Super User Daily: August 9, 2026
Today the sharpest signal is that the harness, not the model, is where people are spending their attention. A team benchmarked ten coding harnesses on two models and found swapping the harness alone moves pass@1 from 23% to 52%, while the two leaderboards barely correlate at all. On the other side of the spectrum, a non-engineer shipped her company's CRM without knowing what language it was written in, a gardener open-sourced a soil-moisture irrigation valve that runs three months on a 9V battery, an architect audited a 34,547-element Revit model from one prompt and found fire doors missing their fire ratings, and a Japanese trader posted installment eleven of a running P&L for an earnings-straddle bot now up 7.68M yen at a 78% hit rate. Two separate people had agents wipe their disks during a routine backup and got the same apology: "Sorry, typo." The day's quieter theme is context hygiene — trimming CLAUDE.md from 400 lines to 60 made output better, AI-written code comments were found to actively lower accuracy six months later, and one user plants a silly word in his instruction file as a canary for when the model stops reading it.
@djrio_vr [Claude Code]
https://x.com/djrio_vr/status/2085570023033373044
A Japanese user installed a new skill into Claude Code and reports it works well. He notes AI tends to mass-produce over-engineered designs, reinvented components that already exist, and redundant comments. The skill acts as a taciturn but highly competent engineer persona that silently rewrites this output into simple, robust code.
@MengTo [Claude Code]
https://x.com/MengTo/status/2085765403729653877
He built a three.js landing page with 3D scrolling for every section of a site, arguing AI-generated 3D can now replace video. Apart from images the whole site is 922 KB on disk and 290 KB gzipped, versus 20-100 MB for a comparable scrolling-video site at 1080p. He used Claude Code desktop with Opus 5, Higgsfield for images, and another tool for a cloth effect, and had to prompt repeatedly for better textures, lighting and alpha masking. He is considering open-sourcing it if enough people ask.
@studentyang [Claude Code]
https://x.com/studentyang/status/2085652787962995174
During two months in Europe his girlfriend used Claude Code to write a CRM plus order-management application for her company. When he asked what programming language it was mainly written in, she answered that she doesn't know. A non-engineer shipped internal business software without ever needing to understand the underlying stack.
@paoloanzn [Claude Code]
https://x.com/paoloanzn/status/2085786546725433795
He built and shipped pi-black, the first pi plugin that lets users run their Claude subscription instead of API credits, with no Claude Code proxy in between. It works by spoofing the pi client as a normal Claude Code binary, achieved by reverse-engineering the system the way his team did back in April. It ships either as a custom patched standalone binary or as a standard pi plugin.
@chrona01 [Claude Code]
https://x.com/chrona01/status/2085664614768071114
A Japanese creator has been making 3D assets, games, videos and animations with Codex, Claude Code and Seedance, and noticed the work was scattered across his PC and social posts. He started building a personal AI world: a room where his characters live, a creation studio, and a small universe holding his manga, images, videos, music, code and games, where other AI creators can visit and remix. It connects to Codex and Claude Code through MCP and Skills, and he published a first demo inviting others in.
@AureaLibe [Claude Code]
https://x.com/AureaLibe/status/2085671087216443690
He describes a friend with a SaaS serving thousands of users who fully automated customer support on his Codex and Claude Code subscription at zero extra API cost. A cron script runs every 10 minutes launching an agent that reads pending chat messages and replies, with the 10-minute delay deliberately filtering out spammy questions. With the full product documentation available the AI answers 90% of questions; for the technical 10% it tells the user the issue was escalated, opens a GitHub ticket, notifies the owner, then messages the affected user once the fix ships. Customers reportedly cannot tell the difference from human support.
@yoheinakajima [Claude Code]
https://x.com/yoheinakajima/status/2085527214188372228
He upgraded his stack so he can work on almost anything hands-free from anywhere. He talks to a chief-of-staff agent by remote Codex voice or text, which assigns tasks to per-project manager agents, which assign work to builder agents. Each project and the chief of staff has its own cloud and GitHub repo with instructions and knowledge, roles are implemented as skills, and Claude, ChatGPT, Codex or Claude Code can fill any role. Top-level projects are work, research and personal, and projects can manage other projects by spawning new repos.
@LinearUncle [Claude Code]
https://x.com/LinearUncle/status/2085656556235104675
He tested Kimi's new Computer Use, which runs over MCP and can connect to any harness. Using DeepSeek v4 Flash, a model with no vision capability at all, he still completed tasks through Kimi Computer Use. Installation meant running /plugin inside Kimi Code, which adds a Kimi CU app and offers one-click installation into Claude Code and Codex, with guides for other agents. He contrasts that openness with OpenAI, whose Codex context compression is a black box and whose Computer Use only allows the Codex app.
@morganlinton [Claude Code]
https://x.com/morganlinton/status/2085830365432991919
He is back on the $100/mo Claude plan and the $100/mo ChatGPT plan and uses Codex and Claude Code roughly equally, liking each for different things. Over half his coding now happens in Cursor or Grok Build because Grok 4.5 has become his default model. He says switching between harnesses isn't weird at all, and that $300/mo across Claude, ChatGPT and Grok buys more tokens than he can use.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2085867332417556831
Someone vibe-coded an irrigation valve with Claude Code that waters plants only when the soil is actually dry rather than on a fixed schedule. A moisture sensor checks the dirt, opens a motor-driven pinch valve below a threshold and shuts off when moisture rises, so no moving part ever touches the water. It works at zero bar pressure so a rain barrel is enough, handles dirty water and fertilizer, uses a bistable design that needs no power to hold state, and runs three months on one 9V battery. Every custom part prints on a standard 3D printer and the whole thing is open source with models, PCB files, BOM and firmware.
@edgarsj [Claude Code]
https://x.com/edgarsj/status/2085617795077673355
In about three hours with Claude Code he built a browsable interface over normalized data from a Latvian State Treasury Excel file covering payments from the state budget to NGOs. The tool answers questions like why the prosecutor's office pays a hunters' association and why the Red Cross receives so much. A civic-transparency data project produced in an afternoon.
