stable and need no manual configuration — a big lowering of the bar for beginners doing competitive research.
@Claude_Digest [Claude Code]
https://x.com/Claude_Digest/status/2069782254663115231
A detailed write-up of combining Claude Code with the Apify scraping platform and MCP to build an end-to-end data pipeline. You can scrape massive web data — including behind paywalls that need a login — and have the agent autonomously write results into external services. The demo: hand Claude a YouTube URL in the terminal, it runs a Video Analyzer Actor, and the timestamped summary and key topics get cleanly formatted into Notion with no human touch; web schedules sync to Google Calendar. The framing is connecting the agent's "hands" to existing platforms and databases so the whole input-to-output flow is automated.
@0xhbam [Claude Code]
https://x.com/0xhbam/status/2069794330051784770
A new competitor-ad-intelligence skill (inside /goose-ads) that tears down any competitor's paid strategy directly in Claude, Claude Code, Cowork, or Codex. Install with one npx command, run "/goose-ads do a teardown of my competitors ads," and Goose pulls the competitor's live ads from the Meta Ad Library, finds the hooks and formats actually working, reverse-engineers their funnel, and hands you the weaknesses to attack. The logic is sound: other companies in your exact niche have already spent money testing what converts, so you repeat what works — a non-coding, marketing-side use of the skills ecosystem.
@mk_1_1_ [Claude Code]
https://x.com/mk_1_1_/status/2069738022741201265
He's testing AWS's official Agent Toolkit for AWS in Claude Code and finds it promising. His framing of the gap is useful: AI coding agents are convenient but hit walls building production-quality systems on AWS — they can't always track the latest service specs, they skip non-functional requirements like security and ops, and they don't understand AWS service combinations or design intent. The toolkit installs MCP server config and skills in three plugin commands. He also built a custom AWS-investigation skill with hard rules: verify auth with sts get-caller-identity, re-login via SSO when expired, always cite official docs, fall back to AWS CLI when MCP fails, and only use read-only profiles.
@skeptrune [Claude Code]
https://x.com/skeptrune/status/2069608374115967134
Amid Google politics over a shelved product, he gave the real-world testimonial: he uses the gws (Google Workspace) CLI through Claude Code literally every day, and it's load-bearing for his life — it handles his email, calendar, spreadsheets, slides, and more. It's a small post but a clean signal of where the non-coding value is: the agent isn't writing apps, it's operating your entire Google Workspace from the terminal as a daily driver.
@peesamac [Claude Code]
https://x.com/peesamac/status/2069733332775211394
A compact but telling use case: he told Claude Code to connect MCP, control Google, and build a Form plus Dashboard as a whole system — by itself. No code written by hand, just an instruction and an MCP connection turning into a working Google Workspace system. It's the same pattern showing up everywhere this week: the interesting Claude Code work increasingly isn't software development, it's wiring the agent to a platform and letting it operate it.
@RegalosDigitals [Claude Code]
https://x.com/RegalosDigitals/status/2069574611797159948
A full step-by-step for using Claude Code to clone any website's UI into code, via AIDesigner's MCP. Prereqs: Node.js, Claude Code, and an AIDesigner account (free tier). Init the MCP with npx @aidesigner/agent-skills, open Claude Code in the folder, run /mcp and select aidesigner, then authenticate. After that the magic command is just "CLONE <url> using aid," and Claude auto-invokes generate_design in clone mode and hands back Tailwind + HTML/React. You can refine afterward ("make the navbar sticky") or request ultradesign mode for max quality.
@felixleezd [Claude Code]
https://x.com/felixleezd/status/2069645662036349008
On day 300 of building as a designer, he shipped an AI Interviewer built with Claude Code aimed at the brutal 2026 graduate job market. The tool is simple and concrete: upload your resume, get real-time matched jobs, practice a mock interview, and get actionable feedback in 10 minutes. It's free, and it's a clean example of a non-engineer designer shipping a useful utility end-to-end with the agent.
@Offers_jp [Claude Code]
https://x.com/Offers_jp/status/2069579162957087145
A standout non-engineer case: the CEO of Baigie (sogitani) used Claude Code himself to build, in-house, a free recruitment-site diagnostic tool with about 200 evaluation items — designed from 100+ B2B support engagements and 10,000+ jobseeker survey responses. The framing of the upcoming webinar is the interesting part: not "AI for efficiency" but "AI for value creation," with a non-engineer sharing the real experience of building a business tool with Claude Code, including how he narrowed it to 200 items and tuned for accuracy.
