Super User Daily: July 20, 2026
The second brain has become the dominant non-coding use case: today's haul is packed with Karpathy-derived setups — one-file stream.md capture systems, Obsidian vaults read natively from the terminal, and one user who turned 4,700 notes into an actual 26-neuron neural network trained on his own click behavior. The money stories got bigger too: a 22-year-old shipped a GTA-flavored wedding-logistics game for ~$100K in sales, a Beijing hutong simulator cleared $151K against $240 in costs, and a disposable camera app hit $20K/month after validating with 250 cold DMs and zero code. Enterprise adoption is quietly compounding — a Japanese tax accountant office moved its whole staff onto the Claude AI Team plan, and Sierra's internal agent has run 75,000 sessions with 70% of PRs opened through it. The dark side kept pace: SlowMist showed that opening a poisoned project can equal granting shell access, researchers demonstrated persistent-memory prompt injection, and OpenClaw's own repo now spends 62% of its PRs on fixes versus 8.5% on features.
@cryptowluha [Claude Code]
https://x.com/cryptowluha/status/2078612256775139406
A 22-year-old Indian developer built "GTA India" on Unity using GTA 5 as the base, with Claude Code and Python doing the heavy lifting on the rewrite. He spent $200 on AI and sold 20,326 copies at $5 each — roughly $100K. Missions are wedding logistics: hijack the limo convoy, fend off the rival family, get the bride to the temple on time. The physics were reworked so piling people onto a bus roof shifts handling, balance, and suspension with the weight.
@AnuragTyagi [Claude Code]
https://x.com/AnuragTyagi/status/2078311759035691290
Solo builder mapped the entire pipeline behind how an app knows a show landed on Netflix before JustWatch does — scrapers, crons, a provider feed — and made it clickable. Built with Claude Code: one person, one AI, covering 8 platforms and 34 markets.
@drillinnair [Claude Code]
https://x.com/drillinnair/status/2078590686853783731
Job-hunting workflow: a master resume in LaTeX made with Claude Pro, then for every application the job description goes into Claude Code and a custom /applykarbsdk command generates a tailored resume and cover letter. That throughput let the user apply to 50-60 companies a day, which turned into 5 interviews and offers from 2 of them.
@CoderNolimit [Claude Code]
https://x.com/CoderNolimit/status/2078474085613412518
French user gave Claude Code a single prompt asking it to generate a three.js experience visualizing a 3D airplane model that you can walk inside. Posted the literal result as a demo — a concrete one-prompt-to-interactive-3D output.
@usutaku_channel [Claude Code]
https://x.com/usutaku_channel/status/2078469182224797786
Japanese user tested a website-styling skill: the result didn't fully clone Apple's site but picked up Apple-like touches such as the typography. Reports Sonnet 5 was underwhelming and the skill only showed its real capability on Fable 5. Notes beginners can just hand Claude Code a GitHub URL and say "make this usable as a Skill."
@bojie_li [Claude Code]
https://x.com/bojie_li/status/2078368794129662052
Author of a book repo ran Claude Code + Kimi K3 with Opus 4.8 as reviewer for a full day: completed 30+ missing experiment code files, fixed missing-dependency issues, and updated most experiments to newer models — noting some experiments must keep weaker models since SOTA models no longer need the harness to patch their flaws. The repo doubled again in a day to 2.6k stars and v1.1 shipped.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2078376607203696786
Connected GA4 and Google Search Console to Claude Code via the official Google MCP and now runs a daily loop: fetch data, find improvements, fix and publish, monitor rankings. What was a 30-minute CSV chore done less than weekly is now a 3-minute "analyze the landing page" prompt done daily — one fix a day compounds to 30 improvements a month. Also documents the 403 permission trap that "99% hit" during setup.
@kamikudaku_wani [Claude Code]
https://x.com/kamikudaku_wani/status/2078382965583999415
Obsidian + Claude Code organization method adapted from a 1,500-bookmark English post: don't design your folder structure yourself — ask the AI to interview you with multiple-choice questions; turn recurring chores into commands like /standup and /tldr; convert PDFs to distilled notes before saving so noise isn't stored as truth; and have the AI draw mind maps inside the vault. The author's own vault grew to 3,196 files in 2 months with the same "AI does the sorting" approach.
@Axel_bitblaze69 [Claude Code]
https://x.com/Axel_bitblaze69/status/2078551452201754980
Describes video-use, a free open-source Claude Code skill (17,000 GitHub stars) that turns editing into a conversation: drop raw footage in a folder, describe what you want, get final.mp4 back. It transcribes every word first and edits off the transcript — cutting umms and dead space, color grading per segment including custom ffmpeg chains, building the cut with parallel editor subagents, then rendering a preview, grading its own work and iterating. It shows its plan in plain English and waits for approval before cutting.
@simonw [Claude Code]
https://x.com/simonw/status/2078343997119172705
simonw publicly asks Anthropic for an automated test suite ensuring Claude Code on the web never blocks cloning or interacting with public repos from existing sessions. He notes "Clone repo X from GitHub to /tmp for reference" is one of his most-used prompting patterns and that this is not the first time Claude Code on the web has regressed this functionality.
@omarsar0 [Claude Code]
https://x.com/omarsar0/status/2078573267015885136
Shared a self-built skill that tracks high-signal X accounts for AI news using X MCP tools: set up the MCP, install the skill, then prompt the agent to build a self-contained local feed.html artifact and validate it. He runs it on a schedule every 4 hours and says it works in Codex, Claude, Hermes, or OpenClaw; source handles must be curated manually and he shares his own list.
@kyroxxxq [OpenClaw]
https://x.com/kyroxxxq/status/2078400541818765775
A founder gave OpenClaw terminal access to act as a 24/7 librarian over his local knowledge base, combining Tiago Forte's PARA method with Karpathy's LLM-wiki concept and removing the manual work. The agent sweeps his files daily, categorizing thoughts, articles, and finished projects without manual tagging, and distills them into an interconnected knowledge graph so context is pre-built when he sits down to work. Published projects get ingested back in, so the graph gets denser and retrieval improves with every project.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2078527096851140670
Claude Code turned a 4,700-note, 6-year, 2.1M-word Obsidian vault into an actual neural network in one evening via 3 prompts: build embeddings and map the link graph; write a 17-input, 26-neuron ReLU network from scratch in ~900 lines of browser JavaScript with no PyTorch; train it on the owner's own behavior — which notes he opens, links he clicks, ideas he drops. The net now pre-loads 5 notes each morning and flagged a link between a 2021 attention note and a 2026 prediction-markets note that became a 214,000-view post.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2078464837953425858
Summary of an Anthropic-sponsored video condensing a developer's 800+ hours of Claude Code operations into 8 minutes: save every instruction into CLAUDE.md and reusable commands so nothing is explained twice, run subagents in parallel for code cleanup, docs, and UI review, and connect Stripe, Supabase, Vercel, and Playwright directly to eliminate documentation copy-paste. The entire production setup is open-sourced; the theme is applying DRY to Claude Code operations so one person runs a multi-person development pipeline.
@sgnoogle [Claude Code]
https://x.com/sgnoogle/status/2078576864268873914
Developer of blockit (500k downloads) rebuilt the app for its 2.0 release in 15 days using Claude Code, talking to it almost 10 hours a day. He announces the biggest update in the app's history and walks through the changes feature by feature in a thread.
@ryanflorence [Claude Code]
https://x.com/ryanflorence/status/2078472215604129920
A small-business owner — not a programmer — replaced his company's CRM with one he built with Claude Code a few months ago. He has never seen the code and doesn't know what stack it uses; it runs on a Mac mini in the office and he says it's way better than what they had.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2078320470852833707
Jyothi built an AI chief-of-staff entirely inside Claude Code: fed it Granola meeting transcripts, strategy docs, org charts, and emails, and had Claude write the architecture spec as a markdown file first, then build from that spec. Every person gets a profile — communication style, motivators, what works — plus a relationship score (ally/friendly/neutral/cautious/friction) that updates after every interaction. The system proactively told its owner to make a specific coworker an ally and warned her to loop in certain people before sending a sensitive message.
@YasLovesTech [Claude Code]
https://x.com/YasLovesTech/status/2078415502242976080
Vibe-coded a finance data pipeline with Claude Code that collects EPS growth forecasts for all S&P 500 companies every day and stores them in a database, producing a weighted-average S&P 500 EPS table. The daily cadence is deliberate — not for the snapshot, but to capture how analyst forecasts change over time.
@Nyra_nx [Claude Code]
https://x.com/Nyra_nx/status/2078579412015190418
An entire company's second brain run from 5 folders of plain markdown that Claude Code reads from the terminal: North Star (one weekly goal), Daily Log (every meeting and name), Clients (one note per account synced to the CRM), Projects (one file each), and Claude Skills (the playbooks the agent executes). The agent works from the same files as the human — same context, zero copy-paste — and a second vault holds the assistant's entire memory as a readable folder. No Notion, no database, no subscriptions.
@Granite0x [Claude Code]
https://x.com/Granite0x/status/2078473107354857968
Ryan Carson (Treehouse founder) open-sourced "ralph," a loop-engineering tool: write your task list once and it runs Claude Code in a loop until every item is done — no prompting or babysitting, you come back to finished work. MIT-licensed, cited at 21,102 GitHub stars.
@tomlife43 [Claude Code]
https://x.com/tomlife43/status/2078318179567432123
Solo dev built "Senbei" with Claude Code — a BeReal-style camera app that records both front and back cameras simultaneously. Unlike BeReal's 15-second limit it records without a cap, and each side can be saved individually if you don't like the combined shot.
@bonsaixbt [Claude Code]
https://x.com/bonsaixbt/status/2078448921869754681
Video roundup of three GitHub repos fixing the two common Claude Code + Obsidian problems — no memory between chats, and generated files that break vault formatting. Obsidian Mind retrieves only task-relevant notes instead of loading the whole vault into context, Obsidian Skills (created by Obsidian's CEO) teaches Claude correct file formats and note structure, and Obsidian Second Brain lets Claude save discoveries, update and link notes, and keep the vault organized in the background. All free and open source.
@vladdubchak_x [Claude Code]
https://x.com/vladdubchak_x/status/2078333187856695779
A team spent days trying to reverse-engineer a video style before finding it already solved — in Chinese, for a completely different pipeline. They translated it, rebuilt it for Claude Code + Maxfusion, and ran it on 90 clips with only 2 needing a redo. They wrote up the full step-by-step including quality gates.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2078328914477473891
Setup where Claude Code writes every session's content into Obsidian, NotebookLM analyzes the accumulated data, and Claude reads the whole vault before answering "how's the company doing today." Example daily answer: 4 new contracts and $11,500 new MRR; 23 leads found, 18 cold emails drafted, 1 meeting booked; $487 ad spend at 3.8x ROAS with the best copy at 8.2% CTR and the worst ad already paused. The vault grows daily and answer precision improves weekly; the deeper design is multiple AIs sharing one memory.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2078503194951372897
skills.sh from Vercel Labs: a package manager plus directory for 930,000+ agent skills. "npx skills add owner/repo" pulls any skill from GitHub, detects which agents you have installed, and drops it into the right folder — it works with Claude Code, Cursor, Codex, Copilot, and 60+ other agents, and skills only load into context when the task calls for them. Managed like any package manager with find/list/update/remove subcommands.
@dotey [Claude Code]
https://x.com/dotey/status/2078481595674112362
Usage tip from a Chinese heavy user: prefer the Codex/Claude Code desktop versions and use Computer Use for automated testing. He reports the desktop apps now proactively invoke Computer Use to test work before delivery and auto-fix problems found, which he says is far more efficient than testing by hand.
