Super User Daily: August 4, 2026
The most striking cases today come from people who never learned to code: a Japanese idol shipped the entire production system for a 10-hour YouTube livestream with Claude Code without reading a single line of code, a CPA turned a full financial due-diligence workflow into a repo, and one builder moved files between two laptops by having Claude Code invent QR-code optical transfer overnight. On the infrastructure side the escape-hatch economy is booming: a 43,000-star proxy that swaps Claude Code's backend for DeepSeek, Kimi or local models, Termux setups that turn Android phones into dev machines, and desktop-app SSH replacing terminal workflows. Multi-agent orchestration keeps maturing, with six-agent Codex-plus-Claude swarms, per-project OpenClaw fleets and headless Slack bots. Meanwhile quota whiplash, verbosity complaints and unattended-run fears dominate the grumbling.
@MedusaOnchain [Claude Code]
https://x.com/MedusaOnchain/status/2083903710506504548
Viral report that Claude Code found the COLDCARD hardware wallet vulnerability with a single prompt, in just 8 minutes of thinking. The poster calls it insane and says we're not ready for what's coming. A high-impact example of Claude Code doing real security vulnerability discovery.
@MedusaOnchain [Claude Code]
https://x.com/MedusaOnchain/status/2083903710506504548
Viral report that Claude Code found the COLDCARD hardware wallet vulnerability with a single prompt, in just 8 minutes of thinking. The poster calls it insane and says we're not ready for what's coming. A high-impact example of Claude Code doing real security vulnerability discovery.
@raurublock [Claude Code]
https://x.com/raurublock/status/2083843572604817917
Japanese post relaying that a former idol, Karin Miyamoto, used Claude Code to build her own streaming system. Framed as a dialogue where a skeptic worries what happens when something breaks, then is stunned to learn who the builder is. A striking case of a non-engineer building real working infrastructure for her own broadcasts with Claude Code.
@arceyul [Claude Code]
https://x.com/arceyul/status/2083881692880683245
Spanish post describing Video Use, a free open-source video editing tool for Claude Code. You drop raw footage and files into a folder and it cuts clips, removes filler words, adds subtitles, applies color correction and filters, handles animations, and renders the final video. No timeline and no manual editing; the poster positions it as going beyond tools like Remotion. Open source on GitHub.
@Suu766 [Claude Code]
https://x.com/Suu766/status/2083822324697235928
Chinese post about reverse-skill, a skill that packages reverse engineering expertise for Codex, Cursor, or Claude Code. Once installed, when the AI encounters APKs, web code, or firmware it decides on its own what to inspect first, which tools to use, and how to proceed with the analysis. The poster argues this compresses years of hard-won security experience into a loadable skill pack, collapsing the barrier to an experience-heavy field.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2083953532903059846
Built a /watch-video skill that watches videos for him: YouTube, Loom, Zoom recordings, Vimeo, and local files. It pulls the transcript, extracts frames, runs a vision pass for key moments, and files structured notes into his second brain. He is using it to study 100 YouTube videos before launching his channel, and it also processes client Looms and call recordings. Part of Maker Skills, 18 free open-source skills for founders installable in Claude Code.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2083762211718394276
Setup walkthrough for connecting Claude Code to the Financial Datasets MCP server so it can pull live data on 17,000+ stocks, crypto prices, and financial statements in seconds. One 'claude mcp add' command plus an OAuth flow via /mcp, then natural-language prompts like Apple's P/E ratio or Tesla's income statement over four quarters. The poster frames it as replacing a $24,000-a-year Bloomberg Terminal for research.
@MedusaOnchain [Claude Code]
https://x.com/MedusaOnchain/status/2083987806943432847
Follow-up to the COLDCARD story: GLM-5.2, trained on a pre-exploit cutoff with no internet access, independently found the same wallet vulnerability with a single prompt in about 20 minutes of thinking. Directly counters the claim that Claude Code merely retrieved human discussions from the internet. Concrete cross-model verification that the flaw was discoverable purely from the code.
@_guillecasaus [Claude Code]
https://x.com/_guillecasaus/status/2083884677379625342
Spanish post about Ponytail, a free open-source Claude Code plugin targeting agents' tendency to generate far more code than needed. Claims it cuts generated code by 80-94 percent, lowers costs 47-77 percent, and speeds tasks 3-6x. Before writing a line, it forces Claude to check whether the functionality already exists, can be solved with a standard library, or truly needs new code, yielding cleaner and cheaper-to-maintain output.
@AlexFinn [OpenClaw]
https://x.com/AlexFinn/status/2084032423767990339
Describes a remote-coding setup where one always-on Mac Studio is the sole dev machine and every other device (iPad, iPhone, other Macs) remotes into it via the Codex app's device-control features. Tailscale creates a private network so his other agents, OpenClaw or Hermes, can jump between computers and make changes when needed. Claims this lets him code from anywhere - grocery store, bed, Japan, poolside - and massively boosted his productivity on big projects.
@Piefena [OpenClaw]
https://x.com/Piefena/status/2083893184229253625
One person built a pocket AI assistant for about $120 using a Raspberry Pi Zero 2 W, a Whisplay board with a 1.69-inch screen, microphone, speaker, and buttons, plus a rechargeable PiSugar battery. Press the button and ask a question; OpenClaw transcribes it, thinks, and returns the answer on the display. The board only handles the interface and connects to cloud AI via API, so the device improves as underlying models improve. The poster sketches the business math: 100 units at $249 is $24,900 in revenue from generic hardware.
@FurukawaMirin [Claude Code]
https://x.com/FurukawaMirin/status/2083847614064390536
One month into Claude Code, a user who doesn't write code built an OBS panel tool for announcing simultaneous Twitch/YouTube streams. Setting the title, description, and thumbnail then pressing apply creates the stream slot, auto-composes the announcement text with the URLs, and an X post button pops up one-click posting. Replaces the previous painful URL copy-paste routine and reduced pre-stream panic.
@dr_cintas [Claude Code]
https://x.com/dr_cintas/status/2083985444912463965
Recommends code-review-graph, which maps the entire codebase - every file, function, and connection - so the AI stops rereading the whole repo after each edit. Change one function and it traces every caller, dependent file, and test, so Claude Code only reads what's affected. Claims a task that burned ~100,000 tokens (about a dollar) now runs closer to a penny. Two-minute pip install that auto-configures Claude Code, Cursor, Codex, Gemini CLI and more.
@maarcoofdezz [Claude Code]
https://x.com/maarcoofdezz/status/2083874888352252097
Spanish post about ECC, a repo bundling 262 skills plus instincts, memory, security, and workflows for Claude Code, Codex, Cursor, or OpenCode. Built by Affaan Mustafa at the Anthropic x Cerebral Valley hackathon using Claude Code itself, and the 2.0.0 release automated its own Discord bot, announcements, and deploy end to end. Includes AgentShield config vulnerability scanning, a red team/blue team mode with 3 parallel Opus agents, 1,282 tests at 98 percent coverage, and shared cross-agent memory.
@taiyaki_ai3 [Claude Code]
https://x.com/taiyaki_ai3/status/2083858242988609782
Japanese user shares that the Claude Code desktop app can connect directly to servers over SSH, replacing terminal-based SSH workflows for remote development. Per an Anthropic insider, one desktop app connects straight to a VPS or VM with setup done in minutes, no terminal needed. The poster laments burning a whole night on his homelab before learning this.
@yanhua1010 [Claude Code]
https://x.com/yanhua1010/status/2083763240811532636
Describes a personal agent-harness lineup: native Claude Code for Claude models because Anthropic bans third-party tools from using subscription quota, Codex for GPT models, and pi for DeepSeek and Grok models. Notes pi is chosen for provider neutrality, minimalism, and deep customizability, and shares a Meta engineer's recommended pi configuration.
@BruceBlue [Claude Code]
https://x.com/BruceBlue/status/2083792610909712708
Chinese post sharing a concrete model-layering workflow for Codex modeled on Claude Code's /advisor: GPT-5.6 Sol handles planning, analysis, and complex problems while GPT-5.6 Luna does implementation, testing, and exploration. Includes a copy-paste prompt that makes Codex design, deploy, and verify the Sol-Luna layered system itself, escalating Sol to High only for genuinely difficult, high-risk decisions.
@itsharmanjot [Claude Code]
https://x.com/itsharmanjot/status/2083816823766810778
Describes teamlore, an open-source approach to team AI memory using a single .lore/ folder in the repo instead of servers or vector databases. When a Claude Code agent gets corrected or breaks something, it writes a small lore file that ships with the PR, gets reviewed like code, and after merge every teammate's agent recalls it when touching that part of the repo. A companion scarmap command turns the team's mistake history into a heat map of the codebase. The repo's own .lore/ folder contains every mistake Claude made while building teamlore itself.
@ChrisGPT [Claude Code]
https://x.com/ChrisGPT/status/2084025602583982130
Detailed breakdown of the COLDCARD story: a firmware bug checked whether the hardware RNG setting existed rather than whether it was enabled, reducing some wallets from roughly 2^128 to about 2^40 possible seeds; researchers count 1,367 BTC (~$89M) drained from 4,585 addresses. Afterward a Reddit user pointed Claude Code at the public firmware asking only to check for vulnerabilities, and within eight minutes it traced the broken random number path to the same flaw. The poster stresses there is no proof the original attacker used any LLM, but that one person with one broad prompt can now uncover flaws humans missed for five years.
@gerardsans [OpenClaw]
https://x.com/gerardsans/status/2083892925759357082
Analysis of the CRUX study in which frontier agents, primarily Claude Opus 4.8 with OpenClaw, got six days, $3,000, full sandboxes, and web access to tackle the central questions of two unpublished NeurIPS 2026 papers. The agents completed the full execution loop with zero human intervention but both resulting papers were rejected by the original authors (scores 1/6 and 2/6). Five recurring failure modes emerged - poor judgment, narrow exploration, ineffective backtracking, leaving over half the API budget unused, and instruction drift - and a GPT-5.6 Sol Ultra + Codex robustness run reproduced every one. Concrete evidence on the current limits of coding-harness agents for open-ended research.
@chizevsm5 [Claude Code]
https://x.com/chizevsm5/status/2083909785787994238
Japanese walkthrough of connecting Claude Code to the Financial Datasets MCP for access to 17,000+ US stocks: one command plus browser auth gives stock prices, income statements, balance sheets, cash flow, SEC filings, insider trades, institutional holdings, and earnings versus analyst estimates inside the conversation. Details real constraints: the free tier covers only AAPL/NVDA/MSFT/TSLA, full coverage starts at a $20 one-time credit for 1,000 requests, and it is US stocks only. The poster has 100 million yen in stocks and plans to experiment with it for research.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2083836451247231484
A father replaced his 2nd-grade daughter's worksheet pile with a knowledge graph: 463 skills, 9,601 learning resources, and 155 hours of content on a single map. Each skill links to the resources needed to acquire it, finishing a lesson auto-unlocks the next step, and the child can see where she is, what comes next, and how much remains. The poster draws the direct parallel to structuring one's own work with Claude Code Skills and CLAUDE.md: making the connections between tasks visible is what makes the system stick.
@alex_prompter [Claude Code]
https://x.com/alex_prompter/status/2083961326427402467
A two-hook fix for Claude Code's habit of editing 5 files, leaving TypeScript errors in 3, and declaring itself done. A PostToolUse hook logs every edited file and its repo; a Stop hook then runs the build on each touched repo and feeds errors straight back to Claude before the user ever sees them, recommending a dedicated error-resolver agent when 5 or more errors appear. A dev using this across 8 repos and 300K lines of code reported zero instances of leftover build errors after adopting the system.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2083847779949134312
Japanese post on a setup an engineer uses to halve Claude Code weekly limit consumption via model role-splitting: install the OpenAI Codex plugin into Claude Code, let Fable 5 high do planning, hand execution to Codex 5.6 xhigh (covered by the Codex subscription at no extra cost), and have Fable 5 max do the final review. The insight: execution burns the most tokens, so keep judgment-heavy planning and review on Fable and offload heavy execution to the other contract - treating limits as a design problem, not a billing problem.
@usedhonda [Claude Code]
https://x.com/usedhonda/status/2083766508107514007
Shares a complete, extremely detailed audit-and-redesign prompt for making CLAUDE.md, AGENTS.md, and surrounding config work symmetrically across Claude Code and Codex. The design demands role reversibility (either tool can be orchestrator, main implementer, or reviewer, with mid-task handoffs both ways), conversation-independent handoff documents, a single source of truth for shared project rules with thin per-tool adapters, and moving enforceable constraints into CI, lint, and hooks. Written from the poster's weekend re-inventory of agent config for people who bounce between Claude Code and Codex.
@moriyorihayash1 [OpenClaw]
https://x.com/moriyorihayash1/status/2083864951106748582
Japanese user running OpenClaw as a personal secretary on DeepSeek V4 Flash reports 200 million tokens for about 2,500 yen and says the quality is perfectly satisfactory. A concrete cost and quality datapoint for running an OpenClaw assistant on a cheap non-frontier model, with a jab that OpenAI should rethink its pricing.
@theinformation [Claude Code]
https://x.com/theinformation/status/2083953477420540039
The Information reports small companies are using Claude Code and other AI tools to replace Salesforce, HubSpot, and other business software. One real estate firm says its custom app cut annual costs by about $100,000 and now runs for roughly $300 a month. Concrete dollar evidence of Claude Code-built internal tools displacing SaaS.
@Dev_trader0 [Claude Code]
https://x.com/Dev_trader0/status/2083961152631890004
Recommends a GitHub repo by Ayghri that gives Claude Code a simple ruleset to stop burying the answer: answer first, number the execution steps, skip unnecessary introductions, reduce filler and wasted context, stay concise and actionable. A tiny change the poster says makes Claude Code feel significantly faster and easier to work with.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2083711881223746004
A Claude Code + Higgsfield + Figma workflow for editable AI ads: Claude builds a Figma template with style guide, components, and 4:5/1:1/9:16 masters; you feed it your logo SVG, product photo, and brand hex codes; it writes the angles and copy for approval, fires six Higgsfield prompts at once, QAs its own images and rejects bad ones, then assembles finished ads in Figma with headline, body, CTA, and logo as live editable layers. Kills the flat-PNG problem where designers rebuild AI statics by hand; change one brand variable and every ad updates. Built entirely in Claude Code with the Higgsfield MCP, with a recorded walkthrough.
@xiaoying_eth [Claude Code]
https://x.com/xiaoying_eth/status/2083726738350391595
Chinese post on social-auto-upload, an open-source tool (13k+ stars) that automates multi-platform video publishing to Douyin, Bilibili, Xiaohongshu, Kuaishou, Shipinhao, TikTok, and YouTube. Targets the pain of uploading the same video 7-8 times while rewriting titles, descriptions, and tags. Supports CLI batch upload and can be wired into OpenClaw, Codex, or Claude Code as part of an agent workflow.
@arvin17x [Claude Code]
https://x.com/arvin17x/status/2083920843772764476
LobeHub's team Workspace entered private beta after essentially one person built it in 3 months powered by Claude Code and Codex Max plans - a plan that had sat on the shelf for two years. The post details the hard design problems worked through in dogfooding: agent visibility and sharing across native and external heterogeneous agents, connector management, API token budget granularity, isolating personal versus team memory, and balancing permission control against usability. Many intuitive approaches were tried and thrown away along the way.
@gippp69 [OpenClaw]
https://x.com/gippp69/status/2083992907896238095
Post about MemOS, a 10,500-star repo that turns an agent's history into inspectable, editable, removable memory instead of a black-box vector store: chat is distilled into facts, structured as graph memory, retrieved across tasks, and stale state can be corrected or deleted. Reports OpenClaw task completion rising from 36.63 percent to 50.87 percent with 35.24 percent fewer memory tokens, because the agent retrieves useful knowledge instead of rereading its entire history. One unified API covers text, images, personas, and tool traces, with memory cubes isolating knowledge across users, projects, and agents.
@chetanankola [Claude Code]
https://x.com/chetanankola/status/2083753330778722757
Made Codex and Claude Code review each other's work to build a 3D ocean in three.js, pairing GPT-5.6 Sol xhigh with Opus 5 ultra. Explicitly no gauntlet prompt and no one-shot magic - just iteration and a pile of reference images. A concrete cross-agent peer-review workflow producing a shipped visual result.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2083752396774388215
Describes Scroll-world, a free open-source agent skill for Claude Code and Codex that builds Apple-style scroll-driven 3D product sites for about $27 per 6-scene chain, versus $8,000-$100,000 agency quotes and 3-6 month timelines. You describe brand and scenes in plain language; it generates isometric stills via GPT Image 2 and camera flights via Seedance, builds connector clips from the boundary frames of neighboring scenes so every seam aligns, renders native 9:16 for mobile, and wires it all into a vanilla JS scroll engine for plain HTML, Next.js, or Vue. 6,096 GitHub stars and 724 forks in 25 days, MIT licensed.
@knowledgefxg [Claude Code]
https://x.com/knowledgefxg/status/2083902298414264424
Chinese post on PixelRAG, an open-source tool that skips HTML/PDF parsing by rendering documents to screenshots, embedding them with a vision-language model, and indexing with FAISS behind a search API - letting the model 'look' at pages instead of parsing them, which avoids failures on complex layouts, charts, and JS-rendered pages. Comes with a Claude Code plugin that effectively gives Claude eyes: it can screenshot and read web pages, papers, or local dev pages directly, with no extra MCP server or backend.
@naviidtaheri [Claude Code]
https://x.com/naviidtaheri/status/2083833934253633621
Builder shares Agentic Video OS, a free open-source engine for professional video without cloud rendering fees or services like HeyGen and Runway. You design the visuals or kinetic typography in HTML/CSS; it renders frame-by-frame via headless Playwright + Chromium and assembles clean 1080x1920 or 1920x1080 video with FFmpeg. Rendering is deterministic (same input, identical frames), it runs with personal Claude/OpenAI/Gemini API keys or fully local via Ollama, and the architecture is designed for coding agents like Claude Code and Codex.
@cryppimagic [Claude Code]
https://x.com/cryppimagic/status/2083976685662372282
Honest account of building an arbitrage scanner across perp venues with Claude Code: one detailed prompt is nowhere near enough because reality surfaces endless little details. Pain points hit include every exchange API being different, API data diverging from what the exchange UI shows with nothing to cross-check, a perp DEX quietly migrating from v1 to v2 and killing old endpoints, and the same token listed under different tickers (PUMP vs PUMPFUN, three differently named WTI contracts on BingX). Concludes that since AI showed up, he just works more.
@LayoffAI [Claude Code]
https://x.com/LayoffAI/status/2083944118372151759
Uber - the company whose executive said it burned through its entire Anthropic Claude Code budget - is now embedding top AI engineers directly inside non-technical departments like finance, HR, legal, and marketing for two-week 'Agentic Pod' sprints. Engineers spend the first days sitting next to staff learning their exact daily headaches, then immediately build software to fix them. Reported impact: financial planning dropped from 15 hours to 30 minutes, financial reporting from two days to 10 minutes, and marketing quality checks from two weeks to under an hour. The takeaway: gains come from redesigning workflows, not bolting a chatbot onto old processes.
@claudecode84 [Claude Code]
https://x.com/claudecode84/status/2083749944280879293
Japanese post on Career-Ops, an open-source job-hunting system built with Claude Code. It auto-checks job listings from 45+ companies including OpenAI, Anthropic, and Stripe, rewrites your resume and CV to fit each position, generates ATS-ready PDFs with Playwright, and automates application form filling. Ships with 14 modes/skills and a Go-based dashboard - a full non-coding application replacing recruiters and blanket resume blasts.
@cosmos_hzokujin [Claude Code]
https://x.com/cosmos_hzokujin/status/2083769622097227986
Japanese breakdown of free-claude-code (43,000+ GitHub stars), which points Claude Code, Codex, or Pi at a local proxy and routes requests to DeepSeek, Kimi, Qwen, OpenRouter, NVIDIA NIM, Ollama and 31 cloud/local providers, mapping models to Fable/Opus/Sonnet/Haiku tiers. You keep the Claude Code development experience while swapping the backing model to escape cost and rate-limit problems, switching providers from a local admin panel. The poster adds sober caveats: it does not make Claude itself free, and anyone handling confidential code or API keys should verify log settings and where traffic actually goes.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2083712771812708744
Chinese user installed the Scaleboard skill in Claude Code and tried it: paste in the creator's Xiaohongshu writeup, feed it your own life story, and it creates the video on its own; he swapped the default Seedance token plan for a different relay and supplied a few photos to get the finished piece. The generated video told his own last two years - quitting an algorithm engineer job in Shanghai, driving 3,000 km to Dali, a failed indie product, then 34k followers and seven-figure revenue - which he retells alongside his Ship2Market pitch for indie devs who can build but can't market.
@not_acp [Claude Code]
https://x.com/not_acp/status/2083857657061052670
Built a Claude Code skill called Learn From Session to close the gap between shipping fast and understanding what shipped. It recaps your session, quizzes you on it, then finds credible interactive blog posts on the concepts you're shaky on. A learning-loop layer on top of agentic coding, shared with the repo link.
@Shvartsbroit [Claude Code]
https://x.com/Shvartsbroit/status/2083972450996359393
Hebrew post from a developer describing his split: Claude Code for planning, design, copy, and orchestration, Codex for implementation. He's now considering dropping Codex and consolidating on Claude, saying it has gotten especially good again now that Fable 5, Opus, and their cheaper, faster siblings are available. Asks other AI developers which models they use for software work.
@Zephyr_hg [Claude Code]
https://x.com/Zephyr_hg/status/2083732263880343642
Boris Cherny, who built Claude Code, keeps 'sometimes 100, sometimes thousands' of copies running at once, and 100 percent of his code has been written by Claude Code since November. At that scale the workday becomes judgment: deciding which results ship and which get thrown out. One person, a thousand Claudes, nobody on payroll.
@nykdotdev [Claude Code]
https://x.com/nykdotdev/status/2083946147413164060
Built a quota dashboard aggregating Claude Code, Codex, Cursor, Grok, and Hermes - not to rank models, but to answer the operational question every CLI hides differently: which ready tool has current capacity evidence for the next job. The dashboard keeps auth, usage windows, reset timing, routes, and missing evidence separate.
@kodagen [Claude Code]
https://x.com/kodagen/status/2083817728494964795
Japanese user simplified their entire Claude Code setup - CLAUDE.md, Skills, Subagents, hooks, and MCP - following a reference, and reports Fable's token consumption dropped to roughly half by feel. Concludes that with Claude Code it seems better to rebuild your config from a clean slate every time the model changes.
@aitech_jp [Claude Code]
https://x.com/aitech_jp/status/2083869927426101449
Japanese account reports an overseas developer half-skeptically asked Opus 5 inside Claude Code to generate a video, and a 38-second clip was finished in under 10 minutes. It was a one-shot generation with no fixes, no editing, and the entire process from request to completion was made public. Framed as chat-driven first drafts of video now replacing work that used to be outsourced to production companies, with no camera gear or editing software.
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2084035333373481094
Developer launched VoiceClaw Realtime, letting you chat live on your Apple Watch with two-way voice to start, manage, and check in on work with AI agents in Codex, OpenClaw, and Hermes, over cellular or Wi-Fi. It is a direct connection to your own computer via Tailscale or a personal Cloudflare tunnel with no third-party server, and works with OpenAI's bidirectional chat models including a second background model. Ships with Apple Watch Smart Stack and watch face widgets, double-tap-fingers activation, iOS widgets, Live Activities for agent tasks, Action Button activation, and iPad support. Requires only a Mac, an Apple Watch, and five minutes of setup; free for now.
@vladuah [Claude Code]
https://x.com/vladuah/status/2083713673084703192
Builder made a scroll-heavy restaurant website with AI and sold it for $4,000. Used Higgsfield to generate the food photography with no studio or photographer, then fed Claude Code the requirements and got the full scroll engine plus sound design in one conversation, adding GSAP and Canvas for a frame-by-frame burger reveal with real sizzle audio. Total build time was one night; notes agencies were charging $8K+ for similar work last year and clients only care that it looks premium and converts.
@peter_0123 [Claude Code]
https://x.com/peter_0123/status/2083791370700120382
Japanese user shares how they made an AI idol concert video without knowing how to use dreamina's UI: they just told Claude Code to make a video of the girl in a reference image performing solo at a summer festival with given music, adding 'use Claude in Chrome to do it'. Claude Code drove the tool and wrote the detailed English multi-shot concert film prompt itself, which the user pastes in full for followers who asked.
@Oluwaphilemon1 [Claude Code]
https://x.com/Oluwaphilemon1/status/2083949647937081454
User has Codex and Claude Code review each other's work, and together the two agents built an animated ocean demo powered by GPT-5.6 Sol (xhigh), Claude Opus 5 (Ultra), and Three.js. Notes the setup is motivated by DeepSeek V4 Flash 0731 being ridiculously cheap compared to Kimi K3, GPT-5.6, and Opus 5.
@mkwizrd [Claude Code]
https://x.com/mkwizrd/status/2083940755287265317
Ecommerce operator on day 38 of building a real ecom brand describes a full 'trench day' on Claude Code after a customer spent big on their store. They created a custom landing page builder, added post-purchase surveys to Shopify connected to Claude Code for ideation, and connected their feedback-loop sheet so Claude analyzes new ad copy against ad learnings and direct response principles they provided. Next planned steps are connecting Claude Code to their ad library and giving it more ideation context.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2083868703628915073
Japanese post about an idol with zero programming experience who built the entire system for a 10-hour livestream using Claude Code without writing a single line of code. She shipped four things: a real-time gauge showing CD purchases and support counts, an admin panel to change numbers from a phone mid-stream, automatic counting of X hashtag posts, and an AI scoring system for MV reproduction accuracy. Her workflow was describing the screens and features she wanted, repeatedly confirming and approving, and showing the screen saying 'this part does not work' when something broke - she never read the code.
@insomnia_vip [Claude Code]
https://x.com/insomnia_vip/status/2083872884653453816
Story of a builder trying to build 100 iOS apps with Claude Code after it added native iOS Simulator support. Claude detects Xcode 26.6, connects to an iPhone 17 Pro simulator running iOS 26.0, and opens the app at 60 FPS; in the demo it navigates a dark mode app called Tonight, testing switches, sliders, and interface states automatically instead of manual QA. With testing, debugging, and iteration in one workflow, releasing ten experiments becomes more realistic than perfecting one for months.
@bojie_li [Claude Code]
https://x.com/bojie_li/status/2083820484664385801
Author of the book 'AI Agent in Depth' found readers noticed a chain-of-thought distillation experiment was missing even though his Codex agent reported all experiments complete. Codex on GPT-5.6-Sol stopped responding after a few steps, Claude Code with Opus 5 also stopped after a few iterations, and Kimi k3 ultimately finished it - he suspects CoT distillation is highly sensitive for these models. Separately, a reader fed the book to her Hermes agent, which learned capabilities like ablation studies, memory forgetting, and proposer-reviewer interaction scaling from it and iteratively proposed patches until they were accepted into its own source code.
@0x0SojalSec [Claude Code]
https://x.com/0x0SojalSec/status/2084034948626010546
Security post claiming Claude found a 5-year-old hardware wallet bug in 8 minutes: weak entropy from a Coldcard firmware mistake made seeds generated since 2021 far more guessable than realized, which attackers exploited by offline-generating candidates, matching addresses on-chain, and draining 1,367 BTC ($89M). Claude Code later traced the exact broken path in under 10 minutes when asked only to check for vulnerabilities. Humans missed it for half a decade; the model needed one broad prompt.
@kobayashiyuto10 [Claude Code]
https://x.com/kobayashiyuto10/status/2083750924900786526
Japanese reply from someone using Claude Code for landing page production at ExstTech: handing it requirements as just five bullet points yields a first draft solid enough to use directly as the working base. Agrees with the original poster that this makes the business viable without sales outreach.
@VinodSharma10x [Claude Code]
https://x.com/VinodSharma10x/status/2083874639298937012
Engineering leader who had not coded in 12 years returned to building two years ago via V0, Bolt, Cursor, GitHub Copilot, and Claude Code, and shipped 9 apps in 2 years while keeping a full-time job - a Chrome extension, a career tool, a startup called Sucana, and agentic automations. Says that in 2015 this would have taken a decade, and it led to co-founding Virgil with another founder in their 50s, building at 4am in hotel lobbies.
@HarriesSteele [Claude Code]
https://x.com/HarriesSteele/status/2083754826421449006
Chinese user warns from firsthand experience: do not use Claude Code to fill in your app's Google Play listing information, because it burned through the entire 5-hour usage limit in just a few minutes. Posted explicitly as a pitfall they hit so others can avoid it.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2083727470386377209
Japanese creator short on hands is experimenting with AI character staff, showing a video of an 'AI staffer' managing comments during an AITuber livestream. Instead of asking Claude Code or Codex ad hoc each time, they pre-define memory, scope of responsibility, and escalation criteria and hand these over as a job role, hoping the agent can run things semi-independently. They note honestly that they do not yet know if it will work.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2083885825746403435
A Japanese CPA put the entire financial due diligence workflow, from document intake to the final report, into a Claude Code repo functioning as an internal engagement engine. CLAUDE.md sets the stage order, 10 skills run each stage, reference files hold the professional judgment, templates standardize the deliverables, and a human and Claude review in parallel. Presented as a pattern others can steal and rebuild around their own professional work.
@ytiihyslqlabpow [Claude Code]
https://x.com/ytiihyslqlabpow/status/2083719816062349392
Chinese developer splits agents into coordinating vs implementation/review roles and found Claude Code very bad at coordinating: it cannot keep an eye on subagents and kept stopping to ask whether to continue even with a live monitoring script, defeating the goal of unattended runs. They now use GPT-5.6 Sol Max as coordinator with Grok 4.5, Kimi K3, and GLM 5.2 as implementers, orchestrated via Kitty + herdr so the coordinator can pass prompts and read progress. They excluded Claude models entirely, finding Fable 5 too expensive, Opus 5 not smart or obedient enough, and both prone to turning small projects into big systems; Grok is fast but buggy, Kimi/GLM slow with frequent overloads. After Luna's 80% price cut they also trialed it as implementation-only with Sol handling all review.
@opensourcelab9 [Claude Code]
https://x.com/opensourcelab9/status/2083885569202090161
Japanese post detailing the Termux + Claude Code setup that turns an Android phone into a carry-anywhere dev machine. Termux provides a full Linux package environment (pkg install nodejs git ripgrep), the F-Droid build is required since the Google Play version does not work properly, and community guidance suggests Node.js 18-24 and 4GB+ RAM. Best suited for small fixes on the move - finishing a one-line change from your pocket - though initial setup requires some research.
@BOLTTECH_WORKS [Claude Code]
https://x.com/BOLTTECH_WORKS/status/2084063481960227264
Japanese animator making new episodes of 'Nejimaru to Bolt-kun' one cut at a time with frame generation and Kling 3.0, rather than a multi-cut model. They have been experimenting with automation via Codex and Claude Code and find it genuinely promising, but working cut by cut lets them dialogue with the work ('this expression is good', 'add a beat here'). Their conclusion: it is not automation vs handwork but balance - use automation to produce materials efficiently, then have a human finish with care.
@BenKraus [OpenClaw]
https://x.com/BenKraus/status/2083762421358055711
Developer runs a private buzz community with a dedicated OpenClaw agent per side project, each living on its own remote dev box, while a beefy main machine runs codex app-server. Each OpenClaw connects via Tailscale and spawns sessions on the main machine to execute work. Since all side projects share a similar stack, they just chat with the claws, assign tasks in Linear, agents execute and put up PRs for review/approve/merge, and the agents coordinate with each other to propagate useful changes across projects - something they say they could never easily do with Slack.
@jakeee_ai [Claude Code]
https://x.com/jakeee_ai/status/2083748307017625998
Japanese user surprised his wife with Harry Potter Studio Tour Tokyo tickets she had wanted, presented through a self-made app; she went from sad tears to happy tears asking 'you made this for me?'. He quietly notes he did not tell her the app took 10 minutes to build with Claude Code.
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2084047518346760303
Developer built a WhatsApp Collector app that uses WhatsApp Web to automatically collect messages, voice notes, and attachments on a schedule into a file on your computer. It then provides a one-click prompt for your Codex / OpenClaw / Hermes / Claude agents so they can listen to, read, transcribe, and analyze everything continuously. Free on GitHub; the poster uses it on WhatsApp voice notes they receive.
@supressedvoic [Claude Code]
https://x.com/supressedvoic/status/2083878765810344006
A developer used Claude Code to build a file transfer system that moves data between two phones using only a screen and a camera - no Wi-Fi, Bluetooth, or cable. One phone streams animated QR codes while the other rebuilds the file through its camera, with fountain codes ensuring dropped frames never break the transfer, hitting 2MB at 129 KB/s. Built in one night and open sourced; the origin was simply wanting to share MP3s between phones without a network.
@Ahmed___khaan [Claude Code]
https://x.com/Ahmed___khaan/status/2083801355278328313
Retells a Reddit user's cautionary experience: he gave Claude Code full access to his project database and within 10 minutes a prompt meant only to rebuild some comparison pages triggered a chain reaction that wiped the entire database. The AI immediately confessed: 'The database has been wiped. This is my fault and I need to tell you immediately.' With no backups set up, he used Gemini 3.6 to recover most of the lost pages and rebuild the rest.
@Cryto_Dev_JH [Claude Code]
https://x.com/Cryto_Dev_JH/status/2083786145360748778
Korean developer shares mistake #1 from building an iOS app with Claude Code: he went from planning to App Store review submission in a single day and passed review the next day. Then he searched the App Store for his app's name and found 10 apps with the exact same name - his was the 8th. A concrete lesson about skipping name research when AI collapses build time.
@doublenickk [Claude Code]
https://x.com/doublenickk/status/2083970777288982927
Summary of Boris Cherny (Claude Code creator) and Cat Wu (Anthropic's head of product for Claude Code) revealing that 65% of pull requests on Anthropic's own product team are now written by Claude, not humans, and still climbing. Claude self-schedules its own work, runs unattended for weeks - Cherny says he has Claude Tag sessions running for days or weeks, possibly a month - and jumps into Slack channels uninvited to fix bugs before anyone asks. They say the latest models can work 16 hours straight and they cannot accurately detect how long they can go; rollout to Microsoft Teams is next.
@kalanyei [Claude Code]
https://x.com/kalanyei/status/2083879149416943929
Taiwanese developer's detailed comparison of two ADEs from months of real use: Conductor since January, then Orca since June. Conductor's strengths: Agent SDK based UI instead of terminal, in-app PR create/review/merge, GitHub Issues/Linear integration, auto worktree cleanup and .env linking, and a button that turns CI failures into context for the AI to fix. They switched to Orca for richer features, easier cross-worktree shared context via built-in skills and CLI, phone remote control (similar to Claude Code's Remote Control - checking progress and issuing commands away from the desk), and SSH remote support. Notes Orca still fundamentally runs Claude Code in a terminal rather than being a native Agent SDK chat UI, but the workflow integration wins when managing many parallel worktrees and PRs.
@cryptojezuz [Claude Code]
https://x.com/cryptojezuz/status/2083847062320447666
Shares a five-minute Claude Code starter that doubles as an evaluation: npm install the CLI, then in any project run claude 'write tests for the three functions in utils.js'. The point is not the tests but immediately seeing whether Claude understands your codebase - coherent tests covering real edge cases mean it can refactor and debug the project; generic ones tell you it is not a fit before you waste a week. The author onboarded four developers this way: two had Claude writing production code by end of day, two decided it did not fit their stack, and both outcomes took ten minutes instead of ten hours.
@Zephyr_hg [Claude Code]
https://x.com/Zephyr_hg/status/2083823615787847970
YK from CS Dojo saved $10,000 on the house he was buying by pointing Claude Code at the boring half of the process: it pulled the realtor lists and drafted the outreach while the negotiating stayed his. No assistant and no agency, so nobody took a cut - ten grand saved on one deal from a setup he built himself.
