Super User Daily: August 6, 2026
Trust became the story today. The most-discussed thread in the Claude Code world was not a workflow or a benchmark but a reverse-engineering find: hidden code in the closed-source binary that covertly flags China-linked environments through invisible prompt characters, landing on top of a reported wave of account bans. Meanwhile the actual work kept getting more physical and more commercial: a structural engineer using Claude Code to verify rebar anchorages in Tekla models, an oil and gas engineer building a full calculation system, a $100k-per-month Meta ads pipeline driven through the Claude Code API, and a solo founder replacing Peatix with a self-built ticket site. OpenClaw's smaller but sharper cohort keeps running Mac mini stacks and $10k MRR SEO operations. The countertrend worth watching: several posts about wrapping or replacing the big-lab harnesses entirely, either for trust reasons or because non-technical professionals need something that looks like the tools they already use.
@AYi_AInotes [Claude Code]
https://x.com/AYi_AInotes/status/2084522269745820010
A user shares an AGENTS.md distilled by a Vercel Next.js team engineer who reportedly burned about 60B tokens (roughly six figures in dollars) to arrive at 8 core rules for coding agents. The rules tell the agent to drop backward compatibility, pick the simplest implementation that meets current needs, get a minimal end-to-end version running first, prefer mature maintained libraries, and check existing dependencies before adding new ones. Dropped into the project root, the file is auto-read by Cursor, Claude Code, Codex, and Windsurf, and saves tokens by stopping the agent from over-engineering and rewriting. The author warns it is only for side projects: the no-backward-compatibility rule once nearly made an agent delete production database tables, almost losing $2,000 of data.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2084539348192829526
Japanese user made a short vertical-format drama entirely with AI: a single prompt generates a video complete with sound effects and dialogue. A 30-second generation costs about 1,400 yen, versus roughly 160,000 yen per episode even from a cheap Chinese studio, making AI production about 1/50 the cost at around 3,000 yen per episode. Editing was fully automated with Claude Code, which burned in the vertical titles, subtitles, sound effects, and end cards entirely via code. The video model was Seedance 2.5, which has an API, so the whole pipeline can be completely automated.
@dperry [OpenClaw]
https://x.com/dperry/status/2084722523909050700
A developer says he still loves coding and is testing OpenClaw. With it he built and shipped a free website that predicts how long you will live. A small but concrete output produced while trying out OpenClaw.
@gengdaJ [Claude Code]
https://x.com/gengdaJ/status/2084537270519164996
Chinese user open-sourced all 8 of the Skills he personally uses, his most complete release yet. They include a web research skill combining search, link downloads, bookmark export, and audio transcription; a shared memory system for Codex and Claude Code stored as editable, self-evolving Markdown; WeChat skills exporting chat logs, contacts, moments, and favorites with local data; a WeChat Official Account article batch scraper; a ChatGPT Pro research skill; an X Article uploader handling covers, images and tables; and a risky WeChat dual-instance skill. He also lists community skills he uses often, including business analysis, Feishu CLI, frontend taste skills, WeChat Reading notes, a PUA skill for agents, and an audit skill to vet skill installs.
@Shuhei_Ohno [Claude Code]
https://x.com/Shuhei_Ohno/status/2084461661927408017
A Japanese finance professional says stock price valuation is now almost fully automated with Claude Code, as is financial due diligence. He notes automation causes a strange problem: staff can produce the work but their knowledge does not keep up, so they cannot grow into reviewers and lose motivation. His solution is to have Claude Code handle the training too: it extracts key points from past deals to generate textbooks, for example on breakpoints for startups with multiple preferred share classes, extending into option models and Black-Scholes explanations. Anything the text does not clarify, staff simply ask Claude Code to explain, so staff development itself is automated.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2084727961417371875
User applies the LLM Wiki inside Obsidian concept (attributed to Andrej Karpathy) to run an AI video production knowledge base. He fed Claude Code hundreds of movies, ads, and anime references, which it ingested and cross-linked so any agent can navigate them. Whenever he makes a video with Seedance 2.5, the agent picks the right format, composition, and lens from the knowledge base. His entire video pipeline lives in an AI agent with access to Higgsfield's CLI, skills, and models for image, video, voice, and music.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2084447979914813716
Report that someone turned a bricked Spotify Car Thing into a physical Claude Code control panel. The device shows every session, handles permission approvals, and lets the owner manage agents from a dial. A concrete hardware hack repurposing dead consumer hardware into an agent-management console.
@vista8 [Claude Code]
https://x.com/vista8/status/2084637837371351222
Chinese user open-sourced a meta-skill: a skill that generates other skills. The original was built by his collaborator after analyzing the leaked Claude Code source code and integrating best practices from OpenAI and Gemini official skills, claiming better trigger rates, accuracy, and format validation than Anthropic's own skill-creator. He iterated on it by adding data sources that first search popular existing skills on skill-sharing sites, so generated skills build on proven work; he says tested output quality is very good. It also supports publishing to GitHub, generating npx install commands, and API leak checks, installable via npx skills add joeseesun/qiaomu-meta-skill.
@halilkutukk [Claude Code]
https://x.com/halilkutukk/status/2084556419915792770
Turkish user realized he had spent 8 months re-explaining the same things to Claude Code because every session starts from zero. He found a free plugin called ECC whose hooks record the fixes, resolved errors, and repeated flows during sessions, while a background Haiku agent extracts patterns without consuming the main context window. Learned behaviors get confidence scores between 0.3 and 0.9, stay project-specific unless the same pattern appears in two projects (then auto-promote to global), can be merged into permanent skills with /evolve, and can be exported to teammates. The package ships 67 agents, 279 skills, and a config security auditor, works across Claude Code, Cursor, Codex, OpenCode, Gemini, Zed, and Copilot, and his biggest gain is never explaining the same thing twice.
@AIEPCSolutions [OpenClaw]
https://x.com/AIEPCSolutions/status/2084617457516544140
An engineering firm automated the piping material engineer role in oil & gas with Apex_OS, a multi-agent system built over 4 months using Claude Code and deployed on their own servers. They first tested open-source frameworks OpenClaw and Hermes as a daily assistant and concluded they are not ready for heavy industry, since standard LLM wrappers hallucinate on precise ASME, API, ISO, and NACE standards work. Apex_OS agents are grounded in international standards, historical project databases, and company specs, and execute full deliverables: vendor document review with non-compliance flags and technical comment sheets, code-compliant material datasheets from scratch, and complete procurement requisitions. The operating model flipped from doers to reviewers: engineers no longer write reports but validate and sign off on AI output, letting teams handle multi-fold project loads without burnout.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2084537318879486017
Story of a 27-year-old who connected Claude Code to his 4,000+ note Obsidian vault built over six years, exploiting that Obsidian stores plain Markdown and Claude Code runs in the terminal. Left running overnight, it reconnected hundreds of orphaned notes into the knowledge graph, found contradictions between beliefs written years apart, surfaced forgotten ideas, drafted multiple essays from unnoticed patterns, and suggested projects from six years of recurring themes. It now runs every night, automatically organizing, tagging, and linking each new note, and every morning his daily note shows what he was thinking exactly a year ago and which assumptions no longer hold. The setup cost almost nothing: a folder path and one well-written CLAUDE.md, with no plugins or subscriptions.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2084648051042632032
Agency founder says they ran their $7M ARR agency's entire go-to-market from Claude Code and is giving the repo away. He calls Claude Code the most powerful GTM orchestrator but flags a major flaw: asked to do the same task 5 times, it finds a different way every time, so the fix is structure — skills, hooks, and dedicated agents so it executes the right way consistently. The repo contains 37 skills, 137 hooks, 11 dedicated agents, and 28 content prompts, with integrations for Apollo, Instantly, Gmail, Slack, and CRMs — the exact automations they used growing to $7M ARR. He frames it as open-source repos democratizing GTM engineering that used to require mastering Clay or hiring coders.