@kasong2048 [OpenClaw]
https://x.com/kasong2048/status/2078505722783187248
Data snapshot on OpenClaw — the best-known fully AI-led development project — as of July 18: fix PRs are 62% of all PRs, feature-iteration PRs only 8.5%, and refactoring PRs 26%. The author's read: a Ship-of-Theseus codebase where heavy patching and refactoring to suppress rot is crowding out the ability to ship new features.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2078301247065460816
A fan rebuilt the PS2 downhill mountain-biking game Downhill Domination — which never got a sequel — as a free browser game with Claude Code: the whole thing is one single HTML file with no separate image, model, or sound assets, the track, bike, and physics all generated in code. Built with Fable, Opus, vanilla JS, and three.js over multiple iterating sessions; someone sent a PR for controller support, he merged it, and the community helped squash bugs.
@sudoingX [OpenClaw]
https://x.com/sudoingX/status/2078471458171539573
Hands-on onboarding test of Hermes Agent vs OpenClaw against a 3.9GB local model served on localhost: Hermes showed a provider menu with local servers, verified 127.0.0.1:8899, auto-detected the model plus its context length, and started firing real tool calls with no key. OpenClaw had no menu, hunted for a Codex login or OpenAI key, printed "no models available" three times, defaulted to an unreachable cloud model, and dead-ended — it only runs local if you hand-wire the endpoint. The author discloses he contributes to Hermes, including the auto-detect code that won the comparison.
@wayen_ai [Claude Code]
https://x.com/wayen_ai/status/2078314190901821907
fireworks-tech-graph, an open-source skill for Claude Code and Codex that generates professional architecture diagrams from natural-language descriptions in Chinese or English, output as SVG + PNG. It offers 12 visual styles (flat, dark terminal, blueprint, Notion-minimal, glassmorphism and more), 14 UML diagram types, dedicated AI/agent modes for RAG and multi-agent architectures, and a generate → structural-validation → visual-check → auto-fix loop, installable via npx skills or git clone.
@polydao [Claude Code]
https://x.com/polydao/status/2078353069025485293
After months of one-message-at-a-time prompting, this user adopted the Claude Code team's loops guide and now runs 4 loop types: turn-based (a SKILL.md defines what good looks like so Claude verifies its own work), goal-based (/goal with an evaluator bouncing work back until "Lighthouse 90" clears), time-based (/loop checks his PR and fixes CI every 5 minutes; /schedule runs it in the cloud with the laptop shut), and proactive (a routine watches, triages, and fixes alone while a second agent reviews overnight). When something breaks he encodes the fix so the next run improves; he now opens Claude to see what got done, not to do the work.
@uche_ui [Claude Code]
https://x.com/uche_ui/status/2078483076515074321
Reply describing a concrete personal workflow: Claude Code runs locally connected to git so pushes to GitHub take one line, a folder holds the images and Claude does the rest through back-and-forth prompting, and the site gets built section by section rather than all at once.
@0xKenny1st [Claude Code]
https://x.com/0xKenny1st/status/2078574001354318267
Setup for design-system-driven output: find the UI UX Pro Max skill on GitHub, paste the URL into Claude Code and say "install this skill," then install the Magic MCP server the same way. Claude then pulls from a real design system instead of guessing — 50+ UI styles, 97 color palettes, 57 font pairings, built-in glassmorphism, gradients and animations. Claims under 5 minutes to set up and that one site built this way has already pulled in $7,000.
@SlowMist_Team [Claude Code]
https://x.com/SlowMist_Team/status/2078402837416231007
SlowMist audited Grok Build CLI — which inherits Claude Code CLI's permission configuration model — and within 24 hours found two attack chains leading to arbitrary code execution without explicit user approval: cargo check misclassified as a safe command combined with malicious AGENTS.md instructions, settings files overriding permission checks, and project-level MCP configs adding attack surface. Their conclusion: when AI coding agents trust project-level files too much, opening a project can become equivalent to granting shell access.
@_andresjasso [Claude Code]
https://x.com/_andresjasso/status/2078322115225538871
A designer's end-to-end stack with an explicit role split: Claude Code desktop app for most creative design, UI, animation, motion, and micro-interactions; Codex for app building, integrations, structure, and backend; Cursor's design mode for targeted UI components and landing-page graphics; Flora using GPT Image 2 for on-brand images as a starting point; and Figma now under 10-20% of the process, never for UI components, mostly for presenting work made in the other tools.
@Nekt_0 [Claude Code]
https://x.com/Nekt_0/status/2078560942745374916
A creator built his entire production chain inside Claude Code with 7 skills: Deep Research finds the angle and facts, Messaging Hooks shapes the opening, Voice DNA keeps the writing recognizable instead of generic AI copy, Remotion turns the script into edited video, ElevenLabs handles narration, and a podcast skill converts the same research into a two-person audio format. One source document yields several finished assets, with context surviving from research to final render instead of being rebuilt across five separate tools.
@thekuchh [Claude Code]
https://x.com/thekuchh/status/2078375718254473489
Solo build timeline: day 1 was a landing page and waitlist before the app existed; day 3 brought 71 signups from one TikTok post; by week 3 the whole product shipped with AI resume scoring, Stripe, and ATS-ready PDFs, built solo on Claude Code. Validation came first — a competitor was already charging $44/year for a worse version, so the builder validated and built an audience before writing any code.
@dair_ai [Claude Code]
https://x.com/dair_ai/status/2078555662133665941
New research tests prompt injection against persistent memory in Claude Code and OpenAI Codex across models. Tricking an agent into overwriting its own memory from untrusted external content is genuinely hard, but payloads already sitting in memory files attack current and future sessions successfully. Opus 4.7 and GPT-5.5 showed 0% attack success on credential exfiltration, while unauthorized tool use landed with high success across almost every model tested; one planted rule quietly pins a pip install to a PyYAML version with known vulnerabilities.
