Super User Daily: July 9, 2026

Matt ran a proxy to intercept and inspect exactly what Claude Code sends to the model at session start, and was horrified by how much cruft sat in the default system prompt. He then used his own settings.json tweaks to strip out all the bloat. The result was a lean 13K-token starting context for each session instead of the padded default. He wrote up the full step-by-step process, proxy included, in an article.

Gyan built reins, an open-source (MIT) tool that lets any shell-capable coding agent (Claude Code, Cursor, Codex) drive your actual logged-in browser with your real sessions and cookies. There is no MCP server to register, no debug port, and no separate automation profile, and everything stays local on 127.0.0.1. You install it with npm i -g @karnstack/reins plus a Chrome extension and the agent is driving in under a minute. As proof, Claude Code wrote and posted this very tweet through reins from his own browser.

Yamariku documented how Okada Polishing, a small factory in Kanazawa, built over 30 internal business systems with Claude Code for roughly 300,000 yen in subscription and token costs, work that would have cost 80 million to 100 million yen if outsourced. The systems include an inspection-report tool that digitized dimension data so workers see target measurements on a tablet, an "OKADA Board" portal aggregating daily checks, production results and CO2 tracking, a lunch-order system, and cost/time visualization. The remarkable part is that a managing director with no programming experience built all this through roughly 1,500 hours of dialogue with Claude Code. His point is that the era is shifting from "outsource everything" to deciding case-by-case what to build in-house versus what to keep on SaaS or SIers, and that this is especially disruptive for regional small and mid-size firms.

Levels got too lazy to order food on UberEats, so he had Claude Code do it for him. Claude simply installed Playwright, he logged into UberEats once on the web (the iOS app being just a web wrapper), and after that he could say "order bananas" and the agent would place the order. It was a quick, low-effort setup that turned a browser automation tool into a personal food-ordering assistant.

Prompted by Marc Kohlbrugge to finally build a Nomads iOS app, Levels didn't want to code locally since he now only runs Claude Code on a VPS. So he asked the agent how to handle it and it suggested renting a MacinCloud instance, which he did, then handed the login to Claude on his VPS. Now Claude is building the iOS app over SSH using Xcode on a Mac Mini in California. His whole motivation is mobility: he works roughly 50-50 on his iPhone and MacBook via Termius SSH to the VPS, so he never has to lug a laptop around to track progress.

This post breaks down an autonomous agent architecture built as a single folder containing CONTRACT.md, schedule.yml, rubrics/, state/checkpoint.json, and receipts/. The claim is that this one folder managed 5,382 shifts of AI work, and that after 20:41 the previous night with zero human keystrokes it produced 14 pull requests, 3 rollbacks, 6 summaries plus log cleanup and state saves before anyone woke up. The idea is that the folder structure itself, not chat, is the agent's operating system: CONTRACT defines rules and boundaries, schedule assigns work, rubrics grade quality and force redos, checkpoint stores memory, and receipts keep an audit log. The author argues this pattern applies directly to running Claude Code and Codex agents and predicts it will become a standard AI development layout.

Arrakis built a company-operations system around Andrej Karpathy's idea of an LLM Wiki, capturing every employee's responsibilities, workflows and operational context so the system understands how work is organized before executing. From a digital-twin workspace he issues a single command, and the system decomposes it into specialized skills per business function, each handling its domain and handing output to the next agent until the workflow completes. Codex is the default execution engine, while OpenClaw and Hermes are reached through a bridge when better suited for a task. For anything needing security permissions a human just approves the request and the agents take over, so his role has shifted from doing the work to managing, approving and supervising the system from anywhere.

Xudong shares a closing habit for Claude Code: before shutting down a long session, have it write a handoff document to HANDOFF.md covering what task you're on, what's done, where you're stuck, the next steps, and which traps never to repeat, written for a fresh session with no context. The next session then opens with "read HANDOFF.md first." He frames this as the lowest-cost way to give AI cross-session memory, since the AI's memory is bounded by the session while a project spans weeks, and a plain markdown file bridges that gap without any plugin. His analogy: a smart but daily-amnesiac employee needs a work journal more than more intelligence.

Kaolti had always loved topographic maps and wanted to recreate that look in Three.js. Using Claude Code with Fable, he built an interactive topographic-map visualization in about three hours. The tweet shares the resulting demo.

To test Fable against Anthropic's Opus 4.8, Contraben ran a blind design bake-off. They built 5 real landing-page and portfolio briefs in Claude Code and had 9 working designers judge the outputs blind across 90 matchups. Fable's best brief won 88.9% of its head-to-heads while its worst output won only 11.1%, and his takeaway was that the quality of the brief drove nearly all of the difference.

Patrick had Claude Code with Fable write a FlashAttention forward kernel targeting the NVIDIA B200 (Blackwell), written entirely in pyptx, his Python DSL that emits raw PTX with no CUDA C++ or CUTLASS. The generated kernel runs at 0.92 to 0.99x of the hand-tuned CUTLASS FlashAttention-4, hitting parity on two sequence lengths at roughly 1350 TFLOPS bf16. Out of the box the pairing reached 0.71x (878 TFLOPS), and an improved harness he built for the task let Fable close the remaining gap. It uses tcgen05 tensor cores, TMEM, warp specialization and mbarrier pipelines, all driven from Python.