@IHayato [Claude Code]
https://x.com/IHayato/status/2084609271782920367
Japanese user had Claude Code generate an introduction video for Rondo in a single shot. His reaction: it is already almost good enough as-is, and a few more back-and-forth iterations will make it better, so he will polish it in spare time. A quick concrete example of one-shot video production from Claude Code.
@WuBlockchain [Claude Code]
https://x.com/WuBlockchain/status/2084497001618747800
Dragonfly managing partner Haseeb Qureshi said the Coldcard wallet incident shows AI is reshaping the economics of crypto security, with product safety increasingly determined by how much companies spend on AI-powered testing. He introduced a Cost of Discovery metric measuring how cheaply a frontier model can reproduce a vulnerability, estimating the Coldcard flaw at roughly $2. He noted a reported eight-minute Claude Code discovery of the flaw may have been influenced by web search, but a separate offline GLM test reproduced the issue in about 20 minutes.
@Cryptor_dot [Claude Code]
https://x.com/Cryptor_dot/status/2084653451506511884
User who has never used Blender or Unreal Engine and is not a game developer got a 3rd-person fighter running in Unreal Engine 5 after a few weeks of on-and-off work. The setup: Fable 5 at high effort as planner/orchestrator, Opus 5 at extra effort as the builder, Blender MCP for assets, and Unreal Engine MCP for scene, MetaHumans, and lighting. The UE MCP is still beta with too few tools, so half the pipeline is Claude writing Python around the gaps. The real bottleneck is that Claude Code cannot see the game — every animation pop, floating prop, and broken foot IK must be caught by eye and described in words; once agents can look at frames, he says weeks of work becomes a weekend.
@BuildingWaves [OpenClaw]
https://x.com/BuildingWaves/status/2084617004569493532
User reports SEO got him to $10k MRR, fully automated with OpenClaw. He is now debating whether to publish the playbook or turn it into a product. A concrete revenue result attributed to OpenClaw automation, though the workflow itself is not detailed.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2084608983558713492
Japanese user says Peatix is no longer needed: he built a full ticket sales site with Claude Code using Cloudflare, Stripe, and Resend. Buyers get a QR code on a passkey-login my-page, staff check people in by scanning, 24-hour limited links let tickets be handed to staff on the spot, canceled tickets have their QR revoked via Stripe webhook, and waitlisted people get automatic emails when cancellations occur, so he as organizer does nothing on the day. For his 3,480 yen x 15 ticket event, Peatix fees would be 4,200 yen versus only Stripe's 3.6% fee self-built, saving 2,377 yen — and the saving continues for every future event.
@FutureStacked [Claude Code]
https://x.com/FutureStacked/status/2084620713319632934
Report that a Reddit user built a rough version of GTA 6 with Claude Code. The build involved running 86 AI agents for 22 hours. A concrete, if second-hand, example of large-scale multi-agent Claude Code output.
@YannDecoopman [Claude Code]
https://x.com/YannDecoopman/status/2084720717535268947
French user shares a cheap long-running agent recipe: take DeepSeek V4 Flash (barely more expensive than free), use the /goal loop of Hermes, Codex, or Claude Code, teach it your task on a sample, and package it as a well-tuned skill with permissions validated so it never sits blocked waiting on you. Then give it a hefty /goal with a solid completion condition and let it run for hours or even days; he says a run costs about 5 bucks and does the work of several people at quality better than open source, Gemini, or Sonnet, though not Fable-level. Suggested uses include laborious SEO audits (log cleaning, GSC, cross-analysis of GA/GSC/logs), building research corpora, website construction (good at code, weak on content), and clearing project backlogs, with the agent optionally asking Fable for guidance to avoid drift.
@Zac_labs [OpenClaw]
https://x.com/Zac_labs/status/2084531260849086721
Founder says he has been living in the Maldives for months while having his most productive stretch in years: reading at 6am, diving with sharks and mantas by 7am, and back by noon with most work already done, needing only review and approval. His agents keep up with 40 Telegram groups, five Slack workspaces, X, LinkedIn, and an overflowing inbox, and his team says he replies faster with better context. He started with OpenClaw, Hermes, and NanoClaw but wanted something easier to manage, less technical, and lower-maintenance, so he built his own agent coworker that friends tried and refused to give back. He is now turning it into a product with a 10-person private beta before launch.
@pulmencr [Claude Code]
https://x.com/pulmencr/status/2084656223316500732
Post about a developer who built an open-source tool exposing how fast Claude Code burns the context window: by message 20 a single agent session can hit a hard rate-limit wall without warning. Tracking prompt costs in real time showed a basic codebase summary costs under a cent while a simple UI refactor spikes to 250 tokens per minute. The tool calculates your exact burn rate before you get locked out for 5 hours. The framing: context engineering is not just prompt quality but aggressive token budget control — whoever fills the window degrades the model.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2084478637504594237
Japanese user made a company introduction video with Claude Code, Remotion, and fal, roughly from a single prompt. He notes that placing just one such video on a homepage greatly increases the information a company can convey, and guesses many companies would want one.
@claudecode84 [Claude Code]
https://x.com/claudecode84/status/2084570815107068226
Japanese post detailing Claude Code autonomously testing iOS apps: point it at the simulator, say test everything, and it traverses every screen, feature, and flow. It reads the iOS accessibility tree to recognize buttons, forms, labels, and structure by meaning rather than coordinates, so the same tests work across iPhone SE, Pro, and iPad and survive UI changes, with screenshot analysis as a hybrid fallback for custom UI at roughly 10-50 tokens per screen. In a demo it toured an entire app in about 8 minutes, found bugs the developer had missed, and auto-generated debug log analysis and structured reports; OSS like claude-mobile-ios-testing and ios-simulator-skill exists, though test flows must be defined in CLAUDE.md so maintenance is not zero. The author has driven the iOS simulator via a Claude Code skill himself and says the accessibility tree precision exceeded his expectations.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2084762735229685784
Developer whose sol-advisor repo reached 1,300+ GitHub stars in 4 days (aimed at Codex users) released fable-advisor for Claude Code users. It uses the same pattern: Opus 5 as orchestrator, Fable 5 as high-complexity implementer, GPT-5.6 Luna Max as routine implementer, and Fable 5 as advisor. He says having a model as capable and efficient as Luna Max is incredible, especially combined with Opus and Fable.