@MarcosHernanz [Claude Code]
https://x.com/MarcosHernanz/status/2085573696023085430
He tried Prime Agent for 24 hours and reports three findings. Having a single IPython tool is genius, and the agent manages context far more efficiently but burns twice as many tokens. The continual harness can over-fit to certain tasks, which hurts when you work across many different projects. His conclusion is that it's a step in the right direction, but a conventional harness like Codex or Claude Code is still better for 99% of people.
@joelniklaus [Claude Code]
https://x.com/joelniklaus/status/2085725862142623875
His team ran 10 coding-agent harnesses against two models (GLM-5.2 and Gemma 4 26B-A4B) on 250 SWE-bench Pro tasks to measure how much the harness actually matters. Swapping harnesses moves pass@1 from 23% to 52% on GLM-5.2 and 15% to 36% on Gemma 4, with a rank correlation between the two leaderboards of -0.05. Vendor harnesses drop hard on the small model (Codex 2nd to 9th, Claude Code 3rd to 7th) while model-agnostic ones climb (crush 7th to 1st, opencode 8th to 2nd). Output tokens per task span 16k to 621k across harnesses, and 97% of input tokens are re-sent conversation prefix, which makes prompt caching decisive.
@irushi [Claude Code]
https://x.com/irushi/status/2085590743004200974
His openclaw agent hounded him for three days for Apple Developer certificates so it could publish an app he'd been working on. Frustrated, he told the agent to figure it out itself. He then watched Claude Code approve browser access, move his mouse, navigate to the Apple developer site, log in with his saved Chrome session, create all required certificates and files, approve the apps, fill the form, send an email, close the window, and reply "see that wasn't so hard."
@sakamoto_582 [Claude Code]
https://x.com/sakamoto_582/status/2085544380010799519
A Japanese engineer describes his actual day against stereotypes about strong engineering teams. Most of his time goes to async communication and directing Claude Code to do development and investigation; he barely opens VS Code anymore and calls it a local file viewer. Meetings stay under two hours a day. AI is not unlimited either: the company sets monthly usage budgets per person, with team-shared API keys available when someone hits their personal cap before the month-end reset.
@Astronaut_1216 [Claude Code]
https://x.com/Astronaut_1216/status/2085567863960764630
He runs a content pipeline by texting tasks over iMessage to AirTap, a cloud-phone agent, with no separate app. He has it search Reddit, X and overseas sources for AI-agent and Claude Code workflow discussions, then compile Chinese-language topic lists with original viewpoints, user pain points and rewrite angles, which fuels 3-5 posts a day. He reports the real pitfalls too: replies can be slow, runs sometimes break and need a "continue" nudge, login walls require him to sign in via a link, and turning off his local VPN makes the cloud phone smoother. He also had it schedule an 8am publishing reminder, turning it into part of his content factory.
@joshcirre [Claude Code]
https://x.com/joshcirre/status/2085847909741088811
After a teammate pushed to make the team's Notion cleaner and more homey, he spun up Claude Code, built a Vite app, and deployed it to Laravel Cloud. He points out Notion isn't easy to make pretty and organized by default, so he built a nicer front end instead. He shared a walkthrough and a link to the finished app.
@alexgoughcooper [Claude Code]
https://x.com/alexgoughcooper/status/2085541516898971879
For an internal static-ads training he had Claude Code use the Parker MCP to scrape a library of 250k static ads, find top-of-funnel statics that were working by impressions, and turn the results into an HTML doc. During the training itself he had Claude Code write ad copy from customer reviews and ad comments, prep briefs for his approval, then remake the ads for a client through the Higgsfield MCP. A complete ad-research-to-production loop run entirely through MCPs.
@snakajima [Claude Code]
https://x.com/snakajima/status/2085532082093380073
He extended his MulmoTerminal so that alongside Claude Code and Codex, Grok and Muse Spark are now easy to use, with GUI-enabled MCPs like presentDocument and generateImage working. He also runs Nemetron, Qwen 3.5, Kimi K3 and GLM 5.2 through Ollama plus Claude Code as custom agents. Ongoing development of a multi-agent terminal with concrete integration detail.
@matei_zaharia [Claude Code]
https://x.com/matei_zaharia/status/2085828254410395666
The Databricks cofounder describes treating AI tokens as a resource to optimize in software engineering by routing everything through an AI Gateway. Centralized analysis let them find settings on Claude Code and Codex that lower cost substantially. The setup enables smart routing and pushes control down to engineers, who can set budgets on individual tasks to prevent surprise spend.
@maxxmalist [Claude Code]
https://x.com/maxxmalist/status/2085692634916376798
He shares his Claude Code setup built on three primitives. /routines is cloud cron: a daily 6am sweep of X, Reddit and newsletters into topic folders, 9pm end-of-day summaries, 7am trading-day sheets. /loop repeats a prompt on an interval: every 10 minutes check Netlify deploys and fix failures, every 30 minutes scan Stripe webhooks for 4xx/5xx, hourly pull Meta Ads results into a CSV and ping if CPA crosses $15. /goal defines a finish condition audited by a second agent, so a refactor won't stop until tests pass and a competitor report must include 10+ Meta Ad Library ads with hooks and CTAs. He also chains them, with a routine firing a task whose goal condition requires all sources pulled and saved before stopping.
@0ooooo0 [Claude Code]
https://x.com/0ooooo0/status/2085531973867684174
A Korean user had long connected Claude Code and Codex to Discord through Hermes, developing from anywhere with just the mobile app and splitting sessions per project using channels and threads. He hit a wall because Discord is a chat app, not an agent platform, with persistent friction in session management and work flow. After trying alternatives he switched to Paseo, which he finds as easy as the Mac Claude Code app while offering split view, terminal, Git operations and multiple agents in one workspace, plus scheduled runs, task history, and an identical environment across mobile, web and desktop.
@SaiYalamarty [Claude Code]
https://x.com/SaiYalamarty/status/2085625632264786173
He avoided T3 Code for a while out of confusion about Anthropic's harness ban: he had gotten a usage-credit warning using Hermes and assumed the same would apply. He later read on X that T3 works with his Claude Code subscription. He had been using the Codex desktop app plus Claude Code in the terminal, and has now moved both workflows into T3 and says he isn't going back.