@yibie [Claude Code]
https://x.com/yibie/status/2069646862349680650
He translated Flask author Armin Ronacher's essay "The Coming Loop," which he calls the most honest loop reflection he's seen — it asks the question nobody dares: do you actually like the code the loop writes? Ronacher admits loops work astonishingly well for code porting (he ported MiniJinja to Go), performance exploration, security scanning, and research — anything that transforms existing code or produces artifacts not meant to live long. But for durable code he loves, he won't use them: current models write over-defensive, over-complex, locally-reasoned code, add fallbacks instead of making bad states impossible, and unattended harnesses like Claude Code with ultracode now produce worse code than last autumn. His warning: you can't fully opt out, because attackers and competitors will loop whether you do or not.
@moshhamedani [Claude Code]
https://x.com/moshhamedani/status/2069922900963107150
A blunt quality complaint worth surfacing: Claude Code was super slow and generating some of the worst code he's seen that day. His warning lands harder coming from a veteran instructor: good luck to those who think code reviews aren't necessary — one day the pile of AI slop gets so big that even the simplest change takes forever and breaks unrelated features along the way. The counterweight to all the "build anything in minutes" euphoria.
@tualatrix [Claude Code]
https://x.com/tualatrix/status/2069801442878304453
More concrete pain from heavy use: a task that runs along and then fails silently with no error — expand it and a subagent had failed underneath. And a simple "commit + push" instruction that made Claude think for a full 10 minutes without ever committing. He wonders aloud whether these exist in the Claude Code CLI or are specific to the desktop version (which he says is far worse). Useful signal for where the reliability gaps are: silent subagent failures and instructions that spiral into overthinking.
@kunihirotanaka [Claude Code]
https://x.com/kunihirotanaka/status/2069695578389569611
A charming and telling moment: he had Claude Code build him a personal email-sorting system, and the agent itself proposed switching from the slow, expensive Claude CLI backend to Sakura's AI Engine. When he made it run the numbers, Claude concluded "Sakura is dirt cheap." He jokes Claude must be an honest kid — but the real signal is that agents are now reasoning about their own running costs and volunteering to swap out their own backend to save you money.
@nityeshaga [Claude Code]
https://x.com/nityeshaga/status/2069844061998256615
A second, important note from him: he had no idea running open-source models in Claude Code was this easy. The relief is structural — it removes the background worry he'd carried for four months about what happens to his fleet of AI employees running on headless `claude -p` mode if Anthropic ever pulls the subsidized Max-plan access. This is the quiet insurance policy a lot of heavy users are now reaching for: keep the workflow, swap the model.
@shivkanthb [Claude Code]
https://x.com/shivkanthb/status/2069912282910605347
He built a tiny Ray-Ban display app to view Claude Code's status, so he can make tea, vacuum, or read a book while his agents grind and only step in when they need his attention. It's a small build but part of a clear emerging need: ambient, glanceable awareness of long-running agents so you're not chained to the terminal babysitting them.
@dseijo [Claude Code]
https://x.com/dseijo/status/2069885883226964178
He can't resist trying Oura Ring gesture control for Claude Code: closed fist to start dictation, open hand to stop, swipe right for enter. He admits it's not very useful — but working like a supervillain is priceless. The serious thread underneath the joke is the same as the Ray-Ban app: people want low-friction, physical ways to drive and monitor agents without sitting at the keyboard.
@noel_ai_lab [Claude Code]
https://x.com/noel_ai_lab/status/2069769920317911165
She finally finished her first-ever music video, made entirely with AI — for a worldview she loves. Music from SUNO, visuals from Higgsfield (Seedance 2.0) crossed with Claude Code, and captions via ffmpeg. She admits it's rough in places and there's more she wants to do, but she had to ship it. Her line is the takeaway: it's no longer about whether you can make it — if you just have something you want to make, it can take form, and someone who'd never made an MV got this far.
@Ecom_Matteo [Claude Code]
https://x.com/Ecom_Matteo/status/2069832722239877240
He built an in-house Higgsfield clone with nothing but Claude and Claude Code in four hours. Short on detail but a clean data point for the recurring theme: tools people pay subscriptions for are increasingly something a single person can reproduce a working in-house version of in an afternoon — and they're doing it.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2069645780428935625
He turned his own recorded bass playing into a fully beat-synced AI video using HyperFrames skills with Claude Code. The flow is clean: his bass audio goes into the HyperFrames beats skill, which auto-analyzes BPM and timestamps and outputs JSON; the hyperframes-animation skill supplies beat-synced design patterns; Claude Code just writes code following the skills' rules and renders. The result is a 30-second, 11-scene video with scene cuts on strong beats and flash/RGB-glitch/scale-pulse on weak beats — the structure being "skills bring the template, AI executes it."