@LandonHypixel [Claude Code]
https://x.com/LandonHypixel/status/2084456116809785854
Developer built and open-sourced Aliax, a Mac app for managing multiple accounts across AI coding tools. It supports Codex, Claude Code, and Cursor, showing all usage in one place with detailed per-account stats. Switching accounts takes one click, and every open terminal switches automatically along with each coding app, including T3 code. It also manages plans and renewal dates in one place, with the stated goal of making it easy to run, say, three $20 accounts.
@itsdizcorvus [OpenClaw]
https://x.com/itsdizcorvus/status/2084563214923338055
Indonesian freelancer says that on every client website job he uses the superpowers skill with his AI agents, whether in Antigravity, OpenCode, OpenClaw, or Hermes. The output is consistently neat and professional, with complete plans and implementation plans covering framework choice, the site or app's purpose, structure, and more. He shares the GitHub link for others to try it themselves.
@yoheinakajima [Claude Code]
https://x.com/yoheinakajima/status/2084647915021336680
Yohei Nakajima finally set up shared memory across ChatGPT, Codex, Claude, and Claude Code, so he can ask any one of them what he has worked on recently across all four. For Codex and Claude Code he uses local history, while for ChatGPT and Claude he uses a scheduled recap skill that regularly sends summaries into the shared memory.
@L_go_mrk [Claude Code]
https://x.com/L_go_mrk/status/2084587902118318576
Japanese post arguing that Claude Code tokens melt away mainly because the AI re-reads the whole codebase on every edit, and recommending the OSS code-review-graph to stop it. The tool turns the entire codebase into a graph of functions and dependencies, automatically tracing from a single function change to its callers, dependent files, and tests, and lets the AI read only files within the change's impact zone. Per measurements in the README, token consumption per question drops by a median of about 65x and up to 376x (on fastapi), and it runs fully locally so code never leaves the machine.
@4nt1p4tt3rn [OpenClaw]
https://x.com/4nt1p4tt3rn/status/2084629610239811904
User built a personal assistant agent that runs within Claude Code, backed by Honcho. It has its own email account, access to his NAS and its docker containers, his routers and switches, his other computers, Slack, and Obsidian; it manages his calendar, tracks work accomplishments and deliverables, sends reminders, works on GitHub repos with him, helps frame PR and issue responses, consults on standards-body interactions and 3D printing, tracks assets, does financial consulting, and helps hunt for real estate. It has its own machine token for Bitwarden where all its credentials (logins, API keys) are stored. He says productivity and task completion improved dramatically, and notes he tried Hermes and OpenClaw with both local and cloud models but neither handled even simple tool use, like setting up a cronjob, correctly.
@IAmAaronWill [Claude Code]
https://x.com/IAmAaronWill/status/2084619916255043689
Solo operator says his main job right now is just checking whether Claude Code runs. Clients get onboarded, follow-ups go out, and leads get scored while he is asleep, leaving him only actual client work. He says he built himself out of his own business and it somehow runs better without him.
@kevskgs [OpenClaw]
https://x.com/kevskgs/status/2084671193018511819
Developer introduces local-search, a small open-source Rust CLI that turns Chrome or Chromium on your machine into structured web search for coding agents — Claude Code, Codex, Cursor, OpenClaw, or any shell-capable agent — with no hosted API key, metered search account, or agent-specific SDK. It searches Google, Bing, Brave, or DuckDuckGo for free. For three results it returned 309 visible tokens versus 8,760.5 from a rendered search-page snapshot, 96.5% less context. In a matched 24-request provider run it used 53.4 normalized tokens per result, recorded 148.7 ms median latency, fulfilled 24/24 requested depths, and spent $0.
@reoreo_1_ [Claude Code]
https://x.com/reoreo_1_/status/2084476375180984684
A Japanese user implemented automatic video uploading to YouTube with Claude Code. They report videos were being scheduled and posted literally while they slept. As a result their channel smoothly passed 100,000 subscribers, earning a second silver play button. They say what they did is simple and anyone can do it.
@Argona0x [Claude Code]
https://x.com/Argona0x/status/2084678743935647825
Describes career-ops, an open-source job-hunting system that runs inside Claude Code and other coding agents. A man in Spain pointed it at 740 job listings, it told him which 68 deserved an application, and twelve interviews later he signed as Head of AI, a story written up by Business Insider. The tool scores roles 1.0 to 5.0 across five weighted dimensions, flags ghost jobs and scams, rewrites your CV for the specific listing, sweeps 100+ company career pages, identifies the hiring manager, and walks contracts clause by clause. It refuses to recommend roles under 4.0 and never presses send. It has 62,731 GitHub stars, is MIT licensed, installs with one npx command, and costs nothing beyond a $20 Claude subscription.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2084442361665388770
A 13-minute video walkthrough of making Meta static ads with Claude Code, Higgsfield, and Figma. Higgsfield generates backgrounds with zero text baked in, while Claude builds a Figma template with brand variables, imports the logo as real vectors bound to a theme variable, and keeps headline, body, and CTA button as normal Figma layers. Claude fires all six image prompts at once instead of one by one and runs a built-in QA pass before anything lands in Figma. Changing one variable updates every ad, and the Claude skill used is free.
@DanielGlejzner [Claude Code]
https://x.com/DanielGlejzner/status/2084697623076753583
A developer tested raw model output by creating a simple Angular 22 app with Claude Code and Opus 5 from a single prompt stating some good practices, deliberately using no skills or extra context. The result was poor: effects everywhere, 182 tests that brought almost no value, logic in constructors, subscriptions, ngOnInit, no splitting of functionality per feature, and Tailwind crammed in for no clear reason. His conclusion: these models are terrible without context.
@harumak_11 [Claude Code]
https://x.com/harumak_11/status/2084484722143478205
Japanese summary of an article on refactoring as a way to save AI tokens. The author built a roughly 150k-line Rust application using agents like Claude Code and Cursor, and the data access layer had ballooned into a single file of over 17,000 lines. After each refactoring stage they ran the identical change task and measured consumption: input tokens for the same task fell 83 percent, from 159k to 27k, because the agent no longer needed to read unnecessary code, while output tokens barely changed. Claude could not judge the appropriate refactoring approach by itself and needed active human guidance, and applying the refactoring took about 8 hours with problems like indentation confusion.
@sentient_agency [Claude Code]
https://x.com/sentient_agency/status/2084572164523720736
Describes Deep Researcher Agent, an open-source agent that runs deep learning experiments 24/7 from a research goal given in one Markdown file. It reads previous results, plans the next experiment, modifies code and configs, runs a quick test, launches training on your GPU, monitors logs, compares against the baseline, and decides what to try next, then repeats while you sleep. The creator reports 500+ autonomous experiment cycles, 30+ days of continuous operation, 4 projects managed across 4 GPU servers, and up to 52 percent improvement over a baseline. Because the LLM sleeps while the model trains and only wakes after training finishes, average LLM cost is around $0.08 per 24-hour cycle. It works with Claude Code, Codex, OpenAI-compatible APIs, remote GPU servers, and Slurm clusters, and is fully open source under Apache 2.0.
@0xDanXbt [Claude Code]
https://x.com/0xDanXbt/status/2084570783934726591
A user reports the first time Claude Code denied their request, saying what they were doing was against API usage policy when it was not. They switched to Codex for the first time and it completed the task without any flaw.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2084614478302162972
SimonHoiberg says ads got easy with his AI team: HeyGen generates the UGC ads and OpenClaw manages the ad accounts. The AI team generates variations, tests different hooks, measures results, and keeps going until something sticks. He supervises and oversees but says the process is otherwise 95 percent agentic, and asks developers who hate marketing why they are not doing this.
@notEgoyard [Claude Code]
https://x.com/notEgoyard/status/2084595990514815020
Describes compressing 4-8 weeks of video production into one Claude Code session with a different stack: Claude Code writes the script, reasons about structure, and orchestrates the pipeline; HeyGen MCP handles avatar generation and voice cloning with 230+ avatars and 175+ languages lip-synced; HyperFrames, HeyGen's open-source HTML-to-video renderer, gives deterministic output. The part most setups skip is a design.md file with actual brand colors, fonts, and logo usage that Claude reads before generating anything, so overlays and motion graphics match the brand instead of generic templates. Real output cited: a 45-second explainer with an executive avatar and brand-correct graphics, translated into any of 175 languages with a single follow-up message, with zero timeline editor touched.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2084476721663778896
After a day of brief power outages killed his local agent sessions, this developer built power_failure_resumer (pfr) to automate the repetitive work of reconnecting them. It detects a cluster of sessions that all ended around the same time, opens a Ghostty tab for each, resumes the Claude Code or Codex session, and ensures the same model and effort level are selected as before. It also detects if you use his MCP Agent Mail project and makes that the first tab, and deliberately never connects to his headless ntm orchestration sessions. It is implemented as a roughly 1,200-line Bash script with Python helpers and AppleScript, tested in his setup, and published on GitHub.
@itsharmanjot [Claude Code]
https://x.com/itsharmanjot/status/2084576620359438581
Describes an ADHD-friendly skill that makes Claude Code responses direct: it stops Claude from padding replies with warm-up phrases and gives the answer first. Ask how to fix a bug and you get the command to run, then numbered steps, then the one thing to do next if it fails, instead of three paragraphs of preamble. The skill enforces 10 simple rules, including lead with the action, number every multi-step task, keep lists to 5 items max, give time estimates in real minutes, and no preamble or recap. It is one MIT-licensed file that works with Claude Code and Codex and can be forked and tuned to your preferences.
@ai_300 [Claude Code]
https://x.com/ai_300/status/2084459309073653814
Japanese post arguing Claude Code has upended web production for landing pages, using a pool company LP as the concrete example: scrolling fills an on-screen pool with water, lighting shifts from day to evening, the water surface reflects, and WebGL drives the visual effects, so users pseudo-experience the finished product instead of reading descriptions. The author's division of labor is Claude for effect design, Claude Code for coding, Cursor for debugging and fixes, and GPT for sales copy. They argue the difference between earning 200k and 1M yen a month is not Photoshop skill but what you hand to AI and where you spend your own time.
@xeneks [OpenClaw]
https://x.com/xeneks/status/2084460037314052569
A user has their OpenClaw agent send them fishing reports via Telegram, because the agent is easy to control programmatically. Telegram is also how they chat with their AI agent when away from the computer. They joke that apparently someone thinks their fishing reports are a national security threat.
@drbarnard [Claude Code]
https://x.com/drbarnard/status/2084444802460823866
An iOS developer describes tiring of his Hermes Agent iOS development setup: everything takes too long, he cannot iterate directly the way he can by marking up the simulator in Codex, and having the Hermes brain manage itself via Codex induces the worst busywork tendencies. His old setup of jumping back and forth between Claude Code and Codex felt too clunky, but he still wants the adversarial loop between the two best models. He is weighing working in Codex and having it run Claude Code for planning, code review, and architectural review, versus working in Claude Code and using Codex as an employee for brute-force work.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2084487487951339606
This developer shares a prompt he uses dozens of times a day across all his dev machines, historically with Claude Code and recently also with Grok Build. It leverages his repo_updater project and a /ru-multi-repo-workflow skill to commit all changed files in every project under /dp in logical groupings with super-detailed commit messages, push, then pull from all repos without losing useful work. The prompt encodes safety rules: never commit obviously destructive things, defer to remote unless it would wipe out useful work, and always aim for the best-of-all-worlds version of each code file. It keeps his many interconnected repos harmonized across machines and GitHub, so a fix made on one machine eventually merges everywhere.
@EWTracker [Claude Code]
https://x.com/EWTracker/status/2084755289396416899
A trader says he paid $100 a month for Anthropic's Claude Code for the first six months of the year. In July he ran a dual process pitting Claude against Grok, and now exclusively uses Grok Build 4.5. He pays nothing extra per month since he just uses the tokens included with his X Premium+ subscription, which he now considers a bargain.
@ayushtweetshere [Claude Code]
https://x.com/ayushtweetshere/status/2084619613044879735
The author vibe-coded their own social listening tool with 2 prompts in Claude Code, replacing market tools that cost $100+ per month. It listens for keywords and competitor mentions on Reddit, X, LinkedIn, YouTube, forums, and news, synthesizes and filters out noise, and even drafts a contextual reply in one click linking back to the original post so they can promote their product Voibe. It uses direct APIs for listening and the OpenAI API for synthesis and replies, costing about $10 a month overall. They would never sell it because of security concerns, and the next step is building a Canny alternative for collecting customer feedback. Their takeaway: SaaS is not dead, but your product should not be replicable in 2 Claude prompts.
@kohack_v [Claude Code]
https://x.com/kohack_v/status/2084461607313346600
A Japanese maker built a tool that prints TEPRA label tape from Discord. They adjusted it with Claude Code so that anyone can use it and published it on GitHub for free use.
@insomnia_vip [Claude Code]
https://x.com/insomnia_vip/status/2084707738248945807
Story of a 32-year-old who built an AI system that produces UGC ads while he sleeps by connecting Claude Code with Arcads. Claude generates a complete character sheet, starting with a 6-view character reference with consistent facial features, outfits, and poses, then automatically builds a cinematic video prompt, places the AI creator into realistic environments, and launches realistic UGC videos using Seedance 2.0 and NanoBanana 2.0, combining product photos with virtual actors. The post argues brands spend thousands per month on UGC content while one such pipeline can create unlimited ad variations and launch entire campaigns without a production team.
@Yannlce [Claude Code]
https://x.com/Yannlce/status/2084672655483932680
A concrete how-to for making ecommerce ads: download the ad you want to copy, ask Claude Code to extract the exact structure from the transcript and a few frames, give it your product image and have it adapt the script to your brand, then ask for a video prompt for Seedance 2.0 to remake the video with the same structure but about your product. Take the prompt to any model provider that serves Seedance 2.0 (FAL, Argil, Replicate) with your product image. Since most models currently limit generation to 15 seconds in one shot, longer ads need two prompts plus some editing afterward.