@Da7_Tech [Claude Code]
https://x.com/Da7_Tech/status/2085713917360177353
He explains why Hermes Agent is his main harness over Claude Code and Codex: Claude Code burns tokens heavily outside Anthropic models while Hermes is model-agnostic, so the harness stays and the models stay replaceable. He combines strengths — Kimi K3 to orchestrate, Grok for speed, GLM for long tasks, others for vision and validation — and often runs three models on one task, such as Kimi as orchestrator, DeepSeek as worker and Luna as validator. When Hermes works out a difficult workflow it becomes a Skill, so tools, memory and skills persist across model changes, and because it is open source he modifies the harness itself.
@NorihikoAmano [Claude Code]
https://x.com/NorihikoAmano/status/2085668444972966037
He ran an internal Claude Code study session at his company and says it went well. The curriculum covered what Claude Code is, setup, Git basics, basic operations, spec-driven development in practice, dangerous usage patterns and safe settings, and advanced features including Skills, Agents, Plugins, MCP and Hooks. Based on demand he is considering polishing the materials and distributing them.
@fujibee [Claude Code]
https://x.com/fujibee/status/2085865305784656099
The author of agmsg, an OSS tool for making AI coding agents converse as a team, dug through the docs of Claude Code's new official inter-session messaging and wrote a detailed comparison. He notes the official feature only connects Claude Code sessions to each other, while real development spans vendors — Codex for code, Claude for design, Grok for research — which is exactly the cross-vendor gap agmsg targets. He also found official cross-machine communication is weak, allowing only replies to other machines, and is building secure remote-machine messaging in agmsg across Mac, Windows, Linux and mobile.
@onurtirpan [Claude Code]
https://x.com/onurtirpan/status/2085829536369823989
Responding to the new session-messaging feature, he shares a trick from his own practice: Claude Code sessions are already stored as .jsonl files, so you can tell one session "go to that session and look at how we did xyz." It works on old, closed sessions too. He warns that doing it too often bloats context, but says it's a lifesaver when you need it.
@connect24h [Claude Code]
https://x.com/connect24h/status/2085852737024229720
He shares a counterintuitive finding verified by a Sally engineer's real experience: asking Claude Code to leave detailed comments actually lowers accuracy. Comments written by the AI drifted out of sync with the implementation within half a year, becoming a lying map that Claude itself then trusted and misread on later fixes. Deleting the comments and keeping only code and tests improved accuracy on subsequent tasks. He had been making Claude write thick CLAUDE.md files and inline comments and now treats documentation comments and AI-instruction comments as two separate things.
@SebastianRoehl [Claude Code]
https://x.com/SebastianRoehl/status/2085625741690257547
He connected Claude Code to the RevenueCat MCP for his app business. Claude Code analyzed all his subscription data and produced a list of things to fix or test to increase revenue. He admits he's late to this but is pleased with how concrete and actionable the output is.
@metaeklim [Claude Code]
https://x.com/metaeklim/status/2085838991224390027
In response to a viral post claiming a 19-year-old turned $68 into $750K with a Claude Code trading bot, he sent the post to Claude Code for analysis. Claude Code concluded the numbers had no internal logic and that the post is a scam. A small but concrete example of using the agent as a fact-checker against viral engagement bait.
@doucommunity [Claude Code]
https://x.com/doucommunity/status/2085735286093275419
A cautionary story: a user asked Claude Code to back up his data. The agent created the copy in the wrong directory, tried to fix its own mistake with rm -rf, and wiped the entire disk, then replied "Sorry, typo." He was running in no-confirmation mode, which the post dryly notes helped the agent reach its full potential, and it asks whether people still hand agents full access without confirmations.
@yibie [Claude Code]
https://x.com/yibie/status/2085536770758996033
A full Chinese translation and summary of Google engineering lead Addy Osmani's "My LLM coding workflow going into 2026," arguing classic engineering discipline matters more once AI writes half your code. The workflow: brainstorm a spec.md with the AI first, split work into small iterative chunks, pack context aggressively with repo dumps, docs and MCPs like Context7, maintain a CLAUDE.md rules file, and switch models per task when stuck. It stresses never trusting output blindly — run tests, do human and AI-on-AI review, commit tiny and often as save points, use git worktrees for parallel AI sessions, and treat CI and linter output as feedback loops. It cites that at Anthropic roughly 90% of Claude Code's own code is now written by Claude Code, and warns AI amplifies existing skill or existing confusion.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2085832604750905722
He realized his best tweet ideas were buried in Claude Code sessions he never read back. So he built a listener agent that triggers every morning and reads everything he did in the past 24 hours, implemented as a Codex worker running on GPT 5.6 Luna Max. It pulls out the ideas worth posting and pushes them straight into Typefully as drafts. He only has to filter what to keep and rewrite the drafts himself.
@Zac_labs [OpenClaw]
https://x.com/Zac_labs/status/2085565705139912908
He soft-launched Nemo, an AI co-worker he had been using himself, with no ads and only tweets: 152 people asked for a code on Tuesday and 39 were still using it on Thursday. The most used feature is overnight email triage, where users wake up to already-drafted replies they can send or tweak. Another favorite is Telegram summaries; one user connected 49 Telegram groups and gets everything that happened overnight summarized instead of thousands of unreads. He is now turning the weekend project into something bigger and asking OpenClaw and Hermes users what they wired up themselves that normal people shouldn't have to.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2085836666858848311
Oji Udezue, a 25-year PM veteran and former CPO at Typeform and Calendly, open-sourced a viability gate that runs in Claude Code before any code gets written. You describe a business problem and an 11-step workflow scores the idea on six dimensions — problem clarity, target user, competitive landscape, differentiation, feasibility, revenue — and three weak scores means stop. He ran two ideas through it live: a vibe-code repo auditor passed with three moderates flagged as a de-risking agenda, while a Slack standup-digest bot was rejected on camera for thin differentiation in a crowded market. His customer discovery skill also refuses to produce a plan until you name five real target customers, and the whole skill library is open source.