Xing pushes back on the hype that tools like Hyperframes or Remotion can one-click auto-edit high-quality video. He points to a creator who used Claude Code plus Remotion to produce a Vox-style special-effects video, the kind that does well on YouTube, as the real way to do AI video editing. He breaks down the workflow with timestamps: planning the script and visual style at 1:30, building scenes with Claude Code plus Remotion at 2:26, assembling and fine-tuning the edit at 8:00, and narration, music and final render at 11:01.

Omar worked through a lot of PRs to add first-class iMessage Polls support to OpenClaw, landing in the next release. You can ask your Claw to send a poll to a group chat, it will send you one, and you can send it one back. He's very happy with how it turned out and thinks OpenClaw now has the most feature-rich iMessage support around.

09pauai argues that even a rough, low-effort short video is enough to achieve the 0-to-1 milestone, showing a short he made with Claude Code in 10 minutes. At that quality level he says you could churn out 300 videos a day, so the play is to mass-produce clips tuned to your niche, whether spirituality, career-change, health, dating, or certification prep. He links to a walkthrough of how the videos are made.

Vida shared three months of his personal Codex and Claude Code token-usage charts, estimating he burned through over $10,000 worth of OpenAI and Anthropic tokens on his own account. He's been building everything he ever wanted to make: company project optimizations, games he wants to play, game cheat tools, whatever pops into his head, and describes being deeply addicted to creating infrastructure and reshaping his "little universe" on demand. Beyond his own use, he gifted several relatives and friends $20-100 Codex subscriptions (buying KYC'd cards to pass along since Codex can't be gifted). He notes he'd have given Claude instead, but every Claude account he gave away got banned.

Om highlights an app someone vibe-coded with Claude Code that turns push-ups into a live 1v1 PvP battle. You get instantly matched with a random stranger online, and both of you race to do as many push-ups as you can while your webcams track and count reps in real time. You see your opponent's live feed next to yours for a genuine head-to-head race, and first to out-rep the other wins. The pitch is that nobody wants to grind push-ups alone, but making it a live contest against a stranger gets people going all out.

Gimu created game visual effects using a pipeline of Fable 5, GPT image 2, and Blender, all orchestrated through Claude Code. Claude Code drove Blender via Blender MCP and called the GPT image 2 API: the 3D models were made in Blender MCP, textures were generated with GPT image 2 and applied onto the models to finish the effects, and Claude Code also built the preview screen. He suspects Fable 5's strong visual perception pairs well with driving Blender operations, and credits @rssk_it for teaching him the method.

Definition has been asking Claude Code to show him a feature before building it, prompting it to create a plan plus a low-fidelity artifact of how the UI will look and how the systems fit together, and to keep both in sync as they iterate. This shifts more of the iteration out of code and into the plan itself. He shares an example of the approach in action.

While Fable was still available, Takasek built out a self-improvement ecosystem for his Claude Code environment and realized improvement comes in about three kinds. First, rot detection across seven existing systems like skills, memory, and hooks. Second, taking lessons from a session's trial and error and routing them into lint rules, CLAUDE.md, or skills so they stick. Third, crawling external distillation-source URLs to pull in new knowledge.

Gaurav interviewed a developer with strong fundamentals, good communication, and reasonable problem-solving, but when asked what AI tools he used daily, the candidate said "using AI is cheating, and I don't cheat." Gaurav gave examples of shipping faster, better quality, and less drudgery, but the developer kept insisting it was against his ethics. He then handed over an assignment the candidate estimated at 2-3 days and solved it in 15 minutes in front of him using Claude Code, refining the solution several times, yet the candidate still called it cheating. Gaurav rejected him, arguing that a developer unwilling to learn AI won't have a job in the future no matter how skilled.

Snowmaker is building a web app that integrates with several dev-tools APIs, and describes the wild new debugging loop in the Claude Code era. Whenever he hits an API issue, his coding agent composes a bug-report email that he sends to the vendor; they paste it into their agent, which sends feedback back to his agent. Humans are still the slow message-passers between the two "superintelligences," but he says progress is 100x faster than the old days. It feels to him like an early taste of a workflow that's about to become common.

Davis makes the case for getting a Linux box or home server to run background agents, saying it's absurdly worth it and offloading work from your laptop feels great (with Devin and Cursor as alternatives if you'd rather skip hardware). He lists the setup: Tailscale on all machines, color-coded tmux that auto-opens on SSH, T3 code running on your "code" machine via "npx t3" with the UI accessed from your laptop anywhere, passwordless SSH between your own devices, and a skill listing all your devices and IPs so Codex and Claude Code know how to SSH between them. He also recommends a KVM for remotely rebooting machines.