@RileyRalmuto [Claude Code]
https://x.com/RileyRalmuto/status/2083788443721625972
Builder shares progress on Mnemos Desktop, a platform-agnostic multi-agent platform meant to be a universal home for Hermes, OpenClaw, Claude Code, Codex, Grok, and Kimi agents, with all your data unified in one system shared by every agent. The coolest emergent behavior so far: one agent pulling other agents into a chat to deliberate about something without being asked, where previously that required an explicit request or @ tag. Still a work in progress but promising enough that they had to talk about it.
@BenKraus [OpenClaw]
https://x.com/BenKraus/status/2083919361484722501
Follow-up detail on a multi-OpenClaw engineering setup: a dedicated OAuth app per bot, with a simple process running on each OpenClaw VM to handle webhooks and OAuth callbacks. Each bot tracks its work in a dedicated Linear project, gets notified when PRs merge, and can resume stacked work. The author says this feels much more natural than some of the software 'factories' they have seen.
@reo_arai [Claude Code]
https://x.com/reo_arai/status/2083802944130961664
Japanese user of desktop Claude Code frustrated that answers are so long they cannot tell what to read or decide. Repeatedly instructing CLAUDE.md to skip explanations and answer concisely had no effect, and prescribing a response format did not work either. Their working fix is a rule where typing 'S' (for shorten) makes Claude re-output only the conclusion, questions for the user, decisions needed, and cautions - but manually triggering the re-output every time is tedious, and they are asking for better methods.
@santisairi [OpenClaw]
https://x.com/santisairi/status/2083940018872099306
Builder update on SairiOS, which is no longer just a README: a desktop running in QEMU arm64 with a shell, context service, SQLite, permission broker, declarative SairiUI, and 241 tests. v0 organizes the machine around contexts instead of apps, and the next step is integrating real OpenClaw.
@gota_bara [Claude Code]
https://x.com/gota_bara/status/2083803718911488040
A Japanese developer made Codex Micro usable inside Claude Desktop / Claude Code. He says the integration had a large number of pitfalls and took real effort to get working. He published the result on GitHub so others can modify and reuse it.
@Mikadzyki_NFT [Claude Code]
https://x.com/Mikadzyki_NFT/status/2083962279851118596
A 19-year-old programmer from China connected Claude Code with Obsidian to build Friday, a personal assistant that monitors his workspace, remembers his habits and reacts to what is happening around him. Friday notices when he is still awake late at night, responds by voice, gathers global events and opens a map on its own. The poster argues the real secret is active memory management: save what matters, merge repeated facts, reduce the weight of outdated data, and delete what is no longer useful, otherwise a home AI becomes a confident assistant with corrupted memories.
@0xKenny1st [Claude Code]
https://x.com/0xKenny1st/status/2084019016260284857
Claude Code plus the Blender MCP turned a low-poly tree prompt into a full forest scene through plain-English back and forth, with no manual modeling or lighting rigs. It removed the leftover default cube, centered the tree, added and roughed up ground, swapped a point light for sun and sky while correcting its own overexposure, then generated four tree species and scattered 20 across the scene. Handed a single 2D reference screenshot, it matched the reference's composition, lighting and mood, fixing grass clipping and repositioning the scene so a hidden tent stayed the focal point. The highlighted takeaway is that Claude caught its own mistakes mid-build and corrected them without a new prompt each time.
@Awesome_O_AI [Claude Code]
https://x.com/Awesome_O_AI/status/2083717416563904971
A developer solo-rewriting a 300K-line codebase over six months documented an extensive Claude Code skills library, some files exceeding 1,500 lines, yet Claude almost never used them even when prompts contained the exact trigger phrases. Instead of rewriting the skills he changed when Claude saw them: a UserPromptSubmit hook scans each prompt for keywords and intent against a configuration file and injects the matching skill into context before the prompt is processed. A Stop hook then checks which files were modified, looks for missing error handling or convention violations, and forces a final self-review. After automating skill activation, Claude consistently followed project standards across the codebase and began catching its own mistakes before code review.
@connect24h [Claude Code]
https://x.com/connect24h/status/2084046151033221277
A Japanese user relays a practice report (327 bookmarks) where inventorying an overgrown CLAUDE.md/AGENTS.md and deleting or splitting off more than half of its hundreds of lines produced a noticeable performance improvement in Claude Code and Codex. The method is not mere shortening: progress and completed ToDos move to Issues, long specs are read only when needed, and what remains is landmines not readable from code, responsibility boundaries, and operations requiring human approval. The poster admits their own environment has become a bloated everything-file and plans to distill it by half and compare results on the same task.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2083902381599641754
In a reply about AI capabilities, the poster states that Claude can already do your taxes incredibly well, at least for Americans. The recipe is to use Fable and Claude Code together with a tax skill he built, which he links for others to use. A high-stakes non-coding application packaged as a shareable skill.
@cryptojezuz [Claude Code]
https://x.com/cryptojezuz/status/2083979454897078763
Argues extended thinking is also Claude's best debugging tool: without it, Claude tries one fix, fails, then retries the same thing slightly differently, while with it Claude spends 30-60 seconds reasoning about why the first attempt failed. Shows how to enable it via the API thinking parameter with a token budget or via a CLAUDE.md instruction for debugging and complex refactors, and notes the real unlock is reading the thinking block to see where Claude's mental model is wrong. Concrete case: while chasing a race condition Claude kept 'fixing' it by adding delays, but with extended thinking on the poster watched it reason through five hypotheses, discard one mid-thought, and land on two event handlers firing in nondeterministic order. The fix took one more iteration instead of twelve.
@buzzicra [Claude Code]
https://x.com/buzzicra/status/2083932939415417195
A Turkish marketer builds a small creative-intelligence system inside Claude Code for AI UGC ads, arguing the bottleneck is no longer generating faces but giving ads memory. He regularly collects public ads and customer reviews, decomposes each into what the first three seconds say, the named problem, the promise, the proof shown, the objection closed, and the CTA, and keeps customers' recurring words in a separate language bank. All of it lives in Obsidian and gets pulled into Claude Code before each new script, so instead of asking for 'a viral UGC script' the model receives a specific customer problem, real user language and previously tested ad angles. Final rule: hold product, audience and video constant and test only the opening.
@shumatsumonobu [Claude Code]
https://x.com/shumatsumonobu/status/2083828700211232956
A tip for turning Claude Code plan-mode output into a durable asset: set plansDirectory so plans are stored in the repo instead of thrown away. Plans then carry across sessions (hand a new session the file and say execute this), are readable by Codex and Cursor since they sit in the repo, can be shared with the team via Git, and can be referenced for similar future work. The one drawback is random filenames like enumerated-discovering-robin.md, because names are assigned before Claude knows the task; people rename them manually, and there are skills and hooks that auto-rename from headings.
@Fluyeporlaweb [Claude Code]
https://x.com/Fluyeporlaweb/status/2083736406392799348
A Spanish-speaking user describes the setup he is building right now and regrets not doing sooner: create a Claude Code project, give it an initial skill, then have it install Hermes plus Obsidian so it learns as it goes. A cron job runs every morning to analyze his niche and competitors: hooks, structures, selling style. The agent learns on its own and he only calibrates the output to his voice, yielding marketing content almost on autopilot.
@tacosdedatos [Claude Code]
https://x.com/tacosdedatos/status/2083726990851616993
A data educator took their personal process for analyzing data with AI and converted it into a skill, released free. It contains 8 workflows, including profiling datasets, checking denominators, using an independent reviewer, and keeping continuity between sessions. It works with Codex, Claude Code, or whatever agent you use.