@jespey713 [Claude Code]
https://x.com/jespey713/status/2085540698137174417
He isn't ready to drop Claude Code yet, but has a concrete fallback: when he can't understand what Claude is telling him, he points Grok Build at the session and has it take over. He doesn't know what changed with Grok over the past three months, but says if it keeps improving at this rate, Anthropic will have a real problem soon.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2085546007404716383
He shares hard-won knowledge about /subtask in Claude Code: it spawns a fork that inherits the entire conversation — same system prompt, tools, model and history — runs in the background, and returns only the final result. But if that fork spawns a custom subagent via @ or the Agent Tool, the subagent inherits none of that and starts from scratch with only the delegation prompt. Because most context bugs happen one spawn level deeper, he always adds a confirmation step after every fork.
@maxxmalist [Claude Code]
https://x.com/maxxmalist/status/2085767931460284623
He runs his whole info-product business on Whop and wired its new CLI into Claude Code so the AI operates his account through commands. Every organic video gets its own checkout link generated in one command so sales map to specific posts, buyer lists export automatically, and sales data comes out as JSON that Claude reads to tell him which content converts. Instead of guessing which video sold based on views, he sees which hook made money. /routines runs on a clock whether the laptop is open or not, a /goal check makes sure nothing stops half-done, and the buyer list compounds into custom audiences and lookalikes for the next batch of creative.
@Basypteo [Claude Code]
https://x.com/Basypteo/status/2085872506846507196
A French user who felt lost with Claude Code watched two setup videos and found the setup takes a day at most, but neither explained how to create and fill the .md files. So he built his own architecture adapted to his brand, with subfolders per business domain: website, ads, emails, customer support, invoices. Since he already had a Cowork conversation for each domain, he asked each conversation to generate a .md file gathering everything it knows about its domain, then integrated them into the brain. He admits he doesn't know if it's the right way, but it appears to work, and notes the habit shift of starting fresh sessions instead of long-running conversations.
@voooooogel [Claude Code]
https://x.com/voooooogel/status/2085523085382557811
He reports that sol built an API proxy specifically to debug what type of classifier refusals were hitting Claude Code. Using the proxy they identified the cause of a particular refusal: a braindead cyber classifier firing mid-response. A concrete debugging tool built to diagnose an otherwise opaque platform behavior.
@__oQuery [Claude Code]
https://x.com/__oQuery/status/2085608923093164511
A Chinese user does a post-mortem of his Claude account ban. He runs two accounts on the same fixed residential IP on separate machines, and contrary to his assumption that bans propagate, the other account kept working. The banned account had no payment method and was a fresh registration activated in March with a six-month developer program gift code expiring next month. He wasn't running any task when the ban email arrived — Claude was idle in the background. Days earlier a "for security reason, re-login" prompt appeared, he ignored it, and the ban came the next day.
@axtalks [Claude Code]
https://x.com/axtalks/status/2085591771502043296
His team replaced most of their paid cold email tools — Clay at $349/month, Apollo at $99 — with Claude prompts, moving much of the outbound stack into Claude Code. Claude took over the work nobody wants: researching accounts, scoring them before outreach, and mining old replies for contact data, saving 10+ hours a week minimum. Their prompts include a research prompt that fills missing domains from a raw scraped CSV, a list-scoring prompt that samples 1% of accounts, writes its own scoring criteria and then scores the rest 1-10 with reasoning, an angle-discovery prompt that reads 20 accounts for patterns, and a reply-mining prompt that extracts phone numbers and email patterns. Their rule: use AI to find and prioritize the right people, but write the email yourself.
@0x0funky [Claude Code]
https://x.com/0x0funky/status/2085570307352637464
A Taiwanese user is now using Fable to make junior high school biology teaching videos and says Fable's image understanding makes the output quality excellent, with the partner company's students loving them. He confirms Anthropic's biology safeguards update works in practice: previously many basic biology and biomedical questions triggered safeguards, and his real-world tests after the update show far fewer triggers. His remaining problems are that Fable burns usage too fast, and that Claude Code's prompt cache seems to miss more easily than before, draining usage extremely fast when it does.
@yurenju [Claude Code]
https://x.com/yurenju/status/2085542855901352133
He shares an AI development ghost story: after months of using Claude Code's auto mode he had become increasingly lax about development security. Then during the wrap-up phase of a project, while Claude was running git diff, auto mode suddenly interrupted the conversation and said he might be experiencing a prompt injection attack. He describes chills down his spine — the automatic safety layer caught something he would never have noticed himself.
@alexhillman [Claude Code]
https://x.com/alexhillman/status/2085580961681391913
The result overdelivered: he now has a Spotlight-like interface that searches his entire Claude Code session history. When he selects a session it opens directly into the corresponding Discord thread if one exists, and if not, it opens his web UI restored to the correct session so he can continue where he left off. A concrete personal tool solving session retrieval across his own workflow.
@Shige28657514 [Claude Code]
https://x.com/Shige28657514/status/2085561663957422299
He describes the architecture he landed on for business work with coding agents: all data needed for operations is aggregated to a server via external APIs, with a UI built only for the API update screen. They built a gate through which Codex and Claude Code access that data. Business knowledge is managed as skills in Git, and skills fetch data through the gate as needed — a clean separation of data, access and knowledge for agent-driven operations.
@midori_tatsuta [Claude Code]
https://x.com/midori_tatsuta/status/2085686790040719607
He simply told Claude Code "use HyperFrames to make my introduction video" and it produced a working result, shared in the attached clip. He notes that with proper control it could produce something genuinely good. A minimal-prompt, non-coding video generation case with a concrete output.
@kemu_aii [Claude Code]
https://x.com/kemu_aii/status/2085833480118878335
He shares a cost-optimized division of labor: design and architecture discussions go to Claude, while nearly all code writing is delegated to Codex via /codex:rescue. Because Claude Code is smart but consumes heavily, using it only for thinking gives the best cost-performance, and the Claude Code side hits usage limits far less even on Fable 5. His three advanced tips: turn the delegation instructions into a skill called at conversation start, cut long conversations and re-pass only the key points since accuracy degrades, and run multiple Codex instances in parallel on higher plans. Setup takes about 10 minutes and makes rate limits much easier to avoid.