Iluciddreaming lays out a near-zero-cost content pipeline for regular people making media in 2026 using Codex/Claude Code plus HyperFrames and Edge TTS. You describe your requirement in one sentence and the AI writes the code, does the voiceover, and renders out an MP4. He ran the whole thing himself and went from zero to a finished video in minutes.

Six months after the Claude Code explosion, Ahall gives a field report from political science and political economy research. He says coding agents keep improving in ways he values for research and can't imagine working without Fable now, with the biggest gains in verifiable mathematical and technical work like automated replications, living research that updates dynamically, and software prototypes, though he's seen little improvement in generating good research questions or insights. He's blunt that the models' ability to actually write papers is horrendous and maybe getting worse, with a "Claudeslop/GPTslop" tone he can't fix, so he still writes papers himself. He notes real uptake of these tools in empirical research and journals adapting slowly, and sees the future in building increasingly complex research that complements human researchers with good judgment.

Spent the weekend doing a thorough overhaul of the PC version of their radiology diagnosis "cafe" web app using Claude Code with Fable 5. The payoff was mostly on performance: they cut the initial load burden by more than 80%, so load speed improved dramatically. They also refreshed the design, which now looks considerably more polished than before. Since this pass focused on desktop, it's hard to notice on mobile, but on PC it's a big visible change, and they're now moving on to the smartphone version.

Recounts how Roy Mann, co-founder and co-CEO of monday.com, treated deploying an AI agent like onboarding a real employee rather than installing software. He bought a new blue Mac and a phone, installed the open-source OpenClaw "lobster" agent on it, and named her Nova, giving her personal equipment from day one. Nova now runs the Google account for monday's agent lab, buys domains, ships releases, and single-handedly built the foundation of a new product: a job board for hiring AI agents. The author's sharpest point is the contrast: Roy handed the identical OpenClaw setup to 20 employees on recycled laptops and nobody used them, and even Roy's own calendar-assistant Claude got forgotten in a tab. The lesson he draws is that an agent survives only when it gets real onboarding, a name, and a seat in the team WhatsApp group, not just an install; monday now runs around 700 OpenClaw copies company-wide.

Wrote up a full build log of an autonomous US-stock trading bot he assembled with Claude Code just by asking it in chat to "build an automated US stock trading bot." The result runs four strategies in parallel (dividend aristocrats, momentum, reversal, low volatility) every weekday at 6am on a VPS and reports to Slack. The clever part is the four-layer safety design: rule strategies generate signals, Claude re-reviews each signal and rejects low-confidence ones with reasons, a human approves the actual order list with a 30% sector-concentration cap, and only then does it report to Slack with company names plus buy/sell rationale. He stresses starting with paper trading before real money and notes results are only one month in and not a profit pitch, but the point is turning AI from one-off chats into a system that runs for you every morning.

In one week, using Fable, he built a new diary app, transcribed and translated 65 hours of Islamist sermons in Australia, and translated a 145,000-word Italian novel he'd wanted to read for years but couldn't because no English translation existed. He's blunt that he still can't code: he just asks Claude to generate Claude Code prompts that he copies and pastes into a Terminal window, and it does everything for him. He frames this as evidence there's no AI bubble, because people with zero math or coding background will suddenly be able to do things like this, creating demand out of nothing. Basically a pure non-technical user living entirely off copy-paste prompts.

Describes how he improves the reliability of his agentic loops with human-in-the-loop escalation delivered over voice and phone rather than being chained to a computer or Slack. All his Claude and Codex agent sessions now use the DialAgent MCP server, which provisions each agent its own phone number for calls, SMS, and iMessage as native tools. When his automations work on PRs and new features, the agents escalate decisions to him via a short phone call, which he says is especially useful when he's on the road. He shares the exact starter prompt to have Claude Code or Codex grab a Dial number and call you to confirm setup.

Uses Claude Code as a personal health manager that already knows his dietary preferences, workout schedule, health goals, and body metrics. It builds his meal plans and then uses the Instacart connector to actually order all the groceries for him. On top of that, it generates charts and graphics of his data that he sends to his trainer. A tidy example of chaining personal context, a real ordering integration, and reporting into one loop.

Got fed up being the human router between Claude Code, Codex, and Cursor, each keeping its own context so he wasted half his time re-explaining the same project to each tool. So he built Commonly, a Slack-shaped workspace where agents from any runtime join as members with a persistent identity and project memory. You chat with them, drop files on them, and they answer and file deliverables back. The key claim is they remember everything into the next session even when you switch tools.

Wanted B-roll footage of himself typing prompts into Claude for a video, and found the usual route (typing, filming, cutting) super annoying. So he wondered whether he could just ask Claude Code to generate the animation instead. He says it one-shotted it perfectly, and all he supplied was a single screenshot of the Claude chat interface plus the five prompts. The tweet shows one of the five resulting videos.

Pointed Fable 5 at his rotting Obsidian vault, 3,312 markdown files with 2,416 orphaned (72% linked to nothing), which doubles as the working memory of his OpenClaw agent Zoe that runs 53 live cron jobs over it. With a plan and a /goal he walked away, and four hours later the vault cleans itself nