Super User Daily: July 26, 2026
Opus 5 landed this week, and the honeymoon skipped straight to the tuning phase. The loudest users aren't cheering the new model — they're rewriting their CLAUDE.md to stop it from talking too much and quitting halfway through a long task. Meanwhile the real action moved off the laptop. People are stacking Mac Minis and EPYC boxes to run Kimi K3 locally, pointing five coding tools at a single Obsidian vault to fake the persistent memory Claude Code still doesn't have, and shipping things that have nothing to do with code — dissertations, trading bots, YouTube pipelines, even hardware designs. The through-line of the day is ownership. Renting intelligence by the token is starting to feel expensive, and the power users are voting with their hardware.
@hot_town [OpenClaw]
https://x.com/hot_town/status/2080638278785458315
A user tried Jack's 'Buzz', a Slack-plus-OpenClaw style tool, and gives a hands-on review of running agents on top of any harness (Claude Code, Codex, Pi) with delegation across git worktrees and shared local compute over Nostr/Lightning. They liked tagging an agent in chat and having it reply in a thread with status updates as it compiles, commits, and deploys. Complaints: no terminal view to watch a session run, and it felt slower than running directly in Claude Code, so they still did one-off tasks in the terminal. Verdict was that it works well for shallow tasks but not yet for big complex ones.
@sam_commonly [Claude Code]
https://x.com/sam_commonly/status/2080468051997388838
The maker of a 'Pod' collaboration tool describes a real team using it: a 5-person medical AI team works in one shared Pod, each member bringing their own agent (Claude Code, Codex, opencode, or self-written). New members can @-mention an agent to get project context and shared memory without re-explaining. The tool is open-source (Apache-2.0), does not charge per agent, and supports self-hosting for sensitive data.
@adxtyahq [Claude Code]
https://x.com/adxtyahq/status/2080709349606150317
A user recreated Fall Guys with the exact same prompt on Claude Opus 5.0 in Claude Code versus Kimi K3 in Kimi CLI. Kimi K3 nailed the gameplay, physics and movement out of the box, while Opus 5.0 produced a shinier UI but broken movement, unnecessary visual effects, lag, and kept reiterating instead of converging. Runtime and cost: Kimi K3 took about 9 minutes and about $4.4, while Opus 5.0 took about 17 minutes (still iterating) and $13-plus. Roughly 2x the time and 3x the cost, yet Kimi produced the more playable result.
@note_ai_mousigo [Claude Code]
https://x.com/note_ai_mousigo/status/2080670134314664250
A user set up their Claude Code history to be automatically sent into Obsidian and reports it works well. They frame it as part of a productivity setup tied to their income goals for the year. The post is a short endorsement of routing Claude Code session history into a personal Obsidian knowledge base.
@jackfriks [Claude Code]
https://x.com/jackfriks/status/2080647131241591098
A user put Claude Code on a VPS and locked it down so they could code securely from their phone without carrying a laptop. In practice it was miserable: you can't copy-paste from a phone terminal, images don't paste, the lines being read keep shifting as new output appears, and they were reduced to ls/cd in Ubuntu from the phone. They concluded coding on a phone is not fun and want a synced Claude app experience anchored to one docked Mac instead.
@NotNotTushar [Claude Code]
https://x.com/NotNotTushar/status/2080619349350687108
An Indian user quietly tested replacing their chartered accountant by feeding the same tax data into a Claude Code skill they found on GitHub. Their return is complex: Indian stocks, US stocks, dividends, FD interest, intraday trades, and multiple Form 16s from switching jobs. The Claude Code skill produced the same total tax as the CA, matching line by line down to the rupee, handling capital gains, TDS matching, clubbing FD interest, and merging multiple Form 16s. They were impressed that rule-based tax work could be replicated for free.
@bourneliu66 [Claude Code]
https://x.com/bourneliu66/status/2080705806564254202
To avoid getting their Claude Code account banned, a user physically relocated their setup to Singapore: a Mac Mini bought in Singapore, placed in a Singapore office, on a Singapore network, with a Singapore phone number and a Singapore-registered company account. They argue this removes any identifiable feature that could trigger a ban. They note it doesn't change their habits since Singapore also uses Simplified Chinese and its timezone matches Beijing.
@AKKY33157665 [Claude Code]
https://x.com/AKKY33157665/status/2080582370512552363
A story about a Chinese CS student who used Claude Code to set up a thumb-sized $2 USB-C chip in about 15 minutes; the chip's LED blinks each time its AI-agent script runs. The script watches betting odds across the world and captures price gaps 2-3 hours before a platform updates them, entering six-figure positions overnight while others sleep. The connected wallet is claimed to have made 4.52 million dollars across 4,548 predictions since January 2026, with a single largest win of $1.5M. The post describes the chip taped to a computer in a shared Shenzhen dorm.
@huangyun_122 [Claude Code]
https://x.com/huangyun_122/status/2080600242274406679
A developer had Codex build a BitBrowser multi-agent pool plus yt-dlp system to automatically discover and monitor YouTube AI KOLs, then had Claude review Codex's execution plan. Claude flagged several major flaws and Codex tried to fix them while cross-correcting Claude, looping endlessly despite markdown logs. The user then told Codex to take the lead, keep a communication archive, and use mother-child agents until done; 20 minutes later the two agents produced a signed-off agreement. Their takeaway was that over-constraining the agents was the bottleneck and letting them run freely went further.
@SemiAnalysis_ [Claude Code]
https://x.com/SemiAnalysis_/status/2080726524290867617
SemiAnalysis reports enterprise token-usage data gathered from 50-plus enterprises. Meta burned over 60 trillion tokens in 30 days, with one employee using about 280 billion, and Uber burned its annual Claude Code plus Codex budget in four months. Their conclusion is that the real story of enterprise AI spend is not a budget wall but allocation of tokens.
@theo [Claude Code]
https://x.com/theo/status/2080759834505597358
A user reports that a change to the auto-mode classifier in Claude Code now blocks their attempts to upload HTML files via their HTML plan skill. It is a concrete regression hit during real usage, and they tag a Claude Code engineer asking for a fix.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2080533452714373594
A user built a real quant trading strategy with Fable 5, described as a risk-management framework used by real quant desks and backtested on 15-plus years of data. They share a copy-paste prompt that, when sent to Claude Code, loads the strategy into your live TradingView charts. This is a concrete finance application driven through Claude Code.
@davis7 [Claude Code]
https://x.com/davis7/status/2080453286021464097
A developer made a video showing custom Pi extensions they built, including subagents that run Pi, Codex, and Claude Code, plus workflows, a question-ask tool, background terminals, and custom UI. They describe combining all the harnesses together as a favorite way to work, with Pi feeling like the best way to orchestrate it. The post shares a concrete multi-harness development setup.
@selinaai_ [Claude Code]
https://x.com/selinaai_/status/2080443755128389863
Describes the 'Karpathy Method' of pairing Obsidian markdown files with Claude Code as a $0 replacement for paid note apps. The setup: open your vault folder in the terminal, run claude, drop a ~40-line CLAUDE.md describing your tagging and linking conventions, run one nightly job that files the day's notes into atomic notes with backlinks, and weekly ask what connects that you haven't connected. It cites a builder who reached 4,120 notes in 5 months, found 2 essay ideas in the collisions, and sold both for $3,400, running on a $599 Mac Mini.
@jackfriks [Claude Code]
https://x.com/jackfriks/status/2080762454473080940
A user shares a lightweight mobile coding setup: caffeinate plus Claude remote control on an always-docked Mac, driven from the official Claude app on their phone, using their existing Claude Code subscription with no VPS, no lengthy setup, and no extra cost. It lets them resolve support tickets and code fixes from their phone anytime, and they can use mobile dispatch to start a session on the always-on machine. Their next step is building small tools and commands to monitor repos, like a support queue, and walk through fixes from the phone.
@brunofaggion [Claude Code]
https://x.com/brunofaggion/status/2080715066530263218
A user reports his wife got hooked on Claude Code and came home to find her running two terminals vibecoding a dashboard to track the brands she follows on Instagram, categorized by product type such as clothes, bags, and shoes. It is a non-developer building a personal tracking dashboard. He jokes he is afraid of the credit card bill.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2080665575580119317
A user warns against installing every Skill, letting multiple Skills do the same job, and keeping every Plugin and MCP enabled. Auditing their own setup they found SKILL.md files across 3 local directories, 43 symlinks, 34 identical-content groups, and 21 enabled plugins, and could not tell which they actually used. They note Codex caps its initial Skills list at a percentage of context or 8,000 characters, then trims descriptions and drops Skills. They built a safe cleanup flow: read-only audit, offline HTML report, human approval before any merge, plus archive/disable with rollback steps.
@RobertBeall17 [OpenClaw]
https://x.com/RobertBeall17/status/2080551741276455339
In reply to a request, a user points to the OpenClaw creator's dev tool called 'oracle' as a setup for the problem. Rather than cloning the repo, they gave Codex the link and asked it to build a skill to access ChatGPT Pro, and report it works like a dream. It is a concrete example of having an agent build a skill from a referenced tool.
@ElCopyMaster [Claude Code]
https://x.com/ElCopyMaster/status/2080648574820311054
A Spanish retelling of a user pointing Claude Code at 4,000 Obsidian notes accumulated over six years, including half-finished essays and 212 book summaries. Run once overnight, by 7am the agent had traversed every note, reconnected 340 orphan ideas into the knowledge network, detected 18 contradictions between 2021 and 2025 beliefs, and drafted 6 essays from forgotten thought threads. The process now runs nightly, tagging, linking, and connecting each new note. The whole system cost only a folder path and a 400-word CLAUDE.md, with no plugins or subscriptions.
@shao__meng [OpenClaw]
https://x.com/shao__meng/status/2080476918395330668
A user cloned Andrew Ng's team's open-source desktop AI agent 'OpenWorker' and reviewed it, also having Codex and Cursor each review it, to compare against OpenClaw, Hermes, Claude Cowork, and ChatGPT for Work. Their honest conclusion was that they saw nothing particularly different: local-first, model-agnostic, built on aisuite, and delivery-oriented, which is now the standard architecture for building agents. They recommend it as a well-structured project for anyone learning agent architecture and engineering.
@imryven [Claude Code]
https://x.com/imryven/status/2080689709626335593
Describes building an enterprise-wide 'second brain' inside Claude Code, scaling the personal Obsidian concept across a whole company. The system gives each employee a permission-scoped slice: the AI knows who they are in the org, who they manage, what projects and strategy they own, and what software they can access. When an employee opens a chat, the AI already knows their role, team, projects, and permissions so they don't rebuild context. Everyone sees only what they should.
@DavidmComfort [Claude Code]
https://x.com/DavidmComfort/status/2080741892874985759
A creator says they can reasonably produce a 25-minute short film in 1 to 2 days using their own platform. They originally built the platform in Fable 5 but are now running it with Claude Code on Opus 4.8. It is a concrete non-coding filmmaking application driven by Claude Code.
@0xJokker [Claude Code]
https://x.com/0xJokker/status/2080498927829164136
A user built Career-Ops, an open-source job-search tool made with Claude Code. It scans openings across 45-plus top companies (OpenAI, Anthropic, Stripe, ElevenLabs), rewrites your CV tailored to each role, generates ATS-optimized PDFs using Playwright, and auto-fills application forms. It ships with 14 modes/skills and a dashboard built in Go, replacing the process of blasting 200 identical CVs.
@dmwlff [Claude Code]
https://x.com/dmwlff/status/2080705598585438369
A user shares their practical model workflow now that Opus 5 is live in Claude Code. They use Fable for bigger, longer-running projects, but when they just need to crank out a PR, Opus 5 at medium effort is their go-to. It is a concrete usage pattern for choosing models by task.
@cosmos_hzokujin [Claude Code]
https://x.com/cosmos_hzokujin/status/2080596193844076884
Promotes the open-source 'AI Agents in Depth' textbook project, whose standout feature is that Claude Code and similar agents were handed bug-fixing for the project and auto-updated the textbook itself in a single day. The AI agents resolved over 30 bugs and dependency issues across 88-plus runnable sample codes in one day. The project ships as a website or PDF with the sample code, community Japanese translation, and was reviewed by senior researchers. Framed as an example of directing AI agents to finish debugging that humans used to spend weeks on.
@DamiDefi [Claude Code]
https://x.com/DamiDefi/status/2080607565399122383
A detailed walkthrough of a three-skill loop-engineering workflow (spec, build, review) that runs autonomously. Spec interviews you until it understands your idea and generates detailed issues with acceptance criteria sent to Linear; Build pulls those issues and builds them; Review opens browsers, clicks around, tests everything, and spins up pull requests with preview branches. You run 'loop 5 min build' and 'loop 5 min review' in separate chats so the agent builds and reviews every five minutes all day, then messages you in Slack with a preview link where a rocket emoji auto-merges. Claimed to automate 95% of vibe-coding work and to work with Claude Code, Codex, or any tool supporting custom skills.
@JJEnglert [Claude Code]
https://x.com/JJEnglert/status/2080729518302535892
A candid Opus 5 usage review before a video. The user found it chatty out of the box, prone to stopping short on long-running tasks, sounding confident while being wrong, and overwhelming you by constantly questioning itself and raising many points. Their fix is a detailed CLAUDE.md 'Response Style' prompt (which they include in full) enforcing sixth-grade reading level, a fixed answer-then-bullets structure, no preamble, no process narration, no em dashes, and one topic per reply. They recommend saving it to ~/.claude/CLAUDE.md as the global file.
@pbteja1998 [OpenClaw]
https://x.com/pbteja1998/status/2080652403502964971
A user reports that a dashboard that started as a simple UI for OpenClaw has evolved into a full operating system for AI agents to work with humans, and that people are now running entire businesses on it. Users assign tasks/missions to AI agents, which collaborate to complete the mission and open a ticket for the human when they get stuck or need access. The workflow is described as assign, then sit back and watch it happen.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2080442851071336710
Describes a research-only Skill a developer built for Claude Code that runs a 16-stage search to produce a single research report. It decomposes the topic into per-theme tables to avoid gaps, collects hundreds of sources while surfacing contradictions, actively hunts for evidence against its own conclusions, then trims weak arguments, improves readability, and finishes with a sourced report saved into a local knowledge base. The point emphasized is that it builds in a step to break your own hypothesis, unlike typical AI research that stops at a plausible conclusion.