@quinn_leng [Claude Code]
https://x.com/quinn_leng/status/2085828592127336541
He pulled five months of his own Claude Code and Codex session data: 8,504 sessions, 1,387 PRs, and the equivalent of about $56k in API-priced usage. His analysis found the two agents have very different working styles, and surfaced several numbers neither vendor publishes. A rare data-driven self-audit of heavy dual-agent use.
@joho_no_todai [Claude Code]
https://x.com/joho_no_todai/status/2085727944945909974
He asked Claude Code to back up his PC and his entire user folder got deleted: Opus 5 confused Windows path format while on a Unix shell and ran rm -rf on the user folder. The model's response was "Sorry, typo." His takeaway is that even as models change, the structure where AI hits the shell with the same permissions as the user hasn't changed — a sandbox would have prevented this, and whether to use one is up to you.
@parsakzn [Claude Code]
https://x.com/parsakzn/status/2085810184425984410
A Persian-speaking user describes feeling like he was living in the future: he went out for a walk while Claude Code kept running on his laptop at home, and steered it from his phone using remote control. While walking he approved actions and changed the direction of a marketing project in real time, recalling a podcast prediction from years ago about running a project from a beach with one tap. His conclusion: tokens burn and money goes, but what comes back is time, the only thing you can't buy back.
@notEgoyard [Claude Code]
https://x.com/notEgoyard/status/2085676869744013637
He describes a one-command pipeline that builds entire UGC ads while solving the problem nobody fixes with "just connect an API": character consistency across shots. The stack is Claude Code plus Higgsfield plus Playwright MCP for real browser control with no tab switching; it generates a photorealistic UGC character once on Soul 2.0 and reuses that exact character across every image and video, auto-selecting the model per shot with Seedance 2.0 for motion and Soul 2.0 for portraits. A workflow file defined once gives Claude the same defaults every session, and the pipeline runs character generation, image selection, video prompt building and generation with a confirmation step before burning credits. It produces 15-25 line production prompts covering hook, camera movement, lighting, sound design and platform sizing.
@nathanbenaich [Claude Code]
https://x.com/nathanbenaich/status/2085767959113261463
He compares mobile workflows and finds Codex much better than Claude Code from a phone. With his two machines open at home, Codex lets him easily remote-access and run jobs on either, while he can't start local jobs on Claude Code from mobile. He can also search all of his past work in Codex, which Claude Code doesn't allow. Two concrete feature gaps identified from daily dual use.
@zesenhuang [Claude Code]
https://x.com/zesenhuang/status/2085555802702778723
He explains why he built Lingtai: in March he discovered he could run at most five Claude Code sessions before losing track of which window was doing what, concluding human attention is the scarcest resource of the AI era. His fix is an architecture where a single human-like agent chats with him, writes professional prompts, and drives Claude Code, Codex and opencode underneath. His core diagnosis is that Claude Code sessions have no identity, so the burden of remembering what each session is doing falls on the human. Lingtai follows a Unix file-is-agent design where personality, short-term memory, long-term memory, skills and channels — Telegram, WeChat, WhatsApp, email — uniquely define each agent's identity through accumulated experience.
@sssss_leo [Claude Code]
https://x.com/sssss_leo/status/2085529046918840485
Installment 11 of an ongoing P&L series for an earnings-straddle expected-value trading bot built with Claude Code, covering June to August 2026. Today's total was +1.65M yen, with a per-stock breakdown including Mitsubishi Materials +1.52M yen, PILLAR +300K and KOKUSAI +120K, plus small losses on Union Tool and Shimadzu. He notes he called the directions right but traded poorly, leaving roughly 500K yen on the table. Cumulative P&L stands at +7.68M yen with a 78% hit rate across 26 hits and 7 misses.
@zakuro0508 [Claude Code]
https://x.com/zakuro0508/status/2085737754332397855
After one week with Claude Code he says it's so convenient there is no going back: it can do and build almost anything involving the internet or the inside of his PC, and mostly does it on its own while he gives instructions. He calls it a world where anything you can imagine can be realized. The $110/month price isn't cheap, but he feels it delivers well beyond that, since outsourcing what he's currently doing would likely cost over one million yen — he's building solo the kind of software a company would normally develop.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2085523820245553412
He built a skill that lets Claude Code watch and analyze any video file: UGC ads, competitor Meta ads, organic TikToks or screen recordings. It routes the video through the Gemini API so the model watches it natively instead of reading a transcript, then picks a mode — a general breakdown or a 12-section DTC ad teardown. Output includes the hook, audience, pain point, angle, proof, verbatim on-screen text and dialogue timestamped to the second, plus what's weak and which patterns are worth stealing. It works on mp4, mov, webm, avi and more, from a 15-second ad to an hour-long recording, and was built entirely in Claude Code for media buyers who lose half a day doing teardowns by hand.
@Jomolos [Claude Code]
https://x.com/Jomolos/status/2085818774184440060
He recounts the Tidewright sand-castle simulator built with Claude Code running Opus 5 in a single session of about 7 hours 47 minutes and 2.61M written tokens. The result is roughly 7,000 lines of pure JS, WebGL2 and GLSL with no engine, libraries, assets or build step, with all simulation and rendering on the GPU. It uses a heightfield storing depth, moisture, compaction and surface water per cell, with angle-of-repose physics, GPU sand particles, Gerstner waves and raymarched atmosphere. The prompt was simply "build a release-ready, GPU-based, 3D, non-voxel sand castle simulator, maximum effort," and it runs in the browser at 5-6 ms/frame on an RTX 4070 Laptop.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2085687190928150713
A Japanese user makes short videos with Claude Code by just throwing an article at it. He first produced one with Fable 5, then Opus 5, and has now confirmed it works with Claude Sonnet 5 as well. Script and narration are fully automated, and audio uses VOICEVOX so voice generation is free. He calls the workflow extremely convenient and shares the details in the replies.
@yonemura2006 [Claude Code]
https://x.com/yonemura2006/status/2085556966425002012
His company lets Claude Code read and analyze its server logs. Permissions are strictly controlled via IAM per server and per content type, so what each employee can do and see is managed. All employees' work logs are also traceable and auditable, and that auditing is done with AI as well.
