Super User Daily: July 31, 2026
The theme today isn't "look what Claude Code can build," it's "look what people quietly turned it into." Solo operators are running whole ad agencies and web studios off a single subscription, non-engineers are closing their monthly books and filing government paperwork with skills, and a striking number of people are pointing the terminal at their own note vaults and letting it find the ideas they forgot they had. Underneath the wins there's a loud, consistent grumble about tokens, cost, and a brutal multi-day outage. On the OpenClaw side, the mood has curdled into "does anyone still run this?"
@buger [Claude Code]
https://x.com/buger/status/2082511681553326480
Burns roughly 30 billion tokens a month and treats it as a leverage number, not a flex. The setup is five agents in parallel plus self-healing loops: agents work features, file GitHub issues when they hit a bug, and a separate "garbage collector" agent fixes those bugs almost in real time. In one week that shipped 100 closed PRs and five releases on his 10-year-old jsonparser, 100% test coverage on rsync that surfaced 20-year-old bugs, and seven PRs on a Kubernetes operator. The agents run 24/7; he sets goals before bed and wakes to finished experiments.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2082492885283352748
A solo SEO consultant pulling €20k/month with a two-tool stack: Claude Code plus Claude Desktop, nothing else. He dumped every technical audit he ever wrote at his old agency into Claude, and now a single skill covers 95% of recurring problems with the "why it's bad" attached. Workflow: a heavy Screaming Frog crawl feeds the audit skill, which emits a .md and then a PowerPoint. For sales he records every call with Plaud, wired via MCP, fed Claude all his proposals since 2021 with win/loss reasons, and got a skill that reshapes his pitch to each client's profile.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2082298062739759223
A $35k-MRR web design agency with zero employees, because the production layer runs on models. Fable 5 is the design engine (brief in, full system out), Claude Code on Opus 5 is the build engine (writes the site, you review against spec), Higgsfield handles hero video and media. 12-14 clients at $750-5,000/month, cost to deliver about $750/month, margin north of 90%. He only touches intake and QA — the labor-heavy middle is gone.
@NXR_NIROX [Claude Code]
https://x.com/NXR_NIROX/status/2082334475539931283
A one-man ad shop running seven Claude Sonnet 4.6 agents through Claude Code Router. It scans Google Maps for businesses in small cities with no site or a 2014-era one, then runs the whole pipeline: mockup, promo video, sales message. Roughly 47 clients at $400 each, about $18,800/month, on ~$480 of API. State is shared over the filesystem via MCP servers, and one agent runs on his iPhone so it replies to leads from the subway — only waking him for deals over $3,000 or when reply rate drops below 12%.
@shin_chan_ai [Claude Code]
https://x.com/shin_chan_ai/status/2082264780526739659
A year as an "AI-Ops manager" turned into a hard results sheet: roughly 80% company-wide adoption of Cursor and Claude Code, inside-sales research and analysis cut by about 1,500 hours a year, release notes going from 2-3 business days to minutes. His one thesis after all of it: the thing that separates AI wins from AI theater isn't tool tricks, it's redesigning how people and work fit together.
@startupideaspod [Claude Code]
https://x.com/startupideaspod/status/2082602070629286033
Non-technical founder whose agent runs his Facebook ad campaigns: two ad sets a day, five ads each, written and uploaded by the agent. It lets them run 2-3 days, kills the worst, and promotes winners into a pool that competes for budget, then studies every ad it has ever made. The hard part is entropy — the agent gets stuck thinking the same way — so he feeds it competitor ads from the ad library and mined YouTube/podcast transcripts as fresh DNA. He just hands Claude Code a transcript and asks it to walk him through setup.
@Mho_23 [Claude Code]
https://x.com/Mho_23/status/2082600596457341398
Built a 24/7 AI UGC video editor that lives entirely inside Claude Code. You give it a reference video and your product info; the Higgsfield analyzer MCP breaks down the reference's structure, pacing, and timing, a custom Claude skill generates the clips with Gemini Omni, then it renders each scene, stitches them, and adds captions — out comes a finished 9:16 ad, no tab-hopping or timeline scrubbing.
@z0rynx [Claude Code]
https://x.com/z0rynx/status/2082502997196624188
Pointed Claude Code at a 7-year-old Obsidian vault — 4,182 notes, 11,600 links — and ran one command. Eighteen minutes later there was a file he never wrote: 63 orphan notes with no links in or out, and 54 of them were the same idea, rewritten across six years without ever being connected. He shipped that idea nine days later; it's the only thing he built this year that people paid for. The graph wasn't a brain, it was a mirror with a bias toward what he already believed.