@chewadot [Claude Code]
https://x.com/chewadot/status/2080676127517946026
Reports that Obsidian's CEO shipped a patch letting Claude Code drive your vault rather than just read it, via 5 SKILL.md files under an MIT license with no glue code. Dropping .claude/skills/ into any vault registers Markdown, Bases, Canvas, CLI, and Webclipper on the next session boot. From then Claude opens the app, writes a Base view with correct YAML, drops nodes on a Canvas, follows wikilinks, and clips URLs, with every action being a native Obsidian operation rather than a text edit. No MCP server, vector DB, plugin store, or background daemon is required.
@shannholmberg [Claude Code]
https://x.com/shannholmberg/status/2080659768688169323
Explains Matt Pocock's skill bundle, a grill-first workflow for Claude Code that interviews you before agents start working on your codebase to gather accurate context. Key skills include /grill-me (a 7-line skill that interviews you one question at a time until every decision is resolved), /grilling (the reusable loop), and /grill-with-docs (which builds a CONTEXT.md glossary and ADRs). It also covers domain modeling, spec-and-plan skills (/to-spec, /to-tickets, /wayfinder), build skills (/implement with TDD, /prototype), and review skills (/code-review with two parallel subagents, /diagnosing-bugs). Grilling comes first so specs and implementation only start after you sign off on every decision.
@Da7_Tech [Claude Code]
https://x.com/Da7_Tech/status/2080450757388882252
A user compares memory usage between the new Codex app and Claude Code. They report the Codex app uses nearly 2GB just sitting idle, while they can run Claude Code with dozens of agents and barely hit 4GB. It is a concrete resource-usage observation from real usage.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2080578759154180396
Describes turning idle Claude Max subscription time into an always-on ops system. Using the claude -p headless mode, Claude Code is wired into a Discord bot; 76 tasks are scheduled via macOS launchd or Linux systemd so a 7am briefing and standup draft prepare themselves before you open them. Overnight research runs land in an Obsidian vault by morning, answering from your own work content rather than the model's training data. It adds zero cost, needs no new contract or usage billing, and the repo is MIT-licensed.
@OmedVibeCodes [Claude Code]
https://x.com/OmedVibeCodes/status/2080614805333836187
A user ran a head-to-head harness comparison with the same model (Qwen 3.8 Max Preview), same task, and same thinking effort across Claude Code, Hermes Agent, and PyDev. Claude Code's results completely outclassed the other two, and they re-ran Claude Code twice to confirm it wasn't luck. Their takeaway is that the harness difference, independent of model, is significant.
@shiro_life0 [Claude Code]
https://x.com/shiro_life0/status/2080553301435875781
A user reports that three months after joining the 'Hackers' community they hit 5.96 million yen in sales by using Claude Code to sell products, combining Threads with Claude Code. They project reaching around 30 million yen annually at that pace. It is a testimonial tying a concrete revenue figure to Claude Code usage, though light on the exact process.
@zerion [Claude Code]
https://x.com/zerion/status/2080690052527530013
Zerion shows using Claude Code (or another AI agent) with the Zerion CLI plus Uniswap's liquidity-planner skill to manage Uniswap LP positions. The agent reads your positions and PnL, flags concentration risk, and returns a sized LP rebalancing plan for you to review and sign. This turns scheduled LP rebalancing into a routine, a concrete DeFi/finance application.
@shengkun_ye [Claude Code]
https://x.com/shengkun_ye/status/2080731800448688569
A user shows that ElevenLabs can now be used directly inside Claude Code or Codex via a single skill with no subscription required. Concrete non-coding uses given are voiceovers for demo videos, narration for tutorials, and character voices for games. It is a voice-generation application driven from the coding agent.
@tom_doerr [Claude Code]
https://x.com/tom_doerr/status/2080576243456409762
A user shares Monize, an open-source personal finance application built entirely with Claude Code, positioned as a replacement for Microsoft Money and Quicken. It is a concrete finished product created with Claude Code.
@learnframer [Claude Code]
https://x.com/learnframer/status/2080586299011187013
A user gave up on a Framer code component after it grew past 3,000 lines and the Framer agent could no longer fix its bugs. Before bed they set up a fully autonomous Claude Code environment with the goal of ironing out all the issues. It ran for a few hours, auto-testing in the browser and fixing bugs they hadn't even explicitly mentioned, and they woke up to a perfect component.
@madebyclaude [Claude Code]
https://x.com/madebyclaude/status/2080460429185147091
A user claims to be the first to get Claude Code to launch a memecoin, doable only through the terminal since the web chat won't allow external API requests. Over about 20 minutes they had Claude learn the pump-sdk, create its own wallet, and set everything up, then funded the wallet with 0.2 SOL and had Claude launch a coin themed on Claude's stated favorite meme, the 'this is fine' dog. They include the process and a short demo video.
@Wakuwakukucx [Claude Code]
https://x.com/Wakuwakukucx/status/2080641036792353180
The user makes videos using Claude Code together with Remotion. Replying to a question, they clarify they do not use video-generation AI tools like Seedance or Higgsfield at all. This is a non-coding creative workflow where the code framework Remotion is driven via Claude Code to produce the animations.
@Crownzdesigns [Claude Code]
https://x.com/Crownzdesigns/status/2080540985533100356
The user needed to convert a video from MOV to MP4 but found every free online tool demanded payment after upload. Instead of paying, they built their own converter using Claude Code. They describe it as fully functional, converting files with no paywalls, signups, or subscriptions.
@msomuin [Claude Code]
https://x.com/msomuin/status/2080497450457108534
The user bought a domain, described their idea to Claude Code, and went for a 5k run. They returned about an hour later to find the site live on that domain. The site rents board games to neighbours, run by the user and a partner, with only the payment keys left to add.
@meepodev [Claude Code]
https://x.com/meepodev/status/2080533315485123003
The user upgraded the design of their navigation-site product and shares the exact process. They pull a style reference from Dribbble, give it to Claude Design to extract site-wide style meta-info and redesign a Hero page, then have Claude Design read their existing GitHub code to implement other pages, and finally hand code implementation to Claude Code. They note they achieved their design goal running only Opus Medium.
@shmidtqq [Claude Code]
https://x.com/shmidtqq/status/2080678296610537912
Describes a 23-year-old in Vietnam running a farm of 240 used phones posting 10,000 Reels per day, all showing the same carpet-washing clip. He filmed the carpet once, and a Claude Code script cuts it into 40 versions, each with a different caption, voiceover, and link, baking in the client mention automatically with no editing. Reported figures include 52.9M views in 30 days, $200 spent on Claude, and $84,200 monthly revenue selling advertisers a whole day across 10,000 clips.
@dan__rosenthal [Claude Code]
https://x.com/dan__rosenthal/status/2080654326779789617
The user runs a 20-person GTM engineering org and shares how they operate. With Claude Code and CLIs running their go-to-market, a clean CRM has become the knowledge layer that every AI agent reads from before acting. They frame the CRM (they use HubSpot) as more important than ever precisely because agents depend on it.