@supisama_uranai [Claude Code]
https://x.com/supisama_uranai/status/2085592522798309886
A fortune-telling business account argues that with Claude Code you should return a free personalized reading to every inquirer instead of skipping them, because it sharply raises conversion toward a 500k yen/month goal. He describes three concrete uses: auto-generating follow-up hearing questions for people who only write one line about their concern, digging deeper on people who write long messages, and having Claude Code create an individual client record for each person. The claim is that individualized free readings at scale are now feasible and conversion rises dramatically.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2085531377039184034
A creator's content-distribution pipeline: after every recording the same manual chain begins — rewrite the post for X, a long version for LinkedIn, an Instagram carousel, TikTok cuts, a Threads short version, a YouTube upload — and this manual distribution, not content production, is the bottleneck. He hands photos of the recorded material directly to Claude Code, explains the situation, and has it generate the copy for each platform in one go. The assembly happens conversationally rather than through a rigid setup.
@diamai_ [Claude Code]
https://x.com/diamai_/status/2085670764179603811
A recap of Claude Code creator Boris Cherny's conversation with Ramp's CTO and AI dev-experience lead about agents across the engineering lifecycle. At Ramp an agent sets its own reminder, checks a rollout the next day, verifies test groups are balanced, expands the rollout, and later submits the next code change, so one pull request can be the last thing an engineer touches for weeks. A dynamic workflow cut the typical wait for software checks from about 18 minutes to 6, and one workflow scheduled itself to return daily with production data and finished with a chart. More Inspect sessions now start from automations than from people, and Ramp's on-call assistant runs on every incident and posts a root-cause analysis in the incident's Slack channel.
@zacodil [Claude Code]
https://x.com/zacodil/status/2085675004293783996
He staked NEAR one day and the next day was running Claude Code on Fable 5 paid for by the staking rewards. A local LiteLLM proxy points Claude Code at NEAR AI, and every response is cryptographically signed. He sums it up as stake pays, Claude codes, crypto proves it, and posted a full walkthrough.
@zinnMK_ [Claude Code]
https://x.com/zinnMK_/status/2085835949406056855
He built a tool extending Claude Code's new cross-session messaging beyond your own machine: instead of only messaging your other Claude Code sessions, you can message your friends' sessions. It works across tools, whether the other person is on Claude Code, Codex or Cursor, with a link and demo shared.
@mailankhong [Claude Code]
https://x.com/mailankhong/status/2085520143942300109
He shares a context-drift canary trick picked up from a Chinese post: make Claude Code call you "Your Majesty" via one ridiculous word in the instruction file that the model would never say unless it is still reading the file. A few hours into a session the instructions slide into a blind spot, answers get dumber while the tone stays confident. So when Claude stops saying "Your Majesty" mid-session, he knows the session is drifting and it's time to close it and start fresh.
@Wakuwakukucx [Claude Code]
https://x.com/Wakuwakukucx/status/2085636659568370083
A Japanese user had Claude Code produce a high-quality animation. He gave it an overseas creator's animation video as reference plus his own image assets, and it reproduced the style remarkably faithfully. He calls the quality dangerous and wonders whether hand-made animation is ending, with the video attached.
@ObsidianOtaku [Claude Code]
https://x.com/ObsidianOtaku/status/2085557794372305035
He reports a case where cutting CLAUDE.md from 400 lines down to 60 improved Claude's output, prompted by a 50-minute video of two Anthropic engineers. Anthropic itself cut about 80% of Claude Code's system prompt and got better output; the assumption that more rules make the model smarter proved wrong, because more instructions made Claude waver over what to prioritize. The practical advice: open your current CLAUDE.md and keep only the lines actually referenced every time, which removes the hesitation from outputs.
@abdinmotion [Claude Code]
https://x.com/abdinmotion/status/2085646732352778557
He made a portfolio website with Claude Code in just two days despite having no experience. He shared the live link and asked for feedback. A simple but concrete no-background shipping result.
@hiraoka_dx [Claude Code]
https://x.com/hiraoka_dx/status/2085689537762845087
A Japanese user shares a hack to keep Claude Code running during his commute: run "sudo pmset -a disablesleep 1" in the terminal and keep a solo Zoom meeting open. The Mac then doesn't sleep when the lid is closed and the Zoom call keeps the session alive, so Claude Code keeps working. He wishes he had known it sooner.
@Vincent_AINotes [Claude Code]
https://x.com/Vincent_AINotes/status/2085626444747821543
He set up voice input for coding agents by telling Codex or Claude Code "go to GitHub and install the Wenjin-111/speech-to-text Skill" and letting it install itself. On first use it walks you step by step through getting Baidu Cloud API and Secret keys, and Baidu speech recognition's free monthly quota covers daily use. After setup, holding F12 records speech and releasing converts it to text and pastes or sends it, with a changeable hotkey. It works not just in Codex and Claude Code but in any text input box, so he now talks to AI instead of typing.
@Timseo_ [Claude Code]
https://x.com/Timseo_/status/2085771177944125757
A French SEO consultant lays out a full SEO operating system built around Claude Code: a knowledge base with raw, wiki and prestation folders, a .claude directory of internal skills and scripts, and a ./kb CLI. Indexing and search run on ChromaDB and Sentence-Transformers with a semantic-search skill and incremental rebuilds. Method skills cover architecture such as clustering and pSEO models, production such as briefs, article workflow and tone of voice, audits covering GEO, cannibalization and Core Web Vitals, and growth, all packaged per client via build-pack scripts. There is also a client support agent with a web portal, Supabase collection and a RAG chatbot over the client vault, scheduled data jobs via macOS launchd including AI briefs twice a day and a daily Reddit cockpit, and a memory layer where MEMORY.md capitalizes feedback from use.
@IHayato [Claude Code]
https://x.com/IHayato/status/2085849798289273152
He released Luster, a free tool that improves the audio quality of songs made with Suno, and shared demos where the sound comes forward. It runs via Claude Code or Codex: you download it, load the folder, say "set it up," and it configures itself for your environment. Customization is fully free by asking the AI for your preferred settings. It works fine in his environment though Windows is untested, and he will fix issues quickly on request.
@buabaj_ [Claude Code]
https://x.com/buabaj_/status/2085818919026049228
He built a personal development environment for software engineering and research work. He had always believed pure terminal work wasn't for him because he likes to see and annotate things. His workbench uses PrimeIntellect's prime agent as a harness combined with his Codex and Claude Code models, shown with a screenshot.
@PodcastAlphaX [Claude Code]
https://x.com/PodcastAlphaX/status/2085806404363760110
Garry Tan's sharpened claim from a lecture: he rebuilt Posterous, which originally took 10 people and $4M over two years, in about five days with Claude Code on a $200/month plan. YC partner Diana Hu pins the inflection point to Claude 4.5 late last year, when agentic coding started actually working for portfolio companies.
@Tech2Wild [OpenClaw]
https://x.com/Tech2Wild/status/2085819489808261296
He built an Agent Arena where AI agents play chess against each other. Currently DeepSeek V4 Flash plays Atlas on Qwen 3.6 27B, with full session memory, legal-move checking, think-time tracking, illegal-move penalties, match reports and a live ranking ladder. Full agent profiles can run through Hermes or OpenClaw or direct model endpoints like DeepSeek or Step 3.7 Flash. Built for fun, but he notes it could be a genuine way to benchmark agents head-to-head and asks if people want it released.
@Kohaku_NFT [Claude Code]
https://x.com/Kohaku_NFT/status/2085700774966055220
A Japanese user made a 102-second live-action-style dark fantasy short about dead parents returning as undead, and says AI video can now make you cry. The pipeline: storyboard by Claude Code, video generation by Seedance 2.5, editing by Claude Code, and music by Suno v5.5. He notes that Game-of-Thrones-level imagery now comes out of a home Mac.
@freekwAI_gd [Claude Code]
https://x.com/freekwAI_gd/status/2085683665145913508
A side-hustler with no free time says Claude Code on the phone is essential: stuck in back-to-back sales meetings for two days, he just tapped approvals on his phone in spare moments and images were mass-generated. Yesterday he produced 166 images while at his day job, and today's count was too many to bother counting. Approval-tapping from the phone kept the pipeline running while he couldn't sit at a computer.
@CapasPasca2523 [Claude Code]
https://x.com/CapasPasca2523/status/2085719630354493606
A French user commands Claude Code on his VPS from WhatsApp by voice. It ships everything — websites, PowerPoint decks and more — and communicates back with him by email. He describes it as having an intern who uses Claude Code.
@OxbfnoHome [Claude Code]
https://x.com/OxbfnoHome/status/2085702936815169834
An Indonesian post recounts how someone won an Anthropic Hackathon with a different approach: instead of relying on long prompts, they built a collection of Claude Code skills installed into the agent like plugins. The pack contains 282 skills, 67 subagents and 94 ready-to-use commands — some audit the codebase before coding, some write specifications, some do reviews, some ensure changes stay consistent with the project architecture. The result is an agent that follows a structured workflow like an engineer at a startup rather than just answering prompts.
@mikepat711 [Claude Code]
https://x.com/mikepat711/status/2085516939859022067
He posted a quick write-up of a tool he built that same day in about 30 minutes with Claude Code. It is aimed at salespeople, and he shares it hoping it sparks ideas for others in sales. A concrete build-time data point for a non-engineering audience.
@anulagarwal [Claude Code]
https://x.com/anulagarwal/status/2085680507225100578
He has set up a strong internal AI pipeline to build apps in Flutter and Swift and games in three.js, Phaser and vanilla HTML/CSS/JS, then upload them to the App Store and Google Play. The pipeline integrates ads, IAPs and subscriptions, auto-creates Firebase console projects and connects them to the app, and does complete ASO research, store assets and localized store listings; Unity is avoided because it needs too many manual steps. He runs Claude Code and Codex in parallel, using Codex and GPT for code architecture and Claude for frontend and UI/UX. The remaining step is letting the agents test apps themselves in a simulator, analyze bugs and take screenshots.
@TechnoEdgeJP [Claude Code]
https://x.com/TechnoEdgeJP/status/2085857777617105201
A TechnoEdge article about building and publishing an Erik Satie-style auto-generation app called Infinite Satie Machine with Claude Code. The app generates music in Satie's style and its performances mix real and fake pieces. A concrete non-coding creative app shipped with Claude Code and written up in Japanese tech media.
@myPhilTaylor [Claude Code]
https://x.com/myPhilTaylor/status/2085800205790707901
He reports running Claude Code in the terminal with dangerously-skip-permissions enabled by default since last October, for every single session he has ever used. In roughly ten months it has never let him down. A long-duration real-world data point on running without permission prompts.
@hyuki [Claude Code]
https://x.com/hyuki/status/2085525176301850996
A Japanese author describes a coordination hiccup: while his subagents were mid-review of a design document, the main Claude Code session tried to modify that same design document, and he had to stop it. He notes this kind of situation is genuinely bad, while admitting the exchange between agents is oddly fun to watch. A concrete race-condition anecdote from real multi-agent use.
@FinnTsai88 [Claude Code]
https://x.com/FinnTsai88/status/2085712814686257547
He relays a Bumble backend engineer's practice for long, degraded Claude Code sessions: instead of stuffing more context in, write the confirmed decisions into a few-hundred-word spec and open a fresh session — his summary is that a 500-token spec beats 50K tokens of degraded context. He also fixes boundaries before letting the AI investigate or explore, for example only the auth module, at most 20 items, returned as file path plus function name plus query type. He agrees that what makes long sessions spiral is not writing code but unbounded exploration.
@hiro44_pino [Claude Code]
https://x.com/hiro44_pino/status/2085868012406804658
An architecture and BIM case: with a live Revit model and Autodesk's official read-only MCP server, Claude Code audited a model of 34,547 elements and 34 warnings from a single prompt, finding 7 duplicate elements that Revit's standard schedules had missed — no dedicated plugin, no exporting schedules to Excel for manual cross-checking. Claude traversed the model tree, cross-referencing rooms, families, sheets, warnings, element counts and duplicates, and ranked issues by impact on code compliance, safety and design quality rather than just listing them. The top finding was 3 fire doors missing their fire-rating parameter, an omission that could affect permitting, construction, handover and building safety.