@degenpiz [Claude Code]
https://x.com/degenpiz/status/2082430030294962412
A four-piece research stack that runs while he sleeps: Claude Code does the hands-on work, NotebookLM does heavy analysis on Google's servers so he burns zero Claude tokens, Skill Creator turns each working flow into a reusable skill, Obsidian stores results as plain markdown. One pipeline skill searches YouTube by topic, pulls top videos by engagement, feeds them to a fresh NotebookLM notebook, generates flashcards/mindmaps/audio, and dumps a dated markdown file to the vault. Weekly Monday 07:00 cron; 14 notes in week one, 92 by week four.
@juanca_mnz [Claude Code]
https://x.com/juanca_mnz/status/2082402085367164931
His best recent purchase is a cheap NUC mini-PC as a home server. He drives Claude Code from his phone via Termius + Tailscale, running side projects, agent loops (automatic SEO improvements every two weeks), a Hermes agent, Jellyfin, and Home Assistant, on 32GB RAM. The whole thing was configured in 30-40 minutes with Claude, almost nothing done by hand.
@marindedix [Claude Code]
https://x.com/marindedix/status/2082581871645565289
A dead-simple "agents work while you sleep" rig: an always-on Raspberry Pi 5 (8GB) with tmux, Tailscale, and Claude Code/Codex installed. Set a goal, close the laptop, and tmux keeps the session alive; wake up to results. It works on his personal projects while he's driving, sleeping, or at his day job — it doesn't stop just because he does.
@lawrencecchen [Claude Code]
https://x.com/lawrencecchen/status/2082359150026318022
Used 367 billion tokens in 30 days — peak 37 billion in a single day — by multitasking with cmux. The 80/20 is boring: hit Cmd+Shift+U whenever Codex notifies, review the artifact/bug fix, respond, and spin up a new workspace whenever there's nothing left to approve. No fancy harness, just vanilla Claude Code/Codex and built-in subagents. Would've cost $85k over API; he paid $7.4k across 37 Pro accounts.
@orogongon [Claude Code]
https://x.com/orogongon/status/2082353368895508488
Month-end admin, gone. One prompt tells Claude Code to pull an employee's timecard scans from a Google Drive folder and assemble January-June into a submission-ready A4 PDF, one month per page. This is the kind of multi-person paperwork that used to eat real hours from an office clerk.
@takechan_lawyer [Claude Code]
https://x.com/takechan_lawyer/status/2082287999900069904
A corporate lawyer set up intake templates for clients with recurring contract work, then built a Claude Code skill that screens the request and attachments and drafts follow-up questions whenever information is missing. Switching to this flow measurably raised his throughput on contract review — a clean non-coding professional use.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2082553012392972794
Wrote a /company-cfo skill that does the agency's monthly close: pulls bank, Stripe, payroll, and cards, reconciles them, and catches the traps — double-counts, internal transfers, a missing partner draw — then computes end-of-month cash to the penny and writes the leadership snapshot. Monthly financials in minutes instead of a bookkeeper backlog. It's one of 18 free open-source Maker Skills he shipped.
@petergyang [Claude Code]
https://x.com/petergyang/status/2082473770741272747
A product manager's repeatable six-step flow: define the problem and a design.md, prototype in Claude Design, draft an interactive HTML spec, then build the core flows with Claude Code. He built Tastemaker (a movies/TV/games sharing app) end to end this way. The lesson he keeps hammering: good upfront planning is the difference between the product you want and slop — resist AI's urge to start coding too soon.
@TokenGremlin [Claude Code]
https://x.com/TokenGremlin/status/2082455564269801549
Opus 5 built a browser desert-graphics demo with no meshes, textures, or downloaded assets — every surface generated in shader code. Procedural dunes on a GPU clipmap, sand that deforms permanently under footsteps then erodes back, GPU cloth, a physically based sky. The interesting part is the workflow: Claude got its own testing harness with headless Chrome, screenshots, and per-subsystem GPU measurements, so it tuned against real telemetry instead of guessing. Runs ~160fps at 1440p on a 5070 Ti.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2082288752227926087
Someone built SNOWFLOW — a browser WebGPU snow and waterbending demo with persistent terrain deformation — in nine hours and about 4 million tokens using Claude Code with Opus 5. It synthesized custom WGSL shaders, WebGPU pipelines, and Babylon.js terrain logic from scratch, then iterated on shader bugs and visual polish directly off browser screenshots of its own output.
@efekorayyy [Claude Code]
https://x.com/efekorayyy/status/2082524786198970514
Zero-cost YouTube automation: 15 videos and 2h40m of content plus a 26-minute music mix, no stock image or voiceover subscriptions. Images pulled automatically from NASA and Wikimedia public archives, narration from Microsoft Edge neural voices, editing in ffmpeg, all glued by six small scripts — each doing one job. He wrote none of the scripts by hand; he described what he wanted, Claude Code wrote them, and he pasted errors back for fixes.
@0x_fokki [Claude Code]
https://x.com/0x_fokki/status/2082442153146364113
A trader stopped storing research and started connecting it. Every setup lives in one Obsidian vault, wired three ways: chart pattern → outcome → market context → what he was thinking at the time. The pipeline: he types one command, Claude Code pulls the relevant sources, runs them through NotebookLM, and saves everything back structured and linked. Most traders screenshot a chart into Discord and forget it in 48 hours; his vault connects the pattern instead of just filing it.
@mave99a [Claude Code]
https://x.com/mave99a/status/2082598208891764939
Open-sourced two skills, design-review and build-phases, born from a real defect: intent-driven development makes plans swell and drift as the agent keeps asking clarifying questions. The fix is loop engineering — each review round spawns a brand-new agent that sees only the files on disk (not the prior chat) and can read but never edit, so it can't grade its own homework. It scores, fixes docs and test coverage, and hands architectural direction back to the human; then build-phases walks the plan one phase at a time with three verification gates before moving on.
@TheGeorgePu [Claude Code]
https://x.com/TheGeorgePu/status/2082536177580577181
Built a full website in a morning: 44 model calls, 727,000 tokens in, 34 cents — and he only knows those numbers because a third-party tool told him. His point, and a lot of people's: Claude Code has never once told him what a session cost, and neither has Codex. Cost visibility is table stakes people don't have.
@arshadkazmi42 [Claude Code]
https://x.com/arshadkazmi42/status/2082529730008326611
Collected $538 and $500 in bug bounties on Meta and HackerOne in a single day, using strix-claude-code plus ichat with Opus 4.8 under the hood. Not fully automated, but the agent did most of the heavy lifting on finding the reports.
@He1s_Sammy [Claude Code]
https://x.com/He1s_Sammy/status/2082436826057298078
Used Claude Code to build and open-source a real quantitative trading strategy around the GARCH volatility model, the kind pro quant desks use. It ships with a TradingView Pine Script indicator that runs the strategy live, step-by-step setup, and a reusable Claude Skill to streamline the workflow.
@himarre [Claude Code]
https://x.com/himarre/status/2082435526204752219
A Reddit user's now open-source AI job-search system for Claude Code scored 700+ applications and actually landed him a job. It scans multiple company career pages, rewrites the CV per job, and even fills application forms, with 14 skill modes, a Go terminal dashboard, ATS-optimized PDF generation via Playwright, and 45+ companies (Anthropic, OpenAI, ElevenLabs, Stripe) preconfigured.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2082454941806047492
A field report on keeping two Claude Code accounts alive without a ban: separate payment rails (a US Apple gift card on one, Google Play with a domestic Visa on the other), a dedicated ~$2 residential IP per account routed through Surge so Anthropic always sees the same exit, and never logging both on one machine. His takeaway from reading the code: it doesn't fingerprint hardware, only account and IP — so the machine barely matters.
@pashmerepat [OpenClaw]
https://x.com/pashmerepat/status/2082583847011799056
Months into building long-running persistent agents, he argues the next paradigm-shifting product after ChatGPT and Codex is the persistent personal/work agent, and the blocker is "reading the room" — deep theory of mind. Concretely, tracking what I know vs what my user knows vs what the room knows, plus compartmentalization (enterprise "tenting"). Harnesses like OpenClaw keep patching this at the app layer with permissions and workspace isolation, but he thinks that's futile as the final answer: discretion has to live in the model.
@manan [OpenClaw]
https://x.com/manan/status/2082297964098461933
Cludgy but working: OpenClaw plus Home Assistant plus Tessie. He plans a road trip over iMessage with OpenClaw, then tells it to schedule a 100% Tesla charge through Home Assistant (connected to the car via Tessie). He's built a bunch of similar automations that run quite well.
@shedntcare_ [OpenClaw]
https://x.com/shedntcare_/status/2082496501935374617
A closed-loop scheduling demo: OpenClaw plus Ling-3.0-flash checks your calendar, suggests open course slots, confirms your pick, and updates your schedule end to end — a small but complete "agent actually finishes the task" loop.
🗣 User Voice
User Voice
Cost and token efficiency is the loudest recurring complaint. Composio's benchmark showed the same task burning up to 30x more tokens in Claude Code than in Kimi Code (about $2.00 vs $0.22 per task), and @TheGeorgePu points out Claude Code has never once shown him what a session cost. People want a leaner harness and real token metering, not guesswork.
Reliability jumped from annoyance to adoption risk. A multi-hour, multi-day outage of Anthropic's models hammered everyone on the 29th, with cascading 529 errors; @Anto1nx got cut off mid-agent-loop, and users are half-joking that the Government tier stayed green while everyone else went red. The ask is simple: credit tokens back on API errors.