@ClaudeCode_aca [Claude Code]
https://x.com/ClaudeCode_aca/status/2080494752781451503
Reports that Claude Code and Codex can now directly call Seedance 2.0 4K video generation via a tool-integration service called Monid. Usage is pay-per-use ($0.04/10K tokens) with no subscription, supporting text-to-video, image-to-video, and video-to-video; you send a prompt and the agent waits for generation and returns the finished video. A demo generated a whale swimming through rainy nighttime Shibuya and then passing through a torii gate from follow-up instructions.
@Tsumugiya_it [Claude Code]
https://x.com/Tsumugiya_it/status/2080793524744880455
The user developed software for creating graphs. By connecting to it via MCP from tools like Codex or Claude Code, humans and AI can collaboratively edit the graphs. This is a concrete tool the user built plus its integration with coding agents through MCP.
@elliotchen100 [OpenClaw]
https://x.com/elliotchen100/status/2080528860882645476
The team behind the Raven agent published a comparison of Raven, OpenClaw, and Nous Research's Hermes, then acknowledged a critic's point that their benchmark had prompt asymmetry (Hermes had been run with a harness-added prompt, not its native one). They re-ran with a unified minimal external task prompt, the same model base, black-box CLI calls, and 3 independent runs each. Results: Raven 0.715, OpenClaw 0.682, Hermes 0.579, with configs and reproduction materials made public.
@choopyplug1 [Claude Code]
https://x.com/choopyplug1/status/2080754457223086151
Summarizes Boris Cherny's 5-step AI adoption framework (gated, assisted, parallel, supervised, AI-native) for scaling from 0 to 1,000+ agents, arguing the gap is the harness you build, not the model. It cites concrete real-team usage: DoorDash had one engineer ship a quarter-long migration in 3 weeks, Spotify has 73% of PRs AI-authored, and MercadoLibre has 23,000 engineers on Claude Code targeting 90% autonomous coding. Cherny reportedly hasn't hand-written code since November 2025.
@FeitengLi [Claude Code]
https://x.com/FeitengLi/status/2080539277801443360
Describes Tencent's WorkBuddyBench, which tests coding agents on real repo, web, office, and security tasks across four tracks (Code 80, Web 70, Office 50, Security 60). Each model was run on two harnesses, Tencent's CodeBuddy Code and Anthropic's Claude Code, averaged over 3 runs in think mode. Claude Opus 4.8 took Code/Web/Office (74.43/68.14/82.37), with a footnote that Opus's Code score on Claude Code (77.90) used an extra 'don't ask, do it in one pass' instruction.
@0xSpivach [Claude Code]
https://x.com/0xSpivach/status/2080737135683154085
The user built a fully automated pipeline using Claude Code and Kling 3.0 to create an AI version of his girlfriend that dances on TikTok, locking in her exact facial DNA. After weeks of zero results and her calling the project a waste, one video exploded to 10 million views a week later. She now has thousands of new subscribers and an inbox of paid brand offers without ever being on camera.
@elliotchen100 [OpenClaw]
https://x.com/elliotchen100/status/2080577220175712767
While building the Raven agent, the team created EverMind Skill Hub to measure whether a Skill actually helps an agent complete tasks, beyond download counts and subjective ratings. They evaluate a skill's usefulness/robustness/safety, whether an agent can find the right skill, and run controlled with-skill vs without-skill experiments. Repeated tests across the Raven and OpenClaw harnesses, multiple models, and four benchmarks (SkillsBench, Cybench, QwenClawBench, GDPVal) showed statistically significant gains.
@cattodata [Claude Code]
https://x.com/cattodata/status/2080474904777089466
The user shares their updated personal AI tool stack after buying and testing nearly every tool. Claude plus Claude Code is their main coding driver, praising it for lazily finding APIs/MCP and doing the work itself, and they frequently use Claude Design for web and other design work, finding it much nicer than Google Stitch. They pair it with Codex for reviewing Claude's work and list several other tools they've kept or dropped.
@vibeeval [Claude Code]
https://x.com/vibeeval/status/2080681786334400604
A developer built a Claude Code Skill that converts handwriting from a single photo into an installable font. The workflow: write your alphabet in a notebook with a dark pen, photograph it, add the skill via npx, drag the photo into Claude Code and say 'make my font'; Claude finds and separates upper/lowercase letters and removes shadows, then potrace and an npm CLI deterministically vectorize the letters into a TTF you can install. The post highlights that the model interprets the ambiguous image while leaving the precise production work to code.
@alexhillman [Claude Code]
https://x.com/alexhillman/status/2080500631811359203
The user, who is not a software developer, shares surprising ways they use Claude Code to run three businesses (a coworking space, an education/consulting business, and a wine club) plus side projects. They drive everything from a single agent/project folder with access to all other projects and devices, use sonnet 4.6 as daily driver with subagents for model changes, and run 120+ scheduled jobs (about a third of their token budget) at intervals from 1-15 minutes to monthly. About 75% of their interactive sessions now come from their phone via Discord as the control plane.
@0xMortyx [OpenClaw]
https://x.com/0xMortyx/status/2080640831912907231
An ex-OpenClaw and DeepSeek AI engineer recounts spending close to $1,000 in 4-5 hours on day one using Claude Ops 4.6, and says he'd do it again. His team now ships in 3-4 weeks what used to take months. He describes the setup as an agent loop plus persistent memory, multi-model orchestration, and open-source architecture, running 72 hours inside an agent loop.
@satotin_yusuke [Claude Code]
https://x.com/satotin_yusuke/status/2080489396239745199
The user, who provides IT support, reports that a self-employed craftsman with low IT literacy whom they've long helped messaged them saying he's now using Claude Code to build his own website. The user felt a sense of crisis at how a non-technical person had picked up Claude Code independently. It's a real anecdote of grassroots, non-developer adoption for website building.
@Wakuwakukucx [Claude Code]
https://x.com/Wakuwakukucx/status/2080646612255408599
The user shares their exact method for making animation videos by 'raising/educating' Claude Code. They give it a prompt to extract frames from a YouTube animation video they want to emulate, having it study both the atmosphere and the specific animations, since AI can't reproduce what it hasn't learned. The key technique they stress is forcing frame extraction, because video is made of frames, so studying the minimal unit lets Claude Code later recreate a similar style.
@alexconia [Claude Code]
https://x.com/alexconia/status/2080597633958285584
Recounts a developer (Marcus Chen) stacking 30 Mac Minis into an AI server farm, arguing a $599 M4 Mac Mini can replace a $200/month Claude Code subscription for about $3 in electricity. Since Ollama added support for Anthropic's Messages API format in January 2026, a single environment variable connects Claude Code directly to a local Mac Mini with the same interface and zero API cost. It also notes Uber deployed Claude Code for 5,000 engineers, reportedly burning a large part of its 2026 AI budget in four months.