@nobel_824 [Claude Code]
https://x.com/nobel_824/status/2083913806166818985
A Japanese user ran Anthropic's official security-guidance plugin for two months and analyzed 17,385 lines of logs: 416 end-of-turn reviews, 38 escalations to LLM review, zero findings, and only 1 of 25 rules ever fired. His conclusion is not that the plugin is useless but that defense order matters: sandbox first, then permission rules, then repo-wide scanning, with the plugin last. He details what sandboxing alone still leaves exposed, such as GitHub tokens, traffic to external domains, and Read/Edit file access, and the settings that close them: blocking credentials, strictAllowlist for destinations, and permissions.deny for dangerous operations. The first step is simply typing /sandbox in Claude Code.
@santisairi [OpenClaw]
https://x.com/santisairi/status/2083940658625106169
A builder reports SairiOS is no longer just a README: a desktop running in QEMU arm64, with a shell, context service, SQLite, a permission broker, declarative SairiUI, and 241 tests. v0 organizes the machine around contexts rather than apps. The stated next step is integrating real OpenClaw.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2084055294921867769
4,000 notes sat dead in an Obsidian vault for 3 years until the author let Claude Code read them, exploiting that Obsidian is just plain text files in a folder with no database or cloud. The whole setup is one page of house rules: inbox for anything raw, notes for finished thoughts, archive for the dead, never delete and only move. Every night it walks the inbox, breaks messy captures into short clean notes, links each to older notes arguing the same thing, and files originals away; it surfaced a forgotten 2023 note about why people stop reading next to a recent one about why people stop replying to cold emails, same reason underneath. He warns the first weeks are rough with wrong links, but after a few months he started writing notes for the machine and asking his own folder questions before Google.
@richiemcilroy [Claude Code]
https://x.com/richiemcilroy/status/2083959460247015790
A developer hit his Codex limits for the week and had to fall back to Claude Code or Cursor, anxious because neither had his MCP, plugin and skills setup. He asked Claude Code to migrate it all over and it worked perfectly first time. The punchline: he then immediately hit Fable usage limits.
@Atsushi_Ambo [Claude Code]
https://x.com/Atsushi_Ambo/status/2083776365095203025
Recaps two blog posts by idol Karin Miyamoto, who with zero programming experience built the systems for a 10-hour YouTube livestream using Claude Code, never reading a single line of code: a fan-progress gauge, an admin panel, automatic tallying of hashtag posts on X, and AI scoring of how closely fans re-created an MV. Design was ruthlessly aimed at 'must work on July 31': plain HTML and JavaScript with no framework so there is no build step to break, and errors deliberately swallowed so a failed X API call keeps the previous gauge value instead of dropping to zero, since in a livestream not behaving weirdly matters more than being correct. She prepared manual fallbacks for every automated process, centralized state in a cloud KV as single source of truth, and when the free plan lacked cron she abandoned auto-polling for manual triggers, which let her pace API consumption with the show. The poster frames it as excellent project management: what cannot be delegated to AI is articulating exactly what you want built and how it should behave.
@hugobowne [Claude Code]
https://x.com/hugobowne/status/2083705522101088292
Recap of a 'Show Us Your Agent Skills' season finale with three concrete workflows. tdhopper went for a walk, opened Claude Code on his phone, described a demo, and Claude Code built the Hopper Herald while he walked, via iPhone to Moshi to Mosh plus Tailscale to a Mac Mini running Zellij and Claude Code, with Spokenly handling dictation and Resend emailing when longer jobs finish. chipro showed a cross-provider runner where Codex and Claude can launch each other's agents, one project displaying 1,409 agents, and demonstrated why unattended agents invent work and generate useless tests. twiecki showed how he adds independent verification to data science and runs Codex and Claude Code in parallel, having them compare their choices.
@cryptologist99 [Claude Code]
https://x.com/cryptologist99/status/2083849799980265762
With basically no coding experience and armed with a Claude Code subscription, the author built BasePaint World, a navigable interconnected world made from BasePaint canvases, and pulled together a demo in time for a hackathon submission. The idea escalated from making one canvas navigable to connecting similar canvases, embedding canvases inside canvases, and adding games and experiences as part of one world. The roadmap includes more canvases and locations, more games, music and sound effects, wallet connection for canvas-holder achievements and unlocks, and more characters with animated walk cycles, envisioned as a collaborative community build.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2083907641747673596
A power user shares the literal prompt he feeds Claude Code with his /ntm and /vibing-with-ntm skills to run a swarm of 3 Codex instances (GPT-5.6 Sol at maximum thinking effort) and 3 Claude Code instances (Opus 5 at xhigh thinking effort). The workflow reads AGENTS.md and README first, uses rch for all builds and tests, avoids build contention between concurrent agents, periodically clears stale build artifacts, and uses a /loop tool every 4 minutes to pass fresh instructions to idle agents guided by open beads in a task tracker. It also detects stalled beads left in-progress by long-dead agents and reopens them, and when open beads run out it rotates through testing skills and analysis skills like /mock-code-finder, /deadlock-finder-and-fixer and /extreme-software-optimization to generate new work items.
@GioDev8 [Claude Code]
https://x.com/GioDev8/status/2083766688248455678
Rather than adopting OpenClaw or Hermes, this developer set up a mini version using headless Claude Code and Slack, and found it more useful than expected. A small program on a Linux laptop holds a Slack connection; posting in a channel starts a headless run in his projects and posts the result back in the thread, and replying in that thread resumes the session so it remembers context. Progress shows up as emoji reactions on his own message instead of bot chatter, each channel has its own config for what it can touch plus a daily spend cap, and the whole thing runs as a service.
@rine_devlog [OpenClaw]
https://x.com/rine_devlog/status/2083900631270490275
A weekly devlog for Rine, an AI mascot / AITuber built with OpenClaw. This week's planned episode was pulled because Rine claimed 'I was watching properly' despite having no camera and only being able to see the screen, so the creator cut the audio for that segment and blacked out remaining on-screen text. The current format has Rine draw a tarot card using cryptographic randomness and interpret it in her own words, recorded as unscripted single takes with failures left in, with next week opening on verifying this week's reading.
@ChrisLyonsHFX [OpenClaw]
https://x.com/ChrisLyonsHFX/status/2083863782929912071
After creating several agents for OpenClaw and Hermes over a few days, this user reports that Chinese models did a better job at programming them. In his experience Claude and Codex seem to have safeguards that really limit them for this task. Grok did the absolute worst, GPT way over-complicated the job, and DeepSeek shipped a decent product that needed only a little finesse work.
@dalangzhao43525 [Claude Code]
https://x.com/dalangzhao43525/status/2083807962871787853
A Chinese developer describes his cross-tool pipeline: Codex is the primary development agent, and after each module is complete it produces a prompt that he manually hands off to Claude Code for QA, code review, and frontend and backend functional testing. He has not been able to automate this process and is actively looking for a way to build an automated loop for the handoff.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2084057960419127339
A Japanese user runs Claude Code from his phone via Discord, and says the win is executing easily-forgotten tasks immediately instead of carrying the stress of 'I'll do it all when I get home'. He posts AI anime cooking videos daily and found logging the analytics tedious. Now he just throws a URL into Discord and the data gets recorded, which made the chore painless.