@milindlabs [Claude Code]
https://x.com/milindlabs/status/2078508667180306561
Developer turned tldraw into "paper that answers back": you can type on the canvas and it responds, ask it to make diagrams, or draw a diagram and ask it to update it. It needs no API key and connects directly with your Claude Code login. He open-sourced it and suggests it could matter for teaching students.
@_stakaya [Claude Code]
https://x.com/_stakaya/status/2078354319569744287
Japanese summary of Sierra's published lessons from internal AI agent adoption: role-specific agents per department failed because valuable work crosses departments, so everything was consolidated into a single agent, "Pinecone," with one Slack account and one URL. The bottleneck was business context, not model intelligence; using Claude Code and MCP they built a data-analysis agent spanning Slack, GitHub, ClickHouse, Salesforce, and PagerDuty that investigates customer issues in minutes, plus a custom MCP gateway inheriting per-employee permissions with audit logs. Since its first commit in March, Pinecone has run 75,000+ sessions, is used by 600+ people, and 70% of PRs are opened through it.
@fabriciocarraro [Claude Code]
https://x.com/fabriciocarraro/status/2078574831466078265
Portuguese-language benchmark: the author gave the same full-site redesign prompt to Fable 5 (Claude Code), GPT-5.6 Sol (Codex), and Kimi K3 (Kimi Code) at max effort on his personal site. Execution times were 35 minutes for Fable, 50 for K3, and 1h10 for Sol; all three independently converged on a warm-paper/ink-blue palette with editorial serif typography, and Fable and Kimi both named the concept "Editorial Poliglota." All removed heavy dependencies for lighter local CSS; Sol cut the deploy from ~382MB to 135MB via media optimization but wrote off-brand coach-style copy with layout errors, while Fable applied the shared palette best in his view.
@thekuchh [Claude Code]
https://x.com/thekuchh/status/2078412712082608321
A disposable camera app reached $20,000+/month in 83 days, validated without writing a single line of code: the builder's rule was that 10 real events had to commit with dates on a calendar or he would not build it. The funnel was 250 cold Instagram DMs, 15 replies, and 12 booked events — only then did he open Claude Code.
@doublenickk [OpenClaw]
https://x.com/doublenickk/status/2078621913048560110
An agency built a local OpenClaw knowledge node to run a client's 15,000-note Obsidian vault offline and kill a $1,720/month cloud bill. The pipeline: a TLS-secured SyncThing bridge ingests raw office files, local Gemma 4 purges duplicates and extracts entities, Qwen 3.6 cross-references notes into an interlinked Obsidian graph, and local models run HIPAA compliance checks with no internet connection. The client saves over $38,000 across three years while the agency collects a $5,000 deployment fee plus a $500 monthly maintenance retainer.
@Liquiddeny [Claude Code]
https://x.com/Liquiddeny/status/2078448038062780700
Applies Karpathy's one-giant-note second brain — new thoughts on top, old ones sink, rereading is the system — with Claude Code fixing the flaw that nobody rereads past 10,000 lines. Keep the note as plain Markdown in Obsidian, dump everything on top, never organize; point Claude Code at the folder and ask things like "what did I think about pricing in March" — it answers from your notes with quotes in ~15 seconds. Once a week one prompt closes the loop: read the last 7 days, surface 5 entries worth pulling back up, and flag contradictions with what was written a month ago.
@CharlieKimmel [OpenClaw]
https://x.com/CharlieKimmel/status/2078445408997765268
A recruiter ran 4 AI sourcing agents simultaneously on the same executive-search brief for a traveling superintendent role: OpenClaw with Opus 4.8 found 60 candidates, Hermes with Opus 4.8 found 56, Hermes with Grok 4.5 found 71, and OpenClaw with ChatGPT 5.6 found 45, for a combined unique pool of 193. Cross-validation became the shortlist: 2 candidates were surfaced by all 4 agents, 8 by three, and 17 by two — 27 cross-validated total. No single model found everything; each searched differently.
@aidoga_lab [Claude Code]
https://x.com/aidoga_lab/status/2078309485484073221
Japanese creator produced a 0.1-second flash-cut video with zero manual editing: he told Claude Code to write an editing script that alternates two AI-generated ocean clips at 0.1-second intervals. Not a single frame was cut by hand, and he shared the copy-paste prompt in the replies for others to try.
@jboogx_creative [Claude Code]
https://x.com/jboogx_creative/status/2078526673956589764
Creator shared "Dream Sketches Vol. 7," made on his Friday night live stream with his community: images generated in MidJourney 8.1, animated in Seedance 2.0, and prompted using Night Shift, his custom Claude Code system built for creators.
@jayclouse [OpenClaw]
https://x.com/jayclouse/status/2078545269621420422
About 7 months in, a creator lists his daily OpenClaw uses: a credit card optimizer skill, business finance projections, a "can we afford this" skill comparing investments against trailing 6-month revenue, personal finance analysis, a workout planner with a logging skill, analyzing member applications and churned-member exit surveys, deriving 6 member personas and building a skill to run synthetic focus groups against them, analyzing webinar chat transcripts, logging all call transcripts via Fathom, essay scoring on a co-created 5-point rubric, and vacation research. It is connected to team Slack so the whole team uses the growing business context, and he switches between Anthropic, OpenAI, Gemini, and MiniMax models with no lock-in — despite "the occasional accidental lobotomy it gives itself."
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2078525280910479491
Concrete workflow for a competitor ad-analysis agent: Apify scrapes the Facebook, LinkedIn, and Google ad libraries; a daily agent extracts net-new ads into a library; Swell AI transcribes UGC ads; a vision model analyzes images; media, ad text, and transcripts land in a database. An agent then analyzes trends, compares against your own ads, and finds market gaps based on Reddit-scraped ICP pain points, exposed as an endpoint the whole team can hit from their daily driver Codex or Claude Code.
@MAXIMISEART [Claude Code]
https://x.com/MAXIMISEART/status/2078512393487835572
Polish-language pro tip from practice: Claude Code's default system prompt loads thousands of mostly-unused tokens into the context window at the start of every session. In settings.json you can disable unneeded pieces — AskUserQuestion, cron-task planning tools, built-in code review, dynamic workflows, remote control, artifacts. Reducing initial context from tens of thousands of tokens to a few thousand noticeably improves the experience and keeps the agent in its smart zone longer.
@oscabriel [Claude Code]
https://x.com/oscabriel/status/2078566660974444576
The author formalized his previously shared process of launching dynamic workflows from Claude Code using pi agents running on GPT-5.6 in focused herdr panes. He moved the workflow instructions out of his root CLAUDE.md and bundled them into a skill called /herd-flow with a routing.md reference file.
@polyesp [Claude Code]
https://x.com/polyesp/status/2078486671759475050
After seeing a post about someone hospitalized from rate-limit stress, a developer who used to stay up until 2am to catch Claude Code limit resets built AgentCal: a local calendar that shows each Claude Code session as a block in a weekly view, with parallel sessions side by side like simultaneous meetings. A coming update queues prompts to auto-run the moment your limit resets so you can sleep while the AI works; it runs on a local API with REST, SSE, and webhooks so you can build Slack alerts on hitting limits, and it already supports multi-account setups.
@Jingyuan_521 [Claude Code]
https://x.com/Jingyuan_521/status/2078521114716950627
Chinese-language sighting: the poster was at a techno bar in Chiang Mai and noticed the lighting engineer vibe-coding with Claude Code, apparently using it to control the venue's lights — a user scenario Anthropic likely never anticipated.
@Gonnector [Claude Code]
https://x.com/Gonnector/status/2078444494408863760
Korean builder runs a 13-member multi-agent team orchestrated through Orca: a leader agent "Jarvis" spawns members into independent terminals, worktrees and sessions with chosen model, effort level, and persona, and pinpoint messaging ensures a specialist working in project A never hears project B traffic. He previously modified the Discord MCP in Claude Code's channel plugin for agent communication but hit cross-project noise; token costs are tracked per task and agents adjust effort mid-session for cost optimization. He is preparing to open-source an "Orca chat" native app and credits Fable 5 at high effort with solving many design problems across 4 months of protocol building.
@kemu_aii [Claude Code]
https://x.com/kemu_aii/status/2078585647242375338
Japanese creator automates Threads with Claude Code and reports making about 300,000 yen per month. The setup is four "AI employees" created inside Claude Code that handle research, drafting, and scheduled posting, requiring little ongoing effort; the build steps are shared in a linked post.
@Abobsterina [Claude Code]
https://x.com/Abobsterina/status/2078526313238077784
A founder runs 8 Claude Code sessions at once with $30,000 in credits: one monitor is filled with eight agents working in parallel, another displays customer session replays showing exactly where users get confused, hit bugs, or abandon the product. He watches usage, identifies each friction point, and assigns each problem to a separate agent — one fixes backend, another rewrites the interface, a third investigates errors, and the rest test features, review code, and prepare the next update, often until 4am.
@indiesoftwaredv [Claude Code]
https://x.com/indiesoftwaredv/status/2078410902118801864
A solo dev with 12 apps and no team advises not delegating tech-stack decisions to Claude Code or Codex, arguing AI performs best where the ecosystem is deepest. His stack: Expo for OTA updates with no store review, RevenueCat (zero StoreKit code in years), Superwall for paywalls and A/B testing, and OneSignal for push notifications.
@Brown_Thunder76 [Claude Code]
https://x.com/Brown_Thunder76/status/2078600918556184638
A self-described non-developer found a bounty for building a price oracle on Keeta and built it with Claude Code: it pulls prices from six independent sources, calculates the median, automatically rejects deviating sources, and marks the price stale if fewer than two sources agree. Every quote is cryptographically signed with aggregation method, contributing sources, confidence metrics and timestamps, publishing to the testnet on a heartbeat. He also built a live transparency dashboard, a full automated test suite with CI, monitoring, abuse protection, a separate verification tool anyone can use to independently verify signatures, and an npm package — all open source.
@Roman9078963816 [Claude Code]
https://x.com/Roman9078963816/status/2078362311727661521
Workflow treating notes like a codebase: Obsidian is the editor, Claude Code is the engineer. Drop in a PDF, article, YouTube video, or random thought and the AI breaks it into connected notes linked to existing ideas; asking "what did I conclude about AI agents six months ago" searches your own thinking and answers in your own words. Once a week an agent finds duplicate ideas, detects contradictions, updates stale notes, and reconnects forgotten pages — and because everything is plain Markdown the vault is portable to any future model. The post also notes the Obsidian team released lightweight skill files letting Claude understand Obsidian natively.
@ivan4th [Claude Code]
https://x.com/ivan4th/status/2078273582409343292
Russian-language post: the author built an expense-tracking harness in Claude Code with Ledger CLI under the hood. Everything lives in git, with YNAB-style envelope-method reports in the browser, automated extraction of his transactions from banks despite their resistance, and a Telegram bot on top.
@claudecode84 [Claude Code]
https://x.com/claudecode84/status/2078376275824316507
Japanese developer built a MacBook app that blinks the Caps Lock LED to indicate Claude agent status — ambient hardware signaling for when the agent needs attention — developed using Claude Code and Kimi K3.
@SimonHoiberg [OpenClaw]
https://x.com/SimonHoiberg/status/2078444274752930167
Simon Hoiberg's approach: instead of making the agent smarter, make everything around it more predictable. OpenClaw sits in the middle wrapped in boring infrastructure — n8n workflows, webhooks, cron jobs, reusable scripts, markdown docs, runbooks, and RAG with a vector DB. His heuristics: if he asks the agent to do something twice it becomes a workflow, if he explains a process once it gets documented, and weekly tasks get scheduled. The result is fewer tokens burned re-explaining context and more predictable output.