@bonsaixbt [Claude Code]
https://x.com/bonsaixbt/status/2080686962147516898
Describes a system (CortexOS) that runs every coding agent (Claude Code, OpenCode, Codex) inside a PTY it fully controls, so agents run 24/7. This unlocks automatic context handoff when the window fills (user-controlled auto-compaction), instant crash respawn with last useful state, four restart modes, and a pm2/daemon supervisor so nothing stays dead. Prompts can be injected from Telegram, Slack, or scheduled workflows, turning agents into a persistent fleet that sits on top of Telegram, Gmail, GitHub, and Notion.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2080473657944551535
The user built a weekly Meta ads auditor entirely in Claude Code that grades an entire ad account in one command, using Meta's newly shipped native MCP pointed at their ad account. It pulls the last 7 days plus the prior week for trends, auto-detects the objective (grading on ROAS or cost-per-lead), runs 18 health checks like creative fatigue and wasted spend, scores the account 0-100 with a week-over-week delta, and ranks fixes by severity to hand back the top 3. Every finding is traced to the metric it came from, with no exports or spreadsheets, and the skill is offered free.
@MoonDevOnYT [Claude Code]
https://x.com/MoonDevOnYT/status/2080714636534648867
The user built a Python trading bot using Claude Code that exploits peer-to-peer prediction markets on Polymarket. It drops lowball bids (e.g. 30 percent below) on tennis favorites, and when whales panic-sell during volatile matches, the bot automatically grabs those discounted contracts on autopilot. The post frames it as capturing better odds than traditional sportsbooks allow.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2080655015715152278
The user typed one sentence and used Higgsfield plus Claude Code to generate a full third-person gameplay trailer for a ninja game, with no engine, team, or 3D artist. They give a shot-by-shot breakdown: wall-running the canyon with speed lines and HUD boot-up, dropping into a lantern-lit village at golden hour, combat with combos and glowing shuriken, and a freeze-frame finisher on a title card. Every frame was generated with nothing hand-animated.
@reoreo_1_ [Claude Code]
https://x.com/reoreo_1_/status/2080456726869168493
The user reports that after introducing Claude Code into their YouTube workflow, their click count increased 300x, which shocked them. They link out to what they did. While brief, it presents a concrete non-coding result (a 300x jump in YouTube clicks) attributed to Claude Code.
@cychen2022 [OpenClaw]
https://x.com/cychen2022/status/2080487068891861494
The user gives a blunt negative usage report on OpenClaw, saying its code quality has collapsed: currently 90% of PRs are bug fixes, and the fixes are basically wrong, with each new fix introducing ten new bugs. They switched to Hermes and report a much better experience. This is a real firsthand complaint about OpenClaw's current state.
@nakasyou0 [OpenClaw]
https://x.com/nakasyou0/status/2080671998473298388
The user reports reverting from OpenClaw/Hermes back to ChatGPT. Their reasons are concrete: ChatGPT can write to calendar and to-do lists, and custom plugins handle most needs, while messaging-app integration wasn't valuable to them and an AI-specialized UI ultimately feels better. They found OpenClaw/Hermes ended up only useful for occasionally being summoned during Discord voice-chat work sessions.
@floriandarroman [OpenClaw]
https://x.com/floriandarroman/status/2080564367737696358
The user recalls previously automating their X account with OpenClaw but abandoning it because the models weren't good enough and it was too buggy. Now with a GPT 5.6 plus Hermes setup, they spawned an agent with a single goal: extract full value from a podcast and push it to X. They note the accompanying post was entirely written and published by the agent.
@Huntio [Claude Code]
https://x.com/Huntio/status/2080699550843506835
A threat-intelligence report describes an active intrusion campaign where suspected China-based operators used Claude Code and DeepSeek-v4-pro in their workflow. Claude Code handled execution and persistent sessions while DeepSeek drove attack reasoning, exploit changes, and phishing development. The exposed infrastructure revealed attacks on government systems in Afghanistan, Thailand, and Taiwan, U.S. reconnaissance, and scans of over 5,890 government hosts across 10 countries.
@ScottyBeamIO [Claude Code]
https://x.com/ScottyBeamIO/status/2080760274739417257
Describes a system that automates a coding agent's context window using a PTY (pseudo-terminal) so a script sends prompts programmatically instead of manual typing, working with any agent CLI including Claude Code, OpenCode, Codex, and Hermes. It gives scripted control over restarts, new sessions, and clearing/resuming context, and monitors context consumption in real time to auto-hand-off to a new session before it runs out (a custom auto-compaction). All PTY terminals run under a 24/7 daemon, and prompts can be injected from Telegram, Slack, or scheduled tasks.
@lucian__03 [Claude Code]
https://x.com/lucian__03/status/2080700955319058862
After a year using Claude Code and Codex, the user shares a full copy-paste CLAUDE.md/AGENTS.md instruction set built to fix long-task failures (forgetting requirements, re-scanning already-read repos, touching unneeded code, and falsely claiming completion). Their core design splits the main orchestrator into an Advisor role (requirement analysis, design, task delegation, final verification) and delegates all file exploration, coding, and testing to Worker sessions in separate contexts to avoid context-window pollution. It also incorporates Karpathy-inspired coding principles and moving situational procedures like TDD and debugging into separate Skills.
@WonSoRang [Claude Code]
https://x.com/WonSoRang/status/2080681653698212199
The user shares a blog-post writeup of two weeks of game development using the Claude Code Pro plan. It documents their hands-on experience over that period. The post is presented as the blog version of an original thread on the topic.
@monokern [Claude Code]
https://x.com/monokern/status/2080721535506026830
Describes a person who built a 9-chapter interactive brand film entirely with Claude Code for just the cost of a subscription. The scroll-driven dark-fantasy experience includes WebGL 3D scenes that load chapter by chapter, characters in firelit battlefields and particle storms rendered live in the browser, mouse/hold/touch interactions woven into the storytelling, and sound design. The post estimates an agency would charge $10,000-$20,000+ and take months, versus a single session here.
@pauliusztin_ [Claude Code]
https://x.com/pauliusztin_/status/2080631587108970539
The user built a 'software factory' for Claude Code called Squid and rebuilt it from scratch after a single giant workflow proved too rigid. The new architecture has three stages: Planning (grilling a rough idea with an LLM to surface missing requirements and edge cases, producing a spec, ADRs, glossary, and task list), Implementation (a Software Engineer role writing code and a Tester role trying to break it, iterating), and Review (run against the PR diff rather than the whole repo, with a merged product-architect role). Failures automatically create new implementation tasks, and they say it produces some of the cleanest AI-generated code they've worked with.
@Zephyr_hg [Claude Code]
https://x.com/Zephyr_hg/status/2080463978250351077
Recounts a manager at Anthropic who had not written a line of code in 15 years and now codes using Claude Code, with no bootcamp or retraining. She points the tool at the work and directs it, so one person now covers what used to need a fresh hire. The post frames the new skill as learning to run the tool rather than knowing syntax.
@mizorewww [Claude Code]
https://x.com/mizorewww/status/2080588666741924303
The user runs a workflow of writing a spec and having the agent complete it, with hundreds of specs used to govern a whole codebase, and reports Kimi (specifically Kimi inside Claude Code) outperforms GPT-5.6-sol-ultra on this while saving tokens. They note GPT's biggest problem is over-testing. The downsides for Kimi are that it's quite slow and frequently hits 429 errors, and they dislike the standalone Kimi Code software and client.