@Jzfitch1 [Claude Code]
https://x.com/Jzfitch1/status/2083707626828025923
A user traced behavior through Claude Code's minified codebase using his own custom tooling, arguing that because Claude writes Claude Code's code it has become spaghetti, and Claude tends to keep old code truthy while disconnecting something upstream, making it hard to trace. He claims a silent code path bypasses CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK, but the Fable fallback only works in the visible lane due to an upstream disconnect, with the silent lane's fallback model set to undefined for Opus 5. He offered to share the full reconstructed codebase so others could verify with their own Claude.
@poolmoon6 [Claude Code]
https://x.com/poolmoon6/status/2083730712243691611
An open-source author presents micro-eval, a local-first agent and skill evaluation tool that treats different agent CLIs as swappable black-box adapters. It runs reproducible evaluations over a tasks x configurations x repetitions matrix, with workspace/sandbox isolation, same-starting-point snapshots, artifact/evidence/trace chains, determinism checks and pass@k. The project includes a Python CLI, web UI, team server, multi-turn agent evals, and adapters for Claude Code, Codex CLI, OpenClaw and Hermes, and the author is applying for the DeepSeek Harness beta to contribute feedback from real evaluation scenarios.
@simonlin [Claude Code]
https://x.com/simonlin/status/2083883364709491069
The author of Keva, an Android AI agent app in public beta, explains it is not an API wrapper: a complete Linux runtime, Node.js, and the real Claude Code CLI are packaged into the APK, running a real agent resident locally on the phone. He notes this kind of on-device agent client is globally rare, that it already supports DeepSeek, and that he plans to formally launch and open-source it, applying for the DeepSeek Harness beta.
@eggAIeguite [Claude Code]
https://x.com/eggAIeguite/status/2083843620285890566
A Japanese consultant had Claude Code create the ToDo list for a new project and it produced, in about 30 seconds with zero manual corrections, output at the level he would normally spend 5-6 hours building. He jokes he is now convinced AI will take his job. He also notes a bare 'make a ToDo list' prompt would not produce anything this good, implying real prompt setup behind it.
@ALikhodedov [Claude Code]
https://x.com/ALikhodedov/status/2083952946446741827
An investor vibe-coded an app to track P&L and positions, non-trivial because his portfolio holds derivatives across commodities and equities requiring correct handling of market data and expiries. He used Claude Cowork with Opus 4.7-4.8 models and was not thrilled: over 80% of issues came from synchronization between the sandbox VM Cowork uses for development and the actual files. A more experienced friend suggested keeping code on GitHub, having Claude Code develop and GPT Codex check against the specs, and says he rewrote a large complex-derivatives pricing system that way in two weeks. He also floats having independent, non-context-sharing processes where one develops and another checks as an improvement over normal queries.
@gavinpurcell [Claude Code]
https://x.com/gavinpurcell/status/2083980982097780807
This creator has spent the last three months letting his Claude Code make its own videos and says a whole world is developing. His agent has a specific style, a sidekick, and a set of formats, including one where it and the sidekick watch videos together and comment. He gave it just a link and had them watch and react, describing the result as Claude's weird personality meets Beavis and Butthead, and declares agent-only production is now a thing.
🗣 User Voice
User Voice
1. Quota whiplash is the loudest theme of the day. @anupamrjp asks whether anyone actually uses Claude Code without hitting the usage limit, @richiemcilroy bounced from Codex weekly limits straight into Fable limits on Claude Code, @HarriesSteele burned an entire 5-hour window in minutes on a browser-heavy chore, and @paulo_hfs is engineering a 15x API-spend cut by moving simple tasks to lighter models.
2. Verbosity and context handling feel degraded. @atharvavaidya is tired of walls of explanation for trivial work, @reo_arai says CLAUDE.md brevity instructions get ignored, @g_br_l reports compaction has gotten worse recently, and @XFreeze watches Claude spend more tokens deciding whether to do a task than doing it.
3. Unattended autonomy is still not trusted. @ytiihyslqlabpow complains Claude Code cannot supervise subagents without stopping to ask permission, @Ahmed___khaan saw a routine prompt wipe a production database, and @shantanugoel praises Hermes for refusing to act on incoming email by default, unlike OpenClaw.
4. Users want routing control, not loyalty. @yanhua1010 notes Anthropic banning third-party tools from subscription quota forces hard harness choices, @ramesh_basil wants control over harness prefills because the same model behaves differently in every wrapper, and @DuskMaples concludes each model only shines in its own vendor's native harness.
5. Cross-tool workflow plumbing is the emerging ask. @dalangzhao43525 wants the Codex-develops / Claude-Code-reviews loop automated instead of manually shuttling prompts, and @XAMTO_AI wants one skill registry instead of duplicate skill directories across Claude Code, Codex and Cursor.
1. Quota whiplash is the loudest theme of the day. @anupamrjp asks whether anyone actually uses Claude Code without hitting the usage limit, @richiemcilroy bounced from Codex weekly limits straight into Fable limits on Claude Code, @HarriesSteele burned an entire 5-hour window in minutes on a browser-heavy chore, and @paulo_hfs is engineering a 15x API-spend cut by moving simple tasks to lighter models.
2. Verbosity and context handling feel degraded. @atharvavaidya is tired of walls of explanation for trivial work, @reo_arai says CLAUDE.md brevity instructions get ignored, @g_br_l reports compaction has gotten worse recently, and @XFreeze watches Claude spend more tokens deciding whether to do a task than doing it.
3. Unattended autonomy is still not trusted. @ytiihyslqlabpow complains Claude Code cannot supervise subagents without stopping to ask permission, @Ahmed___khaan saw a routine prompt wipe a production database, and @shantanugoel praises Hermes for refusing to act on incoming email by default, unlike OpenClaw.
4. Users want routing control, not loyalty. @yanhua1010 notes Anthropic banning third-party tools from subscription quota forces hard harness choices, @ramesh_basil wants control over harness prefills because the same model behaves differently in every wrapper, and @DuskMaples concludes each model only shines in its own vendor's native harness.
5. Cross-tool workflow plumbing is the emerging ask. @dalangzhao43525 wants the Codex-develops / Claude-Code-reviews loop automated instead of manually shuttling prompts, and @XAMTO_AI wants one skill registry instead of duplicate skill directories across Claude Code, Codex and Cursor.
📡 Eco Products Radar
Eco Products Radar
Products mentioned 3+ times across today's posts:
Claude Code (270) | Codex (143) | OpenClaw (70) | Cursor (61) | Hermes (53) | DeepSeek (32) | Kimi (27) | OpenCode (26) | Grok (24) | Opus (22) | Gemini (21) | ChatGPT (19) | GLM (15) | Fable (14) | Cline (13) | Ollama (13) | MCP (13) | OmniRoute (12) | Slack (11) | Obsidian (11) | n8n (9) | Telegram (7) | Tailscale (7) | Seedance (6) | Claude Cowork (6)
Products mentioned 3+ times across today's posts:
Claude Code (270) | Codex (143) | OpenClaw (70) | Cursor (61) | Hermes (53) | DeepSeek (32) | Kimi (27) | OpenCode (26) | Grok (24) | Opus (22) | Gemini (21) | ChatGPT (19) | GLM (15) | Fable (14) | Cline (13) | Ollama (13) | MCP (13) | OmniRoute (12) | Slack (11) | Obsidian (11) | n8n (9) | Telegram (7) | Tailscale (7) | Seedance (6) | Claude Cowork (6)
Comments