@Abobsterina [Claude Code]
https://x.com/Abobsterina/status/2078474347258491168
A 24-year-old Chinese developer turned the GTA formula into a Beijing hutong simulator: $7 price, 21,684 buyers, $151,788 revenue against $240 in development costs for Claude and 3D assets. Players drive rickshaws, deliver soy milk before it gets cold, and complete baijiu-drinking missions; he built it in Unreal Engine 5 over 46 days, using Claude Code for mission logic, NPC conversations, the reward system, and hundreds of side quests, recovering costs within hours of launch.
@09pauai [Claude Code]
https://x.com/09pauai/status/2078585662874497088
Japanese user was stressed by long wait times while working in Claude Code, so they built a small visualization app — a fictional company "株式会社ぱうう" — that shows what their agents are doing. Stress dropped sharply even though the app "generates zero yen." They frame this kind of small tooling as the key to sustaining a side business, and note they run the agents 24 hours without breaks.
@note_ai_mousigo [Claude Code]
https://x.com/note_ai_mousigo/status/2078351441182249225
Japanese creator shares a concrete 7-step content routine: pay for X Premium (Grok for research) and Claude Pro, copy the past month of your own posts, dump them into Claude Code for analysis, turn the analysis into Skills, add your own X articles as Skills too, then post daily. Claims following the sequence automatically produces posts in your own style and that this alone is enough to earn from X.
@peter_0123 [Claude Code]
https://x.com/peter_0123/status/2078506526042714132
A non-engineer Windows user uncomfortable with terminals and unsatisfied with VSCode and Claude for Desktop built their own in-house CLI wrapper to run Claude Code, Codex, and Kimi. Each agent gets a Pokemon avatar — Claude Code is Pikachu, Codex is Totodile, Kimi is Charmander — with a screenshot of the custom interface in use.
@pauliusztin_ [Claude Code]
https://x.com/pauliusztin_/status/2078457256362877087
Engineer building a personal AI assistant for an upcoming Manning book describes the architecture he converged on: make the harness disposable, keep the context. Unified memory lives in one MongoDB database supporting documents, full-text search, semantic search, and knowledge-graph traversal; an MCP server owns the business logic — how memory is searched, updated, when conversations become long-term memory. With this separation, swapping between Claude Code, Codex, OpenCode, or Gemini CLI becomes little more than a configuration change and the new harness immediately knows who you are and what you are working on.
@soboozie [Claude Code]
https://x.com/soboozie/status/2078447968663826553
Describes a free Claude Code skill that ingests a YouTube, TikTok, or Loom link, pulls every frame, reads the transcript, and answers questions like what the hook is and what's on screen in the first 3 seconds. A 60-minute tutorial gets summarized in 2 minutes, and a screen recording of a bug becomes a described error plus a written fix. It supports YouTube, TikTok, Instagram, Loom, and local files, costs $0, and is live on GitHub; the author pitches using it to reverse-engineer competitor videos.
@AIkenkyuFM [Claude Code]
https://x.com/AIkenkyuFM/status/2078359112463843532
Japanese post on Claude Code's new automatic project-file indexing: it pre-scans Word, PDF, and text files in a folder in the background and builds an organized index. The use case is tax work — asked about the requirements for a specific tax provision, it instantly identifies which file and section to read from massive tax manuals. The multi-minute wait per question disappears and document search becomes near zero-time.
@Adea0x [Claude Code]
https://x.com/Adea0x/status/2078541949821124804
Describes a two-model setup running inside Claude Code: Fable 5 as orchestrator and GPT-5.6 as worker via the Codex plugin. The steps: install Codex, add the Codex plugin to Claude Code, paste the repo URL, create the custom skill, type /root. Fable interviews you and builds the plan, GPT-5.6 implements, Fable reviews and sends back changes, and the loop repeats until the project is done.
@diamai_ [Claude Code]
https://x.com/diamai_/status/2078414871142580235
Detailed walkthrough applying Boris Cherny's 4-level AI adoption map to a real workflow: Playwright verification moved into a separate subagent that must produce a recording, code and security reviews in separate subagents, repairs stopped after three rounds, and a missing test environment pauses the loop and alerts a human on Slack. He sends a spec to Percy, his custom Slack bot running Claude Code and Codex in the cloud, and gets an open PR with adversarial review 19 minutes later; a weekly PlanetScale routine hunts expensive queries, and a Sentry instruction spins up about ten workers on issues affecting more than 10 users, with phone alerts when the coordinator is blocked.
@VoxelKei [Claude Code]
https://x.com/VoxelKei/status/2078457720836145349
Japanese creator of the tool Spatialograph Maker states that all code implementation was done by Claude Code and they did not write a single line themselves — coming from someone who describes themselves as having chronically failed high school math. Another data point for non-programmers shipping real software.
@jakemor [Claude Code]
https://x.com/jakemor/status/2078313975435915505
Developer built a new product feature in Claude Code entirely from his phone while his Tesla drove him from the city to Long Island. The feature adds a share button to the agent embedded in his docs and dashboard: ask a question, tap share, get a public link; anyone opening it sees the whole conversation and can fork it to keep asking from where it left off.
@marfinxx [OpenClaw]
https://x.com/marfinxx/status/2078488176180785259
A 27-year-old dev mapped an entire B2B client workspace into one live Obsidian graph coordinated by OpenClaw: 14 client projects, 8 B2B pipelines, and 300+ interlinked wiki pages around one central corporate node. Every client file, contract, and compliance audit updates in real time with no cloud databases or API bills, using local Gemma and Qwen models. The OpenClaw coordinator ingests, synthesizes, and links every document locally at zero recurring cost, replacing cloud APIs, vector databases, and CRM subscriptions.