Model lock-in is being actively routed around. A whole wave of tools — free-claude-code, OmniRoute, Ollama routing — exists purely to keep the Claude Code interface while swapping in Kimi K3, DeepSeek, or local models. @JulianGoldieSEO frames it cleanly: choose the workflow you like, then swap in the best brain for the job. The harness is the moat, not the model.
Trust and verification keep surfacing. Agents say "done" when they aren't, so iFixAI-style auditors and self-check prompts spread fast; @Voxyz_ai's fix is a prompt block that forces the agent to re-read the original ask and list what's MISSING, UNASKED, and UNVERIFIED. People want the agent to prove it finished, not claim it.
OpenClaw is having an identity crisis. A large share of its mentions are literally "does anyone still use this?" as users migrate to Hermes, and @ericelliott_ goes further, calling it a security liability with tens of thousands of RCE-exposed "ZombieClaw" nodes. @pashmerepat's diagnosis lands: the missing piece — discretion, theory of mind — has to move into the model, not the app layer.
Cost and token efficiency is the loudest recurring complaint. Composio's benchmark showed the same task burning up to 30x more tokens in Claude Code than in Kimi Code (about $2.00 vs $0.22 per task), and @TheGeorgePu points out Claude Code has never once shown him what a session cost. People want a leaner harness and real token metering, not guesswork.
Reliability jumped from annoyance to adoption risk. A multi-hour, multi-day outage of Anthropic's models hammered everyone on the 29th, with cascading 529 errors; @Anto1nx got cut off mid-agent-loop, and users are half-joking that the Government tier stayed green while everyone else went red. The ask is simple: credit tokens back on API errors.
Model lock-in is being actively routed around. A whole wave of tools — free-claude-code, OmniRoute, Ollama routing — exists purely to keep the Claude Code interface while swapping in Kimi K3, DeepSeek, or local models. @JulianGoldieSEO frames it cleanly: choose the workflow you like, then swap in the best brain for the job. The harness is the moat, not the model.
Trust and verification keep surfacing. Agents say "done" when they aren't, so iFixAI-style auditors and self-check prompts spread fast; @Voxyz_ai's fix is a prompt block that forces the agent to re-read the original ask and list what's MISSING, UNASKED, and UNVERIFIED. People want the agent to prove it finished, not claim it.
OpenClaw is having an identity crisis. A large share of its mentions are literally "does anyone still use this?" as users migrate to Hermes, and @ericelliott_ goes further, calling it a security liability with tens of thousands of RCE-exposed "ZombieClaw" nodes. @pashmerepat's diagnosis lands: the missing piece — discretion, theory of mind — has to move into the model, not the app layer.
📡 Eco Products Radar
Eco Products Radar
Codex — the constant comparison point; the "switching from Claude Code" narrative runs through dozens of posts.
Cursor — still the default IDE-side agent people benchmark against.
Hermes (Nous Research) — the harness everyone is migrating OpenClaw workflows toward.
Buzz (Block/Jack Dorsey) — Slack-for-agents, open source, plugs into Claude Code/Codex/Goose; the week's most-hyped new stack.
Kimi K3 / Kimi Code — the cheap frontier-adjacent model people run under Claude Code to cut cost.
Higgsfield — the media/video engine behind the solo agency and UGC-ad workflows.
Obsidian — the note vault Claude Code repeatedly turns into a "second brain."
NotebookLM — offloaded heavy analysis so users burn zero Claude tokens.
Tailscale + tmux — the remote/persistent-session backbone for running agents from a phone.
iFixAI — the open-source agent-auditing tool that kept recurring across posts.
OpenRouter / Ollama — the routing layer letting people swap models under the same harness.
Codex — the constant comparison point; the "switching from Claude Code" narrative runs through dozens of posts.
Cursor — still the default IDE-side agent people benchmark against.
Hermes (Nous Research) — the harness everyone is migrating OpenClaw workflows toward.
Buzz (Block/Jack Dorsey) — Slack-for-agents, open source, plugs into Claude Code/Codex/Goose; the week's most-hyped new stack.
Kimi K3 / Kimi Code — the cheap frontier-adjacent model people run under Claude Code to cut cost.
Higgsfield — the media/video engine behind the solo agency and UGC-ad workflows.
Obsidian — the note vault Claude Code repeatedly turns into a "second brain."
NotebookLM — offloaded heavy analysis so users burn zero Claude tokens.
Tailscale + tmux — the remote/persistent-session backbone for running agents from a phone.
iFixAI — the open-source agent-auditing tool that kept recurring across posts.
OpenRouter / Ollama — the routing layer letting people swap models under the same harness.
Comments