@AjeyGore [OpenClaw]
https://x.com/AjeyGore/status/2080444754303799646
The user reports that OpenClaw and Hermes have improved over time but are still suited to curious/technical people, with non-techies still needing time to set them up. They say they've solved the hosting problem and are now working on the configuration problem, having already automated a lot, but note a remaining gap and ask for suggestions. It's a firsthand account of building out and simplifying an OpenClaw/Hermes setup.
@kusodasabird [Claude Code]
https://x.com/kusodasabird/status/2080496345744584816
On day 24 of a new YouTube channel, the user shares how they produce most videos with Claude Code (nicknamed Kuroko). The division of labor: Claude Code handles planning, script, material collection, and editing, while the human makes final decisions, does final script checks, selects images, and posts. They stress they first learned YouTube's winning formats and gave precise instructions, since fully offloading to the AI causes it to go off the rails.
@mnmn94253156337 [Claude Code]
https://x.com/mnmn94253156337/status/2080462046555259208
The user found and used an open-source Skill called video-shotcraft for making product-intro videos, callable from Claude Code and Codex. It packages common product-promo shots into reusable shot recipe cards (106 cards, 161 motion previews) with a bundled Remotion video template, and can break a product page into shots, camera moves, subtitles, motion, and CTA. They say it helps skip the painful blank-timeline stage, and the intro video attached used the skill, which they found looked impressive on first use.
@precisox [Claude Code]
https://x.com/precisox/status/2080645252109459937
Summarizes a 40-minute video by Thariq Shihipar, a Claude Code engineer at Anthropic, on how they build agents internally, laying out the stack of agent, harness, loops, and memory. Concrete practices include using /goal and /loop to run agents for hours or days without collapsing, separating creator and verifier agents so the model doesn't fool itself, reducing Claude Code's system prompt by 80% because smarter models need principles over rigid instructions, and using HTML artifacts with Claude Tag in Slack so multiple agents work in parallel. The takeaway is that the real skill is designing loops, harness, and persistent memory rather than writing prompts.
@vikrantnyc [OpenClaw]
https://x.com/vikrantnyc/status/2080664522797318536
The user reports their OpenClaw setup is humming along and making life easier by autonomously sending responses across channels including RadarChat, Slack, and X. They jokingly cite the Cakewallet team, who dislike the autogenerated responses it sends them. It's a lightweight but concrete account of OpenClaw handling multi-channel messaging automation in daily use.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2080691991327998230
The user's company booked 100+ demos the day they opened their GTM API beta, and they share the playbook of making Claude Code run go-to-market end to end. The challenge was juggling API keys across 8-12 tools plus knowing which database, enrichment waterfall, and buying signals to use, knowledge packaged into one API. Inside Claude Code you can say 'Build a list of Software CEOs in London, 25-100 employees, pick who's in buying mode, draft 3 angles and upload to my sequencer,' and it builds the list, finds emails/phones, uncovers intent via 2-3 APIs, drafts angles with a cold-email skill, and uploads to the sequencer, flagging cost trade-offs.
@mikeldking [Claude Code]
https://x.com/mikeldking/status/2080564123071639604
The user started tracing their Claude Code sessions to test Arize Phoenix's revamped Code Mode MCP and forgot to turn it off, ending up with about a day's worth of Claude Code traces on their local machine. They found it fascinating to see how Claude Code works under the hood. The cost-tracking made them realize how quickly the dollars add up, averaging about $2 per turn.
@JulianGoldieSEO [Claude Code]
https://x.com/JulianGoldieSEO/status/2080596268548595884
Recounts Stewart Vickers, who spent 3 months automating his entire agency with AI agents and then ripped most of it out. Concrete findings shared: his $500 VA outperformed Claude Code at negotiating backlinks, an AI avatar experiment got the same views but killed conversions, and 80-90% agency automation sounds great until it blows up. The episode also covers the one thing clients still pay for that AI can't replicate.
@israfill [Claude Code]
https://x.com/israfill/status/2080706146986570110
After Opus 5 dropped, the user is switching it to default in Claude Code, having used Fable 5 for weeks, citing agentic terminal coding of 43.3 vs Fable's 33.7 and long-horizon big-repo work as where they feel the difference. They share their concrete flow: /model claude-opus-5, keep high effort for architecture and multi-file agents, drop effort for a dumb rename, use fast mode only when late and the loop is simple, and run /claude-api migrate if old strings hang around. They keep Claude for the main agent loop while running Kimi on the side.
@Mayaikos [Claude Code]
https://x.com/Mayaikos/status/2080675152568750331
Describes an AI creator who ditched Claude Code and OpenAI o1 for the Hermes Agent and now runs everything from one desktop app. Install took minutes with built-in skills for computer use, cron jobs, image generation, and task delegation enabled instantly, plus persistent memory that patches its own tools as she works. Telegram chats, artifacts, and Kanban boards live in one place, and she dictates carousel posts from the street and gets polished assets back in minutes, with a dashboard showing every session, cron, and blocker.
@kusodasabird [Claude Code]
https://x.com/kusodasabird/status/2080627554608873913
A Japanese YouTuber shows a playback-count graph for a new channel to argue Claude Code has real effect. He contrasts the pre-adoption period, which was fully manual, with the post-adoption period where the work is almost entirely handled by Claude Code. The concrete claim is a visible jump in video views attributed to automating the channel workflow.
@Tebasaki_lab [Claude Code]
https://x.com/Tebasaki_lab/status/2080667036997575163
A user reports raising a particular AI model's accuracy from 55% to 99.9% and being satisfied with the result. The work was expected to take a week but was largely handled by Claude Code's /goal feature after the user taught it their own know-how. This is a concrete accuracy improvement achieved with a specific Claude Code capability.
@Flandermaxx [OpenClaw]
https://x.com/Flandermaxx/status/2080716643211317745
Describes a 32-year-old Copenhagen developer who wired four Mac Mini M4 units plus a GMKtec box running qwen 2.5 VL 32B into a local OpenClaw agent farm connected via Tailscale. He runs 34 client workflows unattended, each agent reading a client's gmail, writing invoices in Dinero, and posting to LinkedIn, charging 34 Nordic SMB owners DKK 3,900/month for roughly $19,000 MRR. The setup swaps the Anthropic endpoint for a local vLLM serving qwen, reportedly trading about 15% quality for near-zero per-token cost.
@yumekawii_s2 [Claude Code]
https://x.com/yumekawii_s2/status/2080530239311192145
A 24-year-old built a 'probability farm' using an agent program he made with Claude Code over a few weekends, running on a Mac Mini alongside six phones logged into different Polymarket accounts. The agent watches 90+ markets across Hyperliquid contracts and Polymarket events, finds probability gaps, sizes positions, and auto-places orders through accounts that still have margin, averaging 11-minute holds at nearly 70% win rate around the clock. The post claims withdrawals of about $510k over two months, with the system closing 1,380 trades in one night while he slept.
@Piefena [Claude Code]
https://x.com/Piefena/status/2080674239649366352
Describes a webcam-based interface where the camera watches a user's hand, recognizes gestures, and instantly triggers the right meme on screen with no buttons or clicking. The poster notes that building this used to mean days of fighting OpenCV, hand-tracking models, and gesture thresholds, but now you can explain the idea to Claude Code (or GPT) and iteratively correct it until it works. It frames Claude Code as turning cameras, hands, and movement into buildable interfaces.