@EliasMunk [Claude Code]
https://x.com/EliasMunk/status/2078620346387710138
Founder replies that his app, made 100% with Claude Code, already has over 1,000 actors using it — a short but concrete shipped-product-with-real-users data point from the entertainment industry.
@Jacobsklug [Claude Code]
https://x.com/Jacobsklug/status/2078457767069876245
Builder reverse-engineered an autonomous coding loop while building a full app live with Lovable and Claude Code, and breaks it into 5 parts: a spec doc defining every feature and what "done" means; an orchestrator that splits the build into phases with exit criteria (build, test, fix, re-test); a separate verifier agent that runs every user story and documents each error for the builder to fix against; parallel loops — Lovable split one prompt into 12 tasks on its own; and QA-based exit criteria so the loop stops only when work is actually done.
@SakaguchiTax [Claude Code]
https://x.com/SakaguchiTax/status/2078393437678960958
A Japanese tax accountant office announces it contracted the Claude AI Team plan from July 2026 and moved all staff to AI-first workflows. The trigger was a March post titled "how a zero-staff tax accountant runs 60 client companies alone with Claude Code," which prompted them to DM the author immediately. A concrete case of a traditional professional-services firm adopting Claude Code org-wide.
@gowtham_ponnana [Claude Code]
https://x.com/gowtham_ponnana/status/2078445164688027701
Developer released "nexus continuity," a tool for when Claude Code limits hit mid-audit or mid-research: one command continues the same conversation into another harness like Codex or Pi. Built because he bounces between 3 different coding agents depending on the task and hated re-explaining context on every switch, since context does not travel. Fully local with zero dependencies.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2078293662862393425
Japanese creator made a video using only HeyGen's newly free media library — 10,000+ songs, 75,000+ images, sound effects, logos — with Claude Code running Fable 5 as the command center. A /media-use skill fetched all assets for ¥0, and implementation, QA via a Sonnet subagent, and rendering were all completed in a single chat.
@Da7_Tech [Claude Code]
https://x.com/Da7_Tech/status/2078558774508323221
User who pays for both Codex and Claude Code and uses both heavily gives a detailed comparison: Claude Code is better at writing, creativity, planning, judgment, and front-end work; Codex offers a broader subscription — Deep Research, image generation, voice, memory — with more generous limits and more frequent resets. Codex sometimes refuses tasks outright, while Claude Code stays flexible: it may drop from Fable to Opus but finishes the job. His biggest complaint is predictability — OpenAI includes its strongest model clearly, while Fable access feels temporary and promotional, so you never know where you stand.
@calbuldelis69 [Claude Code]
https://x.com/calbuldelis69/status/2078353792748458354
Indonesian user documents day 1 of vibe coding with no PC: full Android plus Termux, running OpenCode and Claude Code through Telegram to build from anywhere without opening a laptop. The first project is migrating an app from Base to Solana, testing how far a phone can replace a laptop for shipping projects.
@TaoOutsider [OpenClaw]
https://x.com/TaoOutsider/status/2078576972331094516
Author open-sourced "Outsider Mining," a free tool you run through Codex, Claude, Cursor, or OpenClaw before mining a Bittensor subnet. Given a public subnet repo it maps miner and validator entry points, traces scoring and reward logic, checks hardware and stake requirements, runs wallet-free simulations with sanitized reports, estimates burn and compute costs, and models P10/P50/P90 reward scenarios, ending with a verdict — RESEARCH, GO-DEMO, WAIT, NO-GO, or GO-REGISTER — so you can evaluate a subnet before committing hardware.
@rohitchauhan [Claude Code]
https://x.com/rohitchauhan/status/2078595985794543866
Trader runs a Claude Code setup for backtesting hypotheses based on price, volume, and fundamentals data. He currently scrapes data with yfinance, which he calls inefficient, and is publicly asking for paid API or MCP-based data services to replace it — real non-dev-tool usage plus a clear unmet data-infrastructure need.
@alexhillman [Claude Code]
https://x.com/alexhillman/status/2078275975406600214
User taught Claude to convert any playlist on any platform into a local data store plus YouTube music playlists, and calls it the most amazing small thing he has built with Claude Code — the personal-utility genre of agent use that never shows up in benchmarks.
@Windowside_Pen [Claude Code]
https://x.com/Windowside_Pen/status/2078323287999046118
Japanese user preparing for the Registered Information Security Specialist exam is using Fable 5 to build morning- and afternoon-section study materials from past exam PDFs spanning 2017-2025, combined with reference books, practice-question sites, Gemini Notebook and Claude Code. A concrete non-coding exam-prep application.
@nikotin30x [Claude Code]
https://x.com/nikotin30x/status/2078561833997291893
A laptop ran Claude Code in a loop for 2 hours with access to exactly 3 tools and produced a full city in Unreal Engine built from scratch — an aerial skyline shot plus a playable character running down the street, with no hand-placed buildings. Three rules made it safe to leave unattended: tool access scoped so drift stayed within a small radius, completion checks that did not rely on Claude judging its own build done, and a hard ceiling to stop the cycle whether it converged on attempt 4 or 40.
@SolanaGambe [Claude Code]
https://x.com/SolanaGambe/status/2078538854047351106
A Karpathy-style second-brain workflow: one running stream.md file captures every idea, quote, and link — capture first, organize later, near-zero friction — then Claude Code with terminal access to the Obsidian vault turns it into searchable memory. Claude can grep years of notes and answer questions like "what assumptions have I repeated across three projects" or "which unfinished idea keeps coming back," making past thinking active context for the next decision.
@gaboweb3 [Claude Code]
https://x.com/gaboweb3/status/2078527878170386671
Three concrete principles for making a Claude Code skill keep improving after you write it: keep a gotcha-list section in SKILL.md where every mistake Claude makes gets logged the moment it happens, so the skill reads it first next time and sharpens weekly; use progressive disclosure, keeping edge cases in the references folder untouched until a task needs them; and avoid railroading, where over-specified instructions make Claude follow the letter into a bad outcome — precision only where it must hold, room to adapt everywhere else.