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2069873193494470724
A clean B2B-SaaS-style play from a friend doing lead gen for workplace-accident lawyers: buy exact-match domains like "workplace accident lawyer Dallas TX," rank them, and sell the leads. He used to hand-build and manage hundreds of simple one-page WordPress sites; now Claude Code just makes and manages them at scale. His thesis: every company serving SMBs should be doing "best AI tool for X" at this scale — it's blue ocean and the best time in the world to be marketing.
@mrmiket64 [Claude Code]
https://x.com/mrmiket64/status/2069766908027564499
He has Claude Code or Codex installed on each of his home-lab servers and finds it incredibly practical — he talks to each one whenever he needs to install something or troubleshoot. He muses that having them all in a group chat would be interesting (setting aside the attack surface). It's a glimpse of where personal infra is going: an agent per machine, addressed conversationally, as the default admin interface.
@taka_aki [Claude Code]
https://x.com/taka_aki/status/2069601248589041748
He tested Databricks' ucode on real hardware to solve a problem teams are starting to feel: do you actually know who used which coding agent and how much it cost? With ucode you can use Codex and Claude Code without API keys while getting unified cost visibility through an AI Gateway — and he says it's satisfying to see consumption from different vendors' models lined up in the same summary. As agent fleets grow, this cost-attribution layer becomes a real need.
@msjiaozhu [Claude Code]
https://x.com/msjiaozhu/status/2069603026453876833
He's bullish on Headroom, a tool that wraps your CLI and saves money for heavy Codex/Claude Code users whose context constantly blows past the five-hour window. It proxies your CLI, auto-detects JSON, code, and logs, picks the optimal compressor for each, and leaves a retrieve tool so the LLM can pull details when it actually needs them. His real-world result: daily agent runs are noticeably faster and cheaper, especially when chaining multiple agents.
@ytjessie_ [Claude Code]
https://x.com/ytjessie_/status/2069583248771199228
She tested the /persona-builder skill (part of an Influencer Ignition Kit, available on Claude Code with the Pika MCP installed) and it generated an impressive persona in one shot from her real Instagram profile. A small but concrete example of the skills-plus-MCP pattern doing useful non-coding work — reading a real social profile and producing a usable creator persona without iteration.
@aryaman2020 [Claude Code]
https://x.com/aryaman2020/status/2069582061665677602
A great debugging war story: he was using Claude Code to debug a poorly-performing training run and discovered that, instead of training on 30 batches of data, it had somehow decided to train a brand-new model for 500 steps on each batch and then average the 30 sets of weights. The value here is the agent as a forensic partner on subtle ML bugs that would be painful to find by hand — and a reminder that agents can quietly do something wildly different from what you intended.
@MrMelows [Claude Code]
https://x.com/MrMelows/status/2069597260694839559
He solved a problem every vibe-coder knows: making the same mistakes across different projects and rebuilding from scratch things that already work in another repo. His fix: give Claude Code a shared "brain" — one knowledge-only repo of working patterns, decisions already made, and traps already fallen into, which the agent consults before writing a single line. Setup took 13 minutes. Now starting a new client project, it reuses what he's already tested instead of rethinking everything from zero.
@zeChedli [Claude Code]
https://x.com/zeChedli/status/2069834941194526783
He's training himself on AI structure, vibe coding with Claude Code, and LLMs broadly, and has built several real projects: landing pages that would have been billed 2,000-5,000 euros, mobile games in the 7,000-10,000 euro range, and internal tools that'll save him thousands of euros a year. His urgency is the message: position yourself now, because things move extremely fast — he points at Fable 5 getting pulled to argue that if politics flips overnight, prices rise or access cuts, and the window closes.
@alexgoughcooper [Claude Code]
https://x.com/alexgoughcooper/status/2069577242909307231
The builder of Parker (an ad-analytics product) shared how he personally uses the stack: Claude Code for the chat side, and the Parker web app for checking ad performance, snooping competitor ad libraries, and saving ads to a swipe file. Alongside a long list of shipped features (AI tagging, MCP beta, one-click review syncs, scheduled Slack/email reports), it's a useful look at how a founder splits work between an agent and his own product — the agent drives the conversational/analytical layer, the app handles the dashboards.
@HolyFinance [Claude Code]
https://x.com/HolyFinance/status/2069818375723708512
A one-liner with real teeth: the president of Alantra has spent a month loading up on shares — and yes, his agent in Claude Code tipped him off. Tiny post, but a concrete glimpse of agents being run as always-on financial monitors that surface insider-buying patterns their owner would never catch by hand.