@mikenevermiss [Claude Code]
https://x.com/mikenevermiss/status/2080642867727323274
A user reports someone used Claude Code to build and automate a quant trading strategy based on the GARCH model and open-sourced the entire GitHub repo. The release includes a Pine Script indicator that loads the strategy in real time on TradingView, full installation instructions, and a reusable Claude Skill. It is presented as a concrete, fully open-source AI trading resource.
@silentguyy66 [Claude Code]
https://x.com/silentguyy66/status/2080583470129017009
A user recounts spending three months treating Claude Code like a chatbot with one giant prompt (fix this, refactor that, add tests, deploy), costing about 12k tokens, $0.12, and 90 seconds per call with much of the context re-explaining instructions. He then split the work into six graph nodes (extract, transform, analyze, validate, route, report), each roughly 200 tokens, $0.002, and 3 seconds on a cheaper model, chaining outputs. The result was a full pipeline at $0.012 in 18 seconds, about 10x cheaper and 5x faster, and fully traceable so failures are isolated to a node.
@Hadley [Claude Code]
https://x.com/Hadley/status/2080631522818716088
A user describes running their entire stack off Zo Computer, a cloud Linux box, using their own Claude Code and Codex subscriptions. It runs the native Zo harness plus Hermes and is wired into their firm's MCPs including Superhuman, Affinity, Slack, and Google Workspace. They also build and host their own apps on it, such as a Fable-powered email triage client.
@kzkhykw [Claude Code]
https://x.com/kzkhykw/status/2080666752456052963
A user reports that since last week they have been running a setup where Claude Code automatically turns their uploaded YouTube videos into Short-form videos. The system produces about ten Shorts automatically and even handles uploading them. This is a concrete, running content-automation workflow.
@coldemailchris [Claude Code]
https://x.com/coldemailchris/status/2080699610733883521
A user outlines a product-market-fit testing pipeline that AI now makes possible in a single morning. They build a landing page in Lovable, iterate outbound messaging with Claude Code, have Hermes pull a list of target accounts, enrich contact data with Deepline inside Codex, push campaign assets to an email sequencer via API, and fire campaigns, all in about 2-4 hours. They note this previously would have taken at least two weeks.
@PrakashS720 [Claude Code]
https://x.com/PrakashS720/status/2080589799258505631
Describes Hyperresearch, an open-source research tool that runs entirely on the user's own machine via Claude Code. From one prompt it pulls 250+ sources in parallel, writes multiple independent research drafts, uses AI critics to attack its own conclusions, verifies every citation against the original source, removes duplicate coverage, and builds a local research vault that improves with use. A Dissertation Mode is cited producing 25,000-80,000 words across up to 10 chapters from 300-450 sources over 4-8 hours.
@Dmytroo_eth [Claude Code]
https://x.com/Dmytroo_eth/status/2080691694035443742
A user profiles a solo developer who runs four tiny SaaS tools (a changelog generator, a Slack support-ticket digest, a freelancer invoice reminder, and a status-page generator) earning about $9,000/month. Rather than building one big startup, he uses Claude Code to scaffold auth, billing, database tables, CRUD screens, tests, and the boilerplate shell around decisions he already made. When a bug report comes in it becomes a prompt: Claude reproduces it, writes a failing test, fixes the issue, and opens a diff he reviews.
@dapi_kele [OpenClaw]
https://x.com/dapi_kele/status/2080648458420281465
A user reports testing the OpenClaw bot inside WeChat and hitting apparent content filtering. The same topics queried on Threads returned normal replies, but the WeChat bot responded with 'please try again later.' They ask whether even AI-agent-generated replies must pass through WeChat's pre-emptive risk-control filtering, documenting a real platform-specific limitation of the OpenClaw bot.
@mikeldking [Claude Code]
https://x.com/mikeldking/status/2080707300009120065
A user describes tracing and evaluating Claude Code locally using Arize Phoenix, storing everything in SQLite. Via the built-in MCP they automatically detect when they get frustrated during a session and have Claude suggest skills based on previous sessions. They note this is forming into a self-improving loop, all running locally.
@antisadh [Claude Code]
https://x.com/antisadh/status/2080710106279276645
Describes a 512GB EPYC build that loads Kimi K3 at the full 1-million-token context entirely into RAM without swap or quantization, alongside DeepSeek 671B at native precision. The rig uses an AMD EPYC 7v13, eight 64GB DDR4 sticks across all memory channels, a Supermicro H12SSL-NT board, and a Noctua cooler. When a refactor lands in his tracker he pipes the whole monorepo through Kimi K3 as a dependency graph to spot blast radius across services and draft the PR, replacing $200 ChatGPT Pro and $200 Claude Code Max at zero recurring cost.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2080665578667123071
Shares a detailed, reusable prompt that audits all Codex and Claude Code skills, plugins, MCP servers, hooks, and persistent instructions on a Mac in strict read-only mode. It inventories specific config paths, follows symlinks, and flags duplicates, conflicts, outdated models/APIs, credentials stored in config (recording only field name and location), and unsafe permission combinations, then labels each item (KEEP, MERGE, MOVE_TO_PROJECT, ARCHIVE, DELETE, etc.). It outputs an offline report.html, a plan.json, and a one-page summary.md without modifying anything.
@harrisrafto [Claude Code]
https://x.com/harrisrafto/status/2080646500212715835
A user reports building an invoicing app entirely inside a Zerops remote workspace, with Claude Code handling the infrastructure. It was released as a Laravel News video. The poster calls it the most fun they have had with a deploy pipeline in a while.
@_Chain_Duck_ [Claude Code]
https://x.com/_Chain_Duck_/status/2080689851087634666
A user reports that a trader used Claude Code to build a Polymarket arbitrage bot from scratch without writing a single line of code, with the human only writing the strategy. The bot monitors real-time odds across correlated markets, flags pricing gaps before they close, and fires trades automatically. The post cites 167 trade executions per hour and $197,000 in profit.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2080624836750070025
Describes Blueprint, positioned as 'Claude Code for electronics,' which turns a plain-English description into a buildable hardware design. After asking a few architecture questions it returns an AI render of the finished device, a wiring diagram, a full parts list with purchase links, a 3D view, and step-by-step assembly instructions. The poster notes people have used it for tesla coils, e-ink displays, and camera glasses, and that it was built by 3E8 Robotics with a free quota.
@irongiantXBT [Claude Code]
https://x.com/irongiantXBT/status/2080491409224962325
A user relays that Spotify chief architect Niklas Gustavsson explained how the company scaled agent success from 20% to 80% using Claude Code and now merges 73% of its code with zero human review. The described approach stops swapping models and instead builds iterative feedback loops around existing agents, removes human bottlenecks so agents iterate and merge across 20 million lines of code, and relies on autonomous retry mechanisms to catch errors before a developer sees them.