🗣 User Voice
User Voice
1. Fable 5 access feels like a promotion, not a product. Heavy users on both sides of the Codex/Claude Code fence say the same thing: OpenAI includes its strongest model plainly, while Fable access shifts between temporary boosts, 50% caps, and silent fallbacks to Opus mid-task (@Da7_Tech, @totoche, @argofowl).
2. Limits are reshaping people's lives, and tooling is emerging around the wound. A developer who used to stay up until 2am for limit resets built AgentCal to schedule sessions like meetings; others released tools to carry a session across harnesses when the cap hits mid-audit (@polyesp, @gowtham_ponnana, @alphinctom).
3. Context portability is the new lock-in battleground. Users are architecting memory outside the harness — one MongoDB + MCP memory layer makes switching Claude Code/Codex/Gemini CLI a config change — and asking for cross-vendor orchestration where Claude Code delegates to codex/pi/hermes as subagents (@pauliusztin_, @jasonzhou1993, @0xMulight).
4. Security researchers are converging on the same warning: project-level files are a privilege-escalation surface. Malicious AGENTS.md plus misclassified safe commands equals arbitrary code execution, planted memory files attack future sessions, and 135,000+ exposed OpenClaw instances leaked API keys (@SlowMist_Team, @dair_ai, @AethirCloud).
5. The non-engineer wave wants friendlier surfaces and less AI-slop. A Windows non-engineer built a Pokemon-avatar wrapper because terminals feel hostile, while designers complain every AI-built site has the same purple gradients and monospace headers — design-taste skills are becoming their own product category (@peter_0123, @andrestaltz, @0xKento_).
1. Fable 5 access feels like a promotion, not a product. Heavy users on both sides of the Codex/Claude Code fence say the same thing: OpenAI includes its strongest model plainly, while Fable access shifts between temporary boosts, 50% caps, and silent fallbacks to Opus mid-task (@Da7_Tech, @totoche, @argofowl).
2. Limits are reshaping people's lives, and tooling is emerging around the wound. A developer who used to stay up until 2am for limit resets built AgentCal to schedule sessions like meetings; others released tools to carry a session across harnesses when the cap hits mid-audit (@polyesp, @gowtham_ponnana, @alphinctom).
3. Context portability is the new lock-in battleground. Users are architecting memory outside the harness — one MongoDB + MCP memory layer makes switching Claude Code/Codex/Gemini CLI a config change — and asking for cross-vendor orchestration where Claude Code delegates to codex/pi/hermes as subagents (@pauliusztin_, @jasonzhou1993, @0xMulight).
4. Security researchers are converging on the same warning: project-level files are a privilege-escalation surface. Malicious AGENTS.md plus misclassified safe commands equals arbitrary code execution, planted memory files attack future sessions, and 135,000+ exposed OpenClaw instances leaked API keys (@SlowMist_Team, @dair_ai, @AethirCloud).
5. The non-engineer wave wants friendlier surfaces and less AI-slop. A Windows non-engineer built a Pokemon-avatar wrapper because terminals feel hostile, while designers complain every AI-built site has the same purple gradients and monospace headers — design-taste skills are becoming their own product category (@peter_0123, @andrestaltz, @0xKento_).
📡 Eco Products Radar
Eco Products Radar
Claude Code — the center of gravity across every use case genre
Codex — the most-cited second harness, increasingly paired rather than versus
OpenClaw — personal-agent workhorse; also the week's biggest security cautionary tale
Fable 5 — the model everyone wants predictable access to
Kimi K3 — the rising third option, weights landing July 27
Cursor — still the IDE-side staple, now with design mode
Hermes — praised for local-model onboarding; the migration destination in comparisons
Obsidian — the de facto file system of the second-brain movement
MCP — the connective tissue in nearly every serious workflow
GPT-5.6 — the worker model of choice inside orchestrator setups
Grok / Grok Build — expanding fast, inheriting Claude Code's permission model (and its risks)
OpenCode / Pi — the long tail of harnesses users hop between
Gemini CLI — frequent fourth harness in portability discussions
n8n — the boring-infrastructure wrapper of choice around agents
Claude Cowork — Anthropic's non-dev surface, showing up in daily-driver stacks
Unreal Engine — where Claude-built games keep landing
Three.js — the default engine for one-prompt 3D experiences
ElevenLabs — narration layer in content production chains
Voicebox — voice interface repeatedly cited in agent stacks
Mac mini — the emerging home for always-on local agents
Claude Code — the center of gravity across every use case genre
Codex — the most-cited second harness, increasingly paired rather than versus
OpenClaw — personal-agent workhorse; also the week's biggest security cautionary tale
Fable 5 — the model everyone wants predictable access to
Kimi K3 — the rising third option, weights landing July 27
Cursor — still the IDE-side staple, now with design mode
Hermes — praised for local-model onboarding; the migration destination in comparisons
Obsidian — the de facto file system of the second-brain movement
MCP — the connective tissue in nearly every serious workflow
GPT-5.6 — the worker model of choice inside orchestrator setups
Grok / Grok Build — expanding fast, inheriting Claude Code's permission model (and its risks)
OpenCode / Pi — the long tail of harnesses users hop between
Gemini CLI — frequent fourth harness in portability discussions
n8n — the boring-infrastructure wrapper of choice around agents
Claude Cowork — Anthropic's non-dev surface, showing up in daily-driver stacks
Unreal Engine — where Claude-built games keep landing
Three.js — the default engine for one-prompt 3D experiences
ElevenLabs — narration layer in content production chains
Voicebox — voice interface repeatedly cited in agent stacks
Mac mini — the emerging home for always-on local agents
Comments