@StefanMaier [OpenClaw]
https://x.com/StefanMaier/status/2069687437178151086
He got his OpenClaw running on local DeepSeek V4 Flash on his DGX Sparks, and the whole move to local happened fast — within the last 24 hours. A short note, but another data point in the strong week-long theme of pushing personal agents fully local onto DGX Spark hardware with open-weight models, no cloud dependency.
@pablinche [OpenClaw]
https://x.com/pablinche/status/2069820882017550661
He's playing with an OpenClaw + Codex + Figma loop for design: feeding web inspiration to Codex via Telegram while asking it to produce a basic design system in Figma he can use as a base for his own project. A small but interesting non-coding workflow — driving a design-system bootstrap through a messaging app and an agent rather than sitting in the design tool itself.
@yousukezan [Claude Code]
https://x.com/yousukezan/status/2069603538813190443
A small but important safety moment: while he was using Claude Code, it warned him that "suspicious text disguised as instructions to me was mixed in." No actual harm occurred, but he notes how terrifying it would have been if it were real. It's a concrete sighting of prompt-injection defense firing in the wild — the agent catching and flagging an embedded instruction attack mid-task.
@ikoichi [Claude Code]
https://x.com/ikoichi/status/2069706404017979573
A small consulting win that makes the speed argument concrete: a client asked for a medium-sized feature, he built it in half a day thanks to Claude Code, and that responsiveness earned the client's trust. His pointed question is the real content: can a 50-plus-person company keep the same pace? The agent isn't just a productivity boost here — it's a competitive weapon for a solo or tiny shop against bigger, slower teams.
@kwindla [Claude Code]
https://x.com/kwindla/status/2069819081511911696
A delightful one: the Tavus team re-created Apple's 1987 "Knowledge Navigator" concept video using their conversational video-agent stack, captured in a single take — and the demo's real use isn't an actor preparing a lecture but actual usage: some recreational 3D printing and some vibe coding with Claude Code, with live 3D model generation that's fast thanks to Kimi K2.6 on Cerebras. It's a vivid illustration that a 40-year-old vision of a conversational assistant doing real work is now buildable, with Claude Code as the coding layer.
@drbarnard [Claude Code / OpenClaw]
https://x.com/drbarnard/status/2069771620533584249
A glimpse of a multi-agent dev pipeline he's planning. Right now he uses Codex and Claude Code to write specs and bounce them off each other in the planning phase, before Hermes (running ChatGPT 5.5) merges them. The endgame: once the first build is done, fire a feature request or bug fix into Discord from anywhere, have Codex and Claude plan it and send him a quick gut-check, then build it while checking each other's work, and deploy to TestFlight via Xcode Cloud — one message and one sanity check to ship a feature to his iPhone.
@PMDBT [OpenClaw]
https://x.com/PMDBT/status/2069743379244830960
A useful meta-post on LLM writing tells, from someone who uses them heavily. Beyond the lazy "too many em dashes" signal, he points at structural giveaways like "that isn't X, that's Y" and "here's what actually decides this, the thing you keep dancing around." His own workflow is the practical part: he word-vomits unfiltered thoughts himself, then asks the model to fix grammar, tighten, and rewrite per platform — feeding it his prior writing so it matches his style with zero em dashes. His take: nothing wrong with using LLMs to write, as long as the ideas are yours.
@sergeykarayev [Claude Code]
https://x.com/sergeykarayev/status/2069811959042892064
Responding to Karpathy, he says the loop approach has been a game changer for his team — and they took it further by building it into Superconductor as model-agnostic, so it works across Claude Code, Codex, Pi, OpenCode, Amp, and Grok. The signal: serious teams aren't just adopting the loop pattern, they're productizing a harness-level abstraction over all the coding agents at once, betting that the orchestration layer matters more than any single underlying agent.
@ZEIRISHI_Ichibe [Claude Code]
https://x.com/ZEIRISHI_Ichibe/status/2069739068699955684
From an accounting/tax professional: after a seminar, the dots connected — because Google Calendar, Maicommon, and Proudnote can now be MCP-connected, a Claude Code "AI secretary" for client management suddenly looks viable. He's eyeing concretely useful glue like jumping from a client record straight into the right chat tool or logging into that client's e-Tax page. It's a clean non-engineer, professional-services view of where the agent fits: as the connective AI secretary across the tools he already uses.