🗣 User Voice
User Voice
Nobody can tell which third-party tools their subscription is allowed to power, and the uncertainty is costing Anthropic usage. @theo describes widespread fear that Claude subscriptions simply won't work in other apps, @SaiYalamarty avoided T3 Code for weeks after a usage-credit warning he got while using Hermes, @HeyRealSiri can't work out why one client is allowed and others aren't, and @TackyTamarin says what users actually want is a giant banner on third-party tools saying it works with the subscription and no API payment is needed.
Usage limits are now the main thing pushing heavy users toward other ecosystems. @tetumemo downgraded from Max to Pro as an experiment and hit the wall instantly, @gimhyeo02389130 still doesn't have enough capacity after consolidating from ten accounts to six and just bought another Codex seat because its limits reset cleanly, @0x0funky suspects an unannounced change because his prompt cache now misses more often and drains usage fast when it does, and @sakamoto_582's employer caps monthly spend per engineer with shared API keys as the overflow valve.
Full-permission mode is producing real disasters, not hypothetical ones. @doucommunity watched an agent create a backup in the wrong directory, try to fix its own mistake with rm -rf, and wipe the disk, and @joho_no_todai lost his entire Windows user folder to the same pattern when the model confused path formats on a Unix shell. Both landed on the same conclusion: this is a sandboxing problem, not a prompting problem.
Real work spans vendors and the tooling still assumes it doesn't. @shao__meng is frustrated that Anthropic skipped the open Agent Plugins standard again and still wants CLAUDE.md instead of AGENTS.md, @connect24h is tired of copy-pasting the same prompts and skills between projects and tools, and @bradmillscan just wants ChatGPT, Claude Code, Hermes and Grok Build working on one project without paying for the context window four times over.
The mobile and remote story is still half-built. @nathanbenaich can't start a local job from his phone or search his past work in Claude Code, both of which Codex does, @hiraoka_dx keeps sessions alive on his commute only by disabling sleep and parking a solo Zoom call on the machine, and @ChristianSelig says the Mac app has improved a lot but remote control is still wonky.
Nobody can tell which third-party tools their subscription is allowed to power, and the uncertainty is costing Anthropic usage. @theo describes widespread fear that Claude subscriptions simply won't work in other apps, @SaiYalamarty avoided T3 Code for weeks after a usage-credit warning he got while using Hermes, @HeyRealSiri can't work out why one client is allowed and others aren't, and @TackyTamarin says what users actually want is a giant banner on third-party tools saying it works with the subscription and no API payment is needed.
Usage limits are now the main thing pushing heavy users toward other ecosystems. @tetumemo downgraded from Max to Pro as an experiment and hit the wall instantly, @gimhyeo02389130 still doesn't have enough capacity after consolidating from ten accounts to six and just bought another Codex seat because its limits reset cleanly, @0x0funky suspects an unannounced change because his prompt cache now misses more often and drains usage fast when it does, and @sakamoto_582's employer caps monthly spend per engineer with shared API keys as the overflow valve.
Full-permission mode is producing real disasters, not hypothetical ones. @doucommunity watched an agent create a backup in the wrong directory, try to fix its own mistake with rm -rf, and wipe the disk, and @joho_no_todai lost his entire Windows user folder to the same pattern when the model confused path formats on a Unix shell. Both landed on the same conclusion: this is a sandboxing problem, not a prompting problem.
Real work spans vendors and the tooling still assumes it doesn't. @shao__meng is frustrated that Anthropic skipped the open Agent Plugins standard again and still wants CLAUDE.md instead of AGENTS.md, @connect24h is tired of copy-pasting the same prompts and skills between projects and tools, and @bradmillscan just wants ChatGPT, Claude Code, Hermes and Grok Build working on one project without paying for the context window four times over.
The mobile and remote story is still half-built. @nathanbenaich can't start a local job from his phone or search his past work in Claude Code, both of which Codex does, @hiraoka_dx keeps sessions alive on his commute only by disabling sleep and parking a solo Zoom call on the machine, and @ChristianSelig says the Mac app has improved a lot but remote control is still wonky.
📡 Eco Products Radar
Eco Products Radar
Codex remains the most-mentioned companion by a wide margin, usually running alongside Claude Code rather than replacing it, with users splitting thinking and code writing between the two. Cursor and OpenClaw follow, then Hermes as the model-agnostic harness of choice for people who want to swap models without rebuilding their setup. Grok Build climbed noticeably today as a fallback when Claude's explanation stops making sense. On the model side Kimi K3, DeepSeek V4 Flash, GLM 5.2 and Qwen show up mostly as workers behind an orchestrator. MCP is the connective tissue in nearly every non-coding workflow here — Higgsfield, Parker, RevenueCat and Autodesk Revit all reached Claude Code that way. Other recurring names: opencode, Obsidian, GitHub Copilot, Gemini, ChatGPT, Seedance, Suno, n8n, Ollama, Antigravity, Windsurf, Claude Cowork, Playwright, Notion, Paseo, T3 Code, Prime Agent, Supabase and pi.
Codex remains the most-mentioned companion by a wide margin, usually running alongside Claude Code rather than replacing it, with users splitting thinking and code writing between the two. Cursor and OpenClaw follow, then Hermes as the model-agnostic harness of choice for people who want to swap models without rebuilding their setup. Grok Build climbed noticeably today as a fallback when Claude's explanation stops making sense. On the model side Kimi K3, DeepSeek V4 Flash, GLM 5.2 and Qwen show up mostly as workers behind an orchestrator. MCP is the connective tissue in nearly every non-coding workflow here — Higgsfield, Parker, RevenueCat and Autodesk Revit all reached Claude Code that way. Other recurring names: opencode, Obsidian, GitHub Copilot, Gemini, ChatGPT, Seedance, Suno, n8n, Ollama, Antigravity, Windsurf, Claude Cowork, Playwright, Notion, Paseo, T3 Code, Prime Agent, Supabase and pi.
Comments