@QuantumTransf [Claude Code]
https://x.com/QuantumTransf/status/2084604936852324690
The maintainer of the obelisk project reports receiving many good-quality PRs that were hard to merge in one pass, not because of contributors but because he struggled to quantify the standards in his head into principles contributors could follow in contributing.md. Obelisk, infrastructure that makes historical sessions queryable, ultimately helped him do exactly that, which he found surprising, and he points to pull request #27 as the final result. He plans an article on what such session-query infrastructure can extend into, and adds that the lody GUI is good enough that he no longer needs to watch Claude Code's TUI panel and it works together with obelisk.
@3chhe [Claude Code]
https://x.com/3chhe/status/2084515125596508457
A Japanese deep-read of qm, the OSS released by Y Combinator's software team, framed as an organization-level agent platform rather than a Slack bot: each employee gets an isolated workspace with separated memory, files, permissions, credential scopes, cron, web apps, and a persistent sandbox, while still collaborating with the agent in Slack channels and projects. The author highlights four design points: separating personal and shared scopes in one system, harness-agnosticism (Pi, OpenCode, Codex, Claude Code all drive the same core), layered safety where Strict/Auto/Dangerous postures coexist with hard refusal rules for destructive operations, and a contribution model where humans write text proposals and ADRs while maintainers implement. They connect this to their own practice of shifting internal development to humans defining requirements and acceptance criteria, AI implementing, and a different reviewer verifying. They also note qm honestly states it is experimental and no substitute for per-deployment security review.
@shion_takk [Claude Code]
https://x.com/shion_takk/status/2084484549703266477
A Japanese architecture practitioner explains why they are building a Rhino add-on wrapper around AI: real-world practitioners with brutal construction workloads cannot realistically adopt Claude Code or Codex on their own, but installing an add-on into familiar Rhino that just shows a side chat lowers the bar. Packaging it as an add-on lets them embed their own skills and workflows, safety design such as file-saving rules and hooks, and collect work logs in a common format that could become organizational knowledge. They acknowledge that developing and distributing an add-on means a thorny road of bug fixes, skill updates, and new Rhino version support, but are taking the challenge anyway.
@Im_IrushiK [Claude Code]
https://x.com/Im_IrushiK/status/2084635327810171037
A user reports that Claude Code randomly started outputting Kimi K2 Thinking details while /model still showed Sonnet 5. They ask whether it is prompt injection, a context leak, a model routing bug, or just another hallucination from a nerfed Claude model.
@nakagomeco [Claude Code]
https://x.com/nakagomeco/status/2084675144489095188
A Tokyo steel-frame fabricator has Claude Code operate Tekla Structures through the Tekla Open API. It first modeled the foundation structure from structural drawings, generating piles, footings, column bases, ground beams, and slabs down to main bars, stirrups, web reinforcement, and pile-head reinforcement. It then cross-checked the rebar against the drawings by extracting the actual rebar coordinates from the model and overlaying beam section cuts with the drawing sections: bar counts, diameters, materials, stirrup pitch, and cover all matched the drawings, but it detected that the L2 (40d) anchorage of main bars was missing in all 24 locations and that some beams were missing two bottom bars at the center. The author concludes that having humans count rebar one by one against drawings is work that can now be left to AI.
@developedbyed [Claude Code]
https://x.com/developedbyed/status/2084714960098132303
A YouTuber says he has been banned for 5 days with a warning plus a strike over a sponsored video. The video is simply him using Claude Code to create internal tooling for himself, such as a pixel art converter and image format converters he uses for making his game, which he calls the most normal video you could imagine. He cannot share it since it is a sponsored video meant for YouTube and he is still in talks with the sponsor about what to do.
@didac_email [OpenClaw]
https://x.com/didac_email/status/2084594678079406259
A Spanish-language post detailing a concrete OpenClaw agent stack running on a Mac mini: persistent memory in Markdown plus SQLite/embeddings to index context. The Google Sheets API via gog serves as the CRM, Meta Graph API/MCP handles ads, the Calendly API manages the calendar, the Fathom API pulls call transcripts, cron jobs automate syncs and reports, and the local filesystem stores knowledge, raw data, and audits. The author says this info will be useless to 99 percent of people, but offers DMs to anyone it helps.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2084677763202728322
A founder made a /pm skill that answers 'what should I work on next?' across every business he runs. It reads one kanban per business in whatever tool each one lives in (Notion, GitHub, Linear, or a markdown file), runs an Eisenhower pass over everything marked Ready, enforces WIP limits, and hands back one answer rather than a list. It is part of Maker Skills, 18 free open-source skills for founders and operators, installable in Claude Code via the plugin marketplace.
@_vmlops [Claude Code]
https://x.com/_vmlops/status/2084596234405503189
Matt Pocock open-sourced his entire agent skills setup, the .agents directory he uses daily for real engineering work. It includes /grill-me and /grill-with-docs to interrogate the agent's plan before writing code and kill misalignment, /tdd to enforce red-green-refactor loops so agents get real feedback, /diagnosing-bugs for a structured reproduce-minimise-hypothesise-instrument-fix loop, /improve-codebase-architecture which scans for entropy and generates a visual report before refactoring, and a CONTEXT.md pattern that builds a shared domain language so the agent stops using 20 words where 1 will do. It can be installed as an auto-updating Claude Code plugin or forked directly.
@techNmak [Claude Code]
https://x.com/techNmak/status/2084569794486186116
Describes 'ADHD for Claude Code,' a skill that changes how Claude Code and other coding agents (Codex, Cursor, Antigravity, Cline, Gemini CLI, Windsurf, and about 50 others) reason through a problem: instead of one thread thinking start to finish, it spins up six isolated threads that cannot see each other, each pushed into a different way of looking at the problem, and only afterward a separate pass scores everything, flags traps, and digs deeper on what held up. It targets the familiar anchoring failure where asking for a few different approaches yields one idea wearing five outfits, which tree-of-thought does not fix because branches share memory. With full isolation during generation, trap detection reportedly went from 1.8 to 9.5 against a single-shot baseline on the same model. It is 100 percent open source with a GitHub repo.
@onaka_yuruyuri [Claude Code]
https://x.com/onaka_yuruyuri/status/2084542710564954384
A Japanese user immediately started using the Artifacts feature inside Claude Code after seeing it explained at a conference talk. Previously they had the twists and turns of an implementation explained at the end as markdown, essentially PR description text, for review. Now Claude explains it presentation-style in HTML, which they find dramatically easier to understand.
@itsivanfalco [Claude Code]
https://x.com/itsivanfalco/status/2084723554126635269
A B2B marketer putting about $100k a month into Meta describes their ABM machine. They pull two CRM lists (Tier 1-4 target accounts and every contact on closed-won deals), then enrich with Prospeo and FullEnrich to get personal emails, phone numbers, and mobile ad IDs, since Meta matches on hashed personal identifiers rather than the work emails a CRM stores. Claude Code then pushes the lists into Meta through the API straight from the terminal, creating three audiences: the full Tier 1-4 list, a 1 percent lookalike off Tiers 1-2, and a 1 percent lookalike off closed-won. They run an objective ladder starting on Lead then MQL then SQL as pipeline data comes in, with 20+ creatives and 50/50 split landing pages, and wire MQL, SQL, and closed-won back from the CRM so Meta optimizes against closed deals rather than form fills.
@sakashiiiii [Claude Code]
https://x.com/sakashiiiii/status/2084480412269834653
A Japanese user made a manga about the Fire Emblem new title announcement using Keisuke's AI Manga Creator, which they are testing pre-release as a monitor. They report the Claude Code Plus plan struggles a bit under it and it affects their other work; going through the API instead, producing 12 pages costs a little under 200 yen, which they note is a real cost. They mention a 500-yen pre-registration discount coupon.
@garrett_makes [Claude Code]
https://x.com/garrett_makes/status/2084465677369229361
Reports that AI code quality has been flat but planning and understanding have gotten much better. A year ago they could not reliably have GPT or Claude drive a browser for smoke testing features. Today they can hand the agents a user story and a branch and get back screenshots on the PR with bug reproduction steps. They credit the combination of tooling (Claude Code and the Codex app) plus model improvements.
@hqmank [Claude Code]
https://x.com/hqmank/status/2084468951640121655
Says one skill made paid PDF tools useless for them. Before, they entered dates and hours by hand and calculated rates and totals, repeating the same invoice work each week. Now the skill fills out the invoice and does the math. They predict agents like Codex and Claude Code will swallow 90% of task-specific software.
@dherealmark [Claude Code]
https://x.com/dherealmark/status/2084520587167424630
Shares their experience using coding agents as orchestrators. They found Codex very complex to use as a conductor and switched to Claude Code (the orange crab) because it is much simpler. They still consider Codex very precise and careful, detailing everything in its sprints, so they keep a GPT swarm running but under the Claude Code harness.
@taishimatsuyama [Claude Code]
https://x.com/taishimatsuyama/status/2084782583725105559
A Japanese rehab facility operator uses Claude Code for financial scenario analysis across three service models: outpatient rehab at 1:1 staffing, short-hour day-care rehab at 3:1 mixing fee-for-service and flat-rate subscription users, and comprehensive community services at 5-6:1. The goal is a flexible structure that never falls below break-even regardless of how demand shifts between services. With productivity per therapist differing by program and too many variables, computing multi-scenario break-even projections by hand in Excel was a struggle, but with well-set assumptions Claude Code visualizes complex scenario analyses in a short time. The time saved now goes into refining assumptions and building persuasive materials for decision-makers.
@agence_sparkana [Claude Code]
https://x.com/agence_sparkana/status/2084541512759538030
French user starts a series on the Claude Code skills they actually use, beginning with the one they open most, called Impeccable. They invoke it when a page works but still looks 'made by a developer': it revisits hierarchy, spacing, and visual rhythm, and explains why. They have run it on their own landing page (le Coffre) and on several client deliverables.
@WiFiMoneyGuy [Claude Code]
https://x.com/WiFiMoneyGuy/status/2084485510253223965
Describes their evolving harness setup over time. When using Opus 4.6 and GPT 5.4 they ran a harness called OMO inside OpenCode. When Fable came out they went back to Claude Code for most of their heavy-duty coding. They also say DeepSeek V4 Flash is at the level of Opus 4.6 and works better on a harness like OMP, fully local.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2084488833198915825
Recounts an attempt to build GTA 6 with agentic loops, 22 hours and 86 agents in. The first attempt got stuck after generating an empty 3D world; the fix was changing what the agent could see, having the game export structured JSON describing its own state instead of Claude Code reasoning over pulled video frames, and having Claude call an image generation API to make its own assets. The builder runs a 'gauntlet loop': give the agent the goal not the architecture, give it an unarguable bar (real GTA 6 trailer screenshots), let it split work into separately judgeable pieces, and never let the builder grade its own work, with a fresh-context critic doing blind A/B comparisons against the reference and no cap on rounds. The run cost around $1,700 of API credits if not on Max.
@bond_ai1 [Claude Code]
https://x.com/bond_ai1/status/2084588908747403506
Japanese info-product seller who uses Obsidian as an external brain that Claude Code reads and writes daily. They store four things: posting tone rules and prohibitions, records of products that sold and did not sell, logs of customer questions and answers, and memos of daily decisions. With this in place, instructions as short as 'do the quote for company A' now work, whereas previously they spent about 10 minutes just explaining context. Their conclusion: the AI was not dumb, it just had nothing to work from.
@C_2049s [Claude Code]
https://x.com/C_2049s/status/2084519819022655797
Heavy user of Codex CLI and Claude Code hit three pain points: manually switching multiple API keys and waiting idle when quota ran out, incompatible protocols locking clients to upstreams, and config-switcher tools still requiring manual edits. So they built OpenAPI Multi-Key Proxy, a transparent network-layer runtime proxy where clients like Codex CLI, Claude Code, Cursor, and Chatbox just point their baseURL at it with no config changes. It does three-way protocol conversion between OpenAI Responses, Anthropic Messages, and OpenAI Chat Completions, smart multi-key scheduling by health score, cooldown state, sliding success rate, and latency percentiles, system-level failover with auto-lockout and recovery, and a full monitoring dashboard with Prometheus and webhooks. It runs as a single Node.js process depending only on ws, open-sourced since June 2026 with 110+ versions, MIT licensed.
@He1s_Sammy [Claude Code]
https://x.com/He1s_Sammy/status/2084521091431571697
Describes a Claude Code setup that calls Opus only twice per turn while a cheap model does everything else, arguing 80% of tokens go to work that never needed the expensive model. Opus plans at the start and grades at the end, and everything in between runs on Sonnet across 8 parallel lanes, with tests deciding when a lane is done rather than how sure the output sounds. They note Sonnet breaks in four places: ambiguous asks, long chains where early instructions fade, new designs with no precedent, and grading its own work, which is exactly what Opus is reserved for. The claimed result is 2 expensive calls per turn instead of forty.
@AaronxShepherd [Claude Code]
https://x.com/AaronxShepherd/status/2084716059194519635
Fed Claude Code five domains and got back a 10,000-person lead list without touching a scraper or enrichment tool. The workflow: hand it five ideal customers, it runs a look-alike search, builds the company list from the matches, then pulls every contact. One pull returned 4,000 founders, 3,200 marketers, and 3,500 business development contacts. Work that used to take a person hours now runs in a single prompt.
@kurono_ai_ura [Claude Code]
https://x.com/kurono_ai_ura/status/2084474411298509207
Japanese post about an overseas creator who fully automates video editing using only Claude Code. Rough cutting uses WhisperX transcription to automatically remove silences, bad takes, and filler; in the demo, 4 minutes 10 seconds of raw footage was cut to 47 seconds, nearly on target in one pass. Graphics are auto-generated by an engine called Hyperframes, fixable with natural-language instructions like 'make it orange' or 'add the logo' with only changed parts re-rendered, and subtitles and BGM are handled inside the same single Claude Code project from import to export. The creator says an edit that normally takes 4-5 hours took 2 hours including fine fixes, and the poster argues editing is shifting from learning software operation to deciding what to instruct Claude.
@aceop_xyz [Claude Code]
https://x.com/aceop_xyz/status/2084592534182076740
Used a multi-agent pipeline to research whether buying and holding memecoins beats simply holding BTC or ETH. ChatGPT refined the research prompt, Grok designed and froze the methodology, and Claude Code handled data collection, surprising them by building an entire research pipeline pulling from CoinGecko, DefiLlama, GeckoTerminal, and CEX APIs with rate-limit handling, caching, resumable workflows, duplicate skipping, and cross-source validation. The final dataset: 5,598 memecoins, 2.59M+ daily price observations, 23,174 raw JSON responses, and over 30,000 API calls, plus QA reports and an audit trail. Surf Research 2.0 then verified the methodology and concluded the evidence did not support a long-term memecoin buy-and-hold strategy over BTC or ETH for the average investor.
@tenobrus [OpenClaw]
https://x.com/tenobrus/status/2084790004086403381
Reply reporting hands-on experience: they set up both Hermes and OpenClaw and used each for a few weeks. Unfortunately they found both incredibly brittle and high friction, with little genuine utility.
@IhorSkiba [Claude Code]
https://x.com/IhorSkiba/status/2084650619000742039
Detailed recap of a 10-minute video by a self-described non-developer with 1,000 hours in Claude Code and 150 agents shipped, running his life from a terminal, mixed with the poster's own takeaways. They were pricing a $200-$500/seat/month stack (IDE seat, copilot sub, SaaS dashboard) and replaced it with one subscription, comparing $20 Pro against $200 Max at 20x usage. Running /init to read the whole folder and write CLAUDE.md killed the ten minutes of re-explaining at the start of every session. They highlight skills as whole jobs you teach once (like a desktop-cleanup skill filing everything into dated folders) replacing the VA they were about to hire, and MCP driving Canva, Gmail, Drive, Calendar, and Higgsfield from the same terminal as ending the app-switching tax.
@LotusDecoder [Claude Code]
https://x.com/LotusDecoder/status/2084440214726987938
Chinese user tracks Codex and Claude Code billing cycle dates and remaining quota by combining Codex background queries with the TickTick (Dida) list API. The agent first checks the official docs, writes cycle dates and remaining allowance into TickTick keyed by due date, then renders a calendar view that refreshes on a schedule. The result is overall usage visible at a glance.
@milloruns [Claude Code]
https://x.com/milloruns/status/2084645208734511285
Discovered their coding agents had been quietly running up a tab via agentacct, a local terminal dashboard that reads the session logs Claude Code and Codex already write on the machine, with no account or cloud. It broke down every session by tokens, cost, files touched, and which tests actually ran. The part that got them was tasks marked 'done' that never ran a single test, which the dashboard labels 'agent reported, not verified'. They now check what actually got verified before believing anything shipped, treating a passing test as proof and everything else as just the agent's word.
@scotthuang_claw [OpenClaw]
https://x.com/scotthuang_claw/status/2084695136886481091
Shares their project Agent Knock Knock, which aims to turn OpenClaw into an orchestrator that better coordinates the work of Codex and Claude Code. It supports coding-agent discovery and routing, task monitoring, approval callbacks, and letting a human take over and hand back control at any time. They are asking to join a beta test.
@hipreetam93 [Claude Code]
https://x.com/hipreetam93/status/2084461906283356331
Found coding from a phone terminal horrible and Slack too hard to track with many parallel chats, so about three weeks ago they built a custom UI on top of the Claude Code, Codex, and OpenCode SDKs running on a Hetzner VPS with access over Tailscale. Every repo is a workspace and every workspace gets multiple parallel windows, as many as VPS RAM permits. They call it ACT, short for AgentChatDeck, and have shipped entire features from their phone, which they were never able to do before.
@MystiqueMide [Claude Code]
https://x.com/MystiqueMide/status/2084669688034758784
Set up a recurring automation (theirs runs every 5 hours) that clears out old Codex cache, thread logs, and temp files that are safe to remove, touching nothing active or recent, only stale data taking up space. They share the exact reusable prompt, which they say also works for Claude Code or any other AI: ask it to set up an automation on a chosen interval that runs automatically without asking each time.
@Rezzi_sol [Claude Code]
https://x.com/Rezzi_sol/status/2084628846570061824
Describes a workflow a developer found that reportedly cuts Claude Code weekly usage by 50% without paying for extra API tokens, by not using the most expensive model for every stage. Install the OpenAI Codex plugin inside Claude Code, then split the work: Fable 5 High handles planning, Codex 5.6 xhigh handles execution through the existing Codex subscription without API usage, and Fable 5 Max reviews the finished work. The logic: planning and review need the most judgment, while execution consumes the most tokens, so moving execution to Codex preserves the Claude weekly allowance. Same workflow, roughly half the Claude usage.
@thomasgauvin [Claude Code]
https://x.com/thomasgauvin/status/2084468011084157086
Practitioner pushback on a blog post: they get the best results from coding agents when they are driving, defining requirements as they prompt Codex, Claude Code, Cursor, or OpenCode, and actively enforcing their quality bar. They can entrust base loops and agent self-reviews for the engineering core, but agents still have no taste on the actual product surface and UX, and no amount of skills or harness engineering fixes that, so they remain the bottleneck. They agree human code review is on its way out and that codebase flux is too high for CI/CD and git to scale, predicting non-GitHub/GitLab collaboration workflows and base loops for security, performance, reliability, and maintainability.
@draprints [Claude Code]
https://x.com/draprints/status/2084759388602962112
Job posting from EagleRev, which runs outbound for 100+ B2B companies with 3M+ LinkedIn DMs sent, its own sending platform, and an agent stack doing the boring parts. They are hiring a GTM Engineer to run email and LinkedIn campaigns end to end for real clients and to build with Claude Code daily, since their whole stack is agents. Requirements include having actually shipped things with Claude Code, not just watched videos about it. Fully remote on US hours with $2-4k/month OTE and a path to Lead GTM.
@rostikdeni [Claude Code]
https://x.com/rostikdeni/status/2084591791269961882
A creator connected Claude Code to Unreal Engine 5 and built a playable game in a repeatable loop. UnrealClaude gives Claude access to the viewport and VibeUE exposes Blueprints and Python; the agent builds one small system at a time, launches the game, inspects screenshots, commits the change, and fixes what breaks. For an indie studio this starts to look like a real prototype service, turning a brief into a playable vertical slice faster while the human keeps control of design and reviews the messy Blueprint logic. The limits are visible: one iteration reportedly took around 15 minutes and 14,000 Opus tokens, so scope and cleanup still decide whether the economics work.
@Fluyeporlaweb [Claude Code]
https://x.com/Fluyeporlaweb/status/2084640816018997525
Spanish post about a two-person team that found a vulnerability worth $250,000 in a single bug bounty, after 9 months hunting High and Critical flaws with AI, finding more than 20 and earning over $1.5M. They just published their complete system on GitHub as open-kritt. Instead of handing a whole repo to a model and hoping (which almost never works), it chops the audit into small specific tasks, launches AI agents in parallel on each, merges and deduplicates findings, classifies them by severity, and validates each with scripts before presenting it as real. It is self-hosted so code never leaves your infrastructure, and compatible with Codex and Claude Code.