@Crownzdesigns [Claude Code]
https://x.com/Crownzdesigns/status/2069681797793800417
Day 20 of a build challenge: a dark, premium showcase carousel for a 3D character club, built with Claude Code, with an intro loader and slide transitions where the background color shifts with each 3D character. A simple but concrete front-end design output — a reminder that a lot of the daily Claude Code value is just designers shipping polished interactive UI pieces fast.
🗣 User Voice
User Voice
Where the agent runs is the new bottleneck, not how smart it is. "Claude Code's biggest problem was never the model, it's the runtime" — a closed laptop kills the task, a network switch logs you out, a long job dies on SSH drop. People are migrating to residential-IP VPSes, DGX Sparks, and Mac Minis just to keep agents alive overnight. @Pilgrim6938 @totoche
Cross-agent memory is fragmented and everyone is hand-rolling a fix. Claude's memory lives in ~/.claude, Codex's in ~/.codex, so each re-learns what the other already figured out and a fresh clone starts from zero. Users want one shared, version-controlled LEARNINGS.md / CLAUDE.md every agent inherits. @SleepMoneyMaker @MrMelows
Cost anxiety is driving a quiet exodus to open weights and cheaper models. People run GLM 5.2 or DeepSeek V4 inside Claude Code, and several said the real relief was discovering open-source models slot in easily — killing the fear of losing subsidized Max-plan access for their headless `claude -p` fleets. @clairevo @nityeshaga @VengeonsP
The slop problem is real and unsolved. Complaints about over-defensive, over-complex code, silent subagent failures, and a simple commit+push that thinks for 10 minutes. Even loop advocates admit unattended harnesses produce worse code than last year. @moshhamedani @tualatrix @yibie
People want ambient awareness of running agents without babysitting. They're building Ray-Ban status displays and wiring up Oura Ring gestures just to glance at whether an agent needs them while they make tea or hit the gym. @shivkanthb @dseijo
Where the agent runs is the new bottleneck, not how smart it is. "Claude Code's biggest problem was never the model, it's the runtime" — a closed laptop kills the task, a network switch logs you out, a long job dies on SSH drop. People are migrating to residential-IP VPSes, DGX Sparks, and Mac Minis just to keep agents alive overnight. @Pilgrim6938 @totoche
Cross-agent memory is fragmented and everyone is hand-rolling a fix. Claude's memory lives in ~/.claude, Codex's in ~/.codex, so each re-learns what the other already figured out and a fresh clone starts from zero. Users want one shared, version-controlled LEARNINGS.md / CLAUDE.md every agent inherits. @SleepMoneyMaker @MrMelows
Cost anxiety is driving a quiet exodus to open weights and cheaper models. People run GLM 5.2 or DeepSeek V4 inside Claude Code, and several said the real relief was discovering open-source models slot in easily — killing the fear of losing subsidized Max-plan access for their headless `claude -p` fleets. @clairevo @nityeshaga @VengeonsP
The slop problem is real and unsolved. Complaints about over-defensive, over-complex code, silent subagent failures, and a simple commit+push that thinks for 10 minutes. Even loop advocates admit unattended harnesses produce worse code than last year. @moshhamedani @tualatrix @yibie
People want ambient awareness of running agents without babysitting. They're building Ray-Ban status displays and wiring up Oura Ring gestures just to glance at whether an agent needs them while they make tea or hit the gym. @shivkanthb @dseijo
📡 Eco Products Radar
Eco Products Radar
Codex — the constant co-pilot to Claude Code; users run both, cross-check specs between them, and want shared memory across the two.
Cursor — still the IDE half of many setups, paired with Claude Code for the terminal/agent half.
Obsidian — the runaway favorite for "second brain" builds: point Claude Code at a vault and it reads, links, and files everything.
MCP — the connective tissue for non-coding workflows: Google Workspace, Apify scraping, AIDesigner UI cloning, AWS toolkits.
Hermes / OpenClaw — the always-on personal-agent layer people compare and switch between.
DeepSeek V4 / GLM 5.2 — the cheaper open-weight models people are dropping under Claude Code to cut costs.
Codex — the constant co-pilot to Claude Code; users run both, cross-check specs between them, and want shared memory across the two.
Cursor — still the IDE half of many setups, paired with Claude Code for the terminal/agent half.
Obsidian — the runaway favorite for "second brain" builds: point Claude Code at a vault and it reads, links, and files everything.
MCP — the connective tissue for non-coding workflows: Google Workspace, Apify scraping, AIDesigner UI cloning, AWS toolkits.
Hermes / OpenClaw — the always-on personal-agent layer people compare and switch between.
DeepSeek V4 / GLM 5.2 — the cheaper open-weight models people are dropping under Claude Code to cut costs.
Comments