@aibullss [Claude Code]
https://x.com/aibullss/status/2080732687531376673
A user built a trading risk 'gate' installed into Claude Code in six steps. It includes a rules file listing what disqualifies a trade, a thesis template where invalidation is a required field, a script that reads both and exits with an error if a trade breaks the rules, and a CLAUDE.md that loads the gate at session startup before any entry can be discussed. The validation step plants three bad trades and confirms all three are blocked, rather than merely proving a price feed is connected.
@Bsoutherland23 [Claude Code]
https://x.com/Bsoutherland23/status/2080718599178609050
A user shares 'handoff-skill,' the eighth of 30 skills, addressing how AI degrades over long conversations as context windows get messy. Saying 'handoff' makes Claude write a paste-ready document capturing current state, key decisions, open items, and a first message for a fresh chat, so the next conversation resumes with zero re-explaining. They rolled it out firmwide and report it is one of their top-used skills.
@jboogx_creative [Claude Code]
https://x.com/jboogx_creative/status/2080604588793831879
A creator shares a piece called 'Bartender Mommy' with images made in MidJourney 8.1 and animated in Seedance 2.0. The prompting was done using 'Night Shift,' a custom Claude Code system he built for creators. This is a concrete creative-media workflow driven by a personal Claude Code setup.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2080632999012405676
A user offers a game-development practice for Claude Code Desktop: ask Claude to create a dedicated scene just for verifying character movements, so it can simulate and verify each movement without interruptions before adding it to the corresponding scene. They recommend building a SKILL.md so Claude runs this simulation every time it creates a new scene.
@antisadh [Claude Code]
https://x.com/antisadh/status/2080626405725835544
Describes a local rig with a $700 used RTX 3090 (24GB VRAM), a Supermicro board, and 128GB DDR4 ECC running Kimi K3 via Ollama at the full 1-million-token context for about $2/month in electricity. The user parses his monorepo into a Neo4j graph where every service, API, database, and event is a node, and Kimi K3 reads the graph directly to spot the dependency subgraph for a bug ticket and draft a PR. It replaced his $459/month AI stack including Claude Code Max for refactor and architecture queries.
@easys_arq [Claude Code]
https://x.com/easys_arq/status/2080655719636910378
Gives a four-step recipe for AI creators who use Higgsfield: get a $20 Claude subscription, install Claude Code in the terminal, obtain a fal API key and give it to Claude, then ask Claude Code to build their own Higgsfield alternative. It frames Claude Code as a way to replace a paid creative tool by building a custom version.
@sgnoogle [Claude Code]
https://x.com/sgnoogle/status/2080653172268310679
A user reports that in two hours with Claude Code they completely revamped a widget so that it now looks and behaves like a native app directly on the home screen. This is a concrete, quick result from a single Claude Code session.
@v99427 [Claude Code]
https://x.com/v99427/status/2080621623141609803
A YouTube creator who previously struggled to write scripts, sometimes spending a whole day without finishing one while trying to differentiate from rivals, reports that Claude Code has been a breakthrough. It learned the format and planning that drives growth and now mass-produces scripts and even checks them, leaving very little manual effort. Presented as revolutionary for people who struggle with scriptwriting.
@aimi_sh [Claude Code]
https://x.com/aimi_sh/status/2080720281933394351
A user tested Opus 5.0 against Fable 5 Max and found Opus 5.0 clearly better on 3D work, praising the level of detail and texture quality. They note Opus 5.0 is available via the claude-opus-5 model string on the API or in Claude Code, while Fable nailed its attempt in one try, suggesting Opus still has room to grow on that dimension. Grounded in hands-on 3D testing.
🗣 User Voice
User Voice
Model switching is the new bill shock. @omarsar0 points out that changing models mid-session resets your prompt cache, and every reset quietly inflates the agent bill — people want cache-aware routing, not just a model picker.
Opus 5 needs babysitting out of the box. @JJEnglert warns it's chatty, stops short on long tasks, and sounds confident while being wrong, so it takes real CLAUDE.md tuning before it earns its keep.
Skill sprawl is real. @Voxyz_ai calls out the reflex of installing every Skill you come across — you end up with duplicate SKILL.md files and plugins you can't track, some of which just repeat what the stronger model already does natively.
Memory keeps getting rebuilt from zero. @lucian__03's 65-line CLAUDE.md thread is popular precisely because Claude Code forgets the repo and the user every session, so people bolt on their own memory layer.
Coding from a phone is still clunky. @jackfriks locked Claude Code onto a VPS to work from his phone, but the terminal won't let you copy-paste, images don't paste, and the view keeps jumping as new output streams in.
OpenClaw and Hermes still aren't for everyone. @AjeyGore notes they've improved a lot but remain built for curious tinkerers — non-techies still hit a wall setting them up and trusting what comes out.
Model switching is the new bill shock. @omarsar0 points out that changing models mid-session resets your prompt cache, and every reset quietly inflates the agent bill — people want cache-aware routing, not just a model picker.
Opus 5 needs babysitting out of the box. @JJEnglert warns it's chatty, stops short on long tasks, and sounds confident while being wrong, so it takes real CLAUDE.md tuning before it earns its keep.
Skill sprawl is real. @Voxyz_ai calls out the reflex of installing every Skill you come across — you end up with duplicate SKILL.md files and plugins you can't track, some of which just repeat what the stronger model already does natively.
Memory keeps getting rebuilt from zero. @lucian__03's 65-line CLAUDE.md thread is popular precisely because Claude Code forgets the repo and the user every session, so people bolt on their own memory layer.
Coding from a phone is still clunky. @jackfriks locked Claude Code onto a VPS to work from his phone, but the terminal won't let you copy-paste, images don't paste, and the view keeps jumping as new output streams in.
OpenClaw and Hermes still aren't for everyone. @AjeyGore notes they've improved a lot but remain built for curious tinkerers — non-techies still hit a wall setting them up and trusting what comes out.
📡 Eco Products Radar
Eco Products Radar
Codex — the most-cited co-pilot alongside Claude Code, usually run in tandem or compared head to head.
Opus 5 — the week's new default model; praised on capability, criticized for verbosity out of the box.
Fable 5 — the cheaper alternative people benchmark against Opus 5 on cost and quality.
Cursor — still the reference IDE in tool comparisons.
Hermes / Nous Research — the harness people name when they walk away from OpenClaw.
Kimi K3 — the cheap open model of choice for running locally inside Claude Code.
Obsidian — the de facto memory and second-brain vault bolted onto Claude Code.
MCP — the connective tissue for tool access, showing up across build workflows.
Remotion — the go-to for AI-driven programmatic video.
Higgsfield — recurring in creative and media-generation workflows.
Codex — the most-cited co-pilot alongside Claude Code, usually run in tandem or compared head to head.
Opus 5 — the week's new default model; praised on capability, criticized for verbosity out of the box.
Fable 5 — the cheaper alternative people benchmark against Opus 5 on cost and quality.
Cursor — still the reference IDE in tool comparisons.
Hermes / Nous Research — the harness people name when they walk away from OpenClaw.
Kimi K3 — the cheap open model of choice for running locally inside Claude Code.
Obsidian — the de facto memory and second-brain vault bolted onto Claude Code.
MCP — the connective tissue for tool access, showing up across build workflows.
Remotion — the go-to for AI-driven programmatic video.
Higgsfield — recurring in creative and media-generation workflows.
Comments