@phyrexia [Claude Code]
https://x.com/phyrexia/status/2084526723480056238
Instrumented 30 days of their own usage (56k Claude Code turns, 256k Codex turns, same projects and kind of work) to quantify why Codex feels slower. Median turns per session were 10 for Claude Code vs 19 for Codex, with 1,162 output tokens per turn vs 464, so Codex takes about twice the turns and produces roughly 2.5x less per turn, moving in many smaller steps. Their initial token-efficiency read (3.61% Claude Code vs 2.77% Codex) turned out to be a Simpson's paradox artifact: standardized to the same session-length mix it flips to Codex 10.14% vs Claude Code 5.50%, so the gap they feel is in step count and wall-clock, not token efficiency. They also flag a fixed measurement bug: Claude Code reports almost all of its prompt under cache_creation with a median input of 2 tokens per turn, which had understated its real input by a wide margin.
@ColinGardiner [OpenClaw]
https://x.com/ColinGardiner/status/2084761383703314613
Asks for outside opinions on OpenClaw versus just using the ChatGPT/Claude apps, grounded in their own experience. They feel desktop usage is much better with the apps and always struggled to get OpenClaw to work well on a cloud-hosted instance. OpenClaw does integrate well with their gbrain and handles their workflows really well, but they feel those could move to the apps too, and they report superior results in certain use cases with the apps versus OpenClaw.
@EvanKirstel [Claude Code]
https://x.com/EvanKirstel/status/2084668180837048642
One person mapped the world's entire electrical grid as a side project, built with Claude Code. The map covers 120,000+ power plants, roughly 3 million miles of transmission line, and hundreds of thousands of substations. Data came from OpenStreetMap, Global Energy Monitor, and EIA, as reported via FastCompany.
@joedevon [OpenClaw]
https://x.com/joedevon/status/2084748167497875892
Wanted to use OpenAI more a while back because it did best in their accessibility benchmarks, but hated the harness and kept drifting back to Claude Code. For what feels like months now the situation has reversed: they hate using Claude Code and much prefer Codex CLI, which they say has gotten way, way better, especially with deep use. They single out computer use as night and day better.
@okirumama_ [Claude Code]
https://x.com/okirumama_/status/2084433717851013151
Japanese user runs an entire Threads account operation on Claude Code. Their listed workflow: get Claude Code on their side, launch the Threads account, then use Claude Code for concept design, post creation, writing note articles, doing the various settings, and keeping management ledgers. Because everything runs through Claude Code, they say the operation keeps going even when their autonomic nervous system acts up.
🗣 User Voice
User Voice
1. Trust and bans broke into the mainline conversation. @heynavtoor and @OlivercrestAI amplified the reverse-engineered discovery of hidden environment-detection code in the Claude Code binary that flags China-linked setups via prompt steganography, @liuren claims over half the Claude accounts in their Chinese circle have been banned since late June, and @Joshuwa points to a two-year account suspended without warning as the argument for building on open alternatives.
2. Memory that survives sessions and tools is the top capability ask. @halilkutukk spent 8 months re-explaining the same context because every session starts from zero, and @GeekCatX wants work memory that follows the project across Claude Code, Codex and Cursor instead of dying at each tool boundary.
3. Token economics anxiety sharpened with concrete numbers. @Mortezabihzadeh flags that Sonnet 5 intro pricing ends Aug 31 while the new tokenizer counts roughly 30% more tokens for the same code, @pulmencr shows sessions hitting hard rate-limit walls by message 20 with 5-hour lockouts, and @L_go_mrk wants dependency-graph-restricted reads because full-codebase re-reads melt tokens on every edit.
4. Multi-model workflows are held together with duct tape. @C_2049s lists manual API-key switching, idle waiting on exhausted quotas and incompatible protocols; @drbarnard wants a proper adversarial loop between Claude Code and Codex instead of clunky tab-switching; @juliankmazo wishes for one CLI that runs any model and spawns background agents on any other.
5. Quality-consistency complaints got specific. @MichLieben shows Claude Code solving the same task five different ways in five runs without a skills-and-hooks structure to pin it down, @DanielGlejzner got 182 useless Angular tests from a no-context run, and @ArielKwiat's blunt "unbelievable how bad in Q3 2026" post drew 27k impressions of agreement.
1. Trust and bans broke into the mainline conversation. @heynavtoor and @OlivercrestAI amplified the reverse-engineered discovery of hidden environment-detection code in the Claude Code binary that flags China-linked setups via prompt steganography, @liuren claims over half the Claude accounts in their Chinese circle have been banned since late June, and @Joshuwa points to a two-year account suspended without warning as the argument for building on open alternatives.
2. Memory that survives sessions and tools is the top capability ask. @halilkutukk spent 8 months re-explaining the same context because every session starts from zero, and @GeekCatX wants work memory that follows the project across Claude Code, Codex and Cursor instead of dying at each tool boundary.
3. Token economics anxiety sharpened with concrete numbers. @Mortezabihzadeh flags that Sonnet 5 intro pricing ends Aug 31 while the new tokenizer counts roughly 30% more tokens for the same code, @pulmencr shows sessions hitting hard rate-limit walls by message 20 with 5-hour lockouts, and @L_go_mrk wants dependency-graph-restricted reads because full-codebase re-reads melt tokens on every edit.
4. Multi-model workflows are held together with duct tape. @C_2049s lists manual API-key switching, idle waiting on exhausted quotas and incompatible protocols; @drbarnard wants a proper adversarial loop between Claude Code and Codex instead of clunky tab-switching; @juliankmazo wishes for one CLI that runs any model and spawns background agents on any other.
5. Quality-consistency complaints got specific. @MichLieben shows Claude Code solving the same task five different ways in five runs without a skills-and-hooks structure to pin it down, @DanielGlejzner got 182 useless Angular tests from a no-context run, and @ArielKwiat's blunt "unbelievable how bad in Q3 2026" post drew 27k impressions of agreement.
📡 Eco Products Radar
Eco Products Radar
Products mentioned 3+ times in today's data:
Codex (125) — the constant comparison target and the destination for refused or rate-limited tasks
Cursor (69)
Hermes Agent (43) — memory-and-skills agent, also cited in brittleness complaints
DeepSeek (29) — the cheap-API engine behind OpenClaw and clone stacks
Gemini (25)
Not Diamond Code (28) — heavy coordinated promo wave today; counted, flagged as astroturf
OpenCode (18)
Slack (18)
MCP (33 incl. servers)
Qwen (16)
Kimi (15)
Grok / Grok Build (17)
Claude Cowork (11)
n8n (10)
Seedance (9)
GLM (8), Obsidian (8), OmniRoute (8), Ollama (8)
Pi (7), Arcads (7), Hostinger Connector (7)
Products mentioned 3+ times in today's data:
Codex (125) — the constant comparison target and the destination for refused or rate-limited tasks
Cursor (69)
Hermes Agent (43) — memory-and-skills agent, also cited in brittleness complaints
DeepSeek (29) — the cheap-API engine behind OpenClaw and clone stacks
Gemini (25)
Not Diamond Code (28) — heavy coordinated promo wave today; counted, flagged as astroturf
OpenCode (18)
Slack (18)
MCP (33 incl. servers)
Qwen (16)
Kimi (15)
Grok / Grok Build (17)
Claude Cowork (11)
n8n (10)
Seedance (9)
GLM (8), Obsidian (8), OmniRoute (8), Ollama (8)
Pi (7), Arcads (7), Hostinger Connector (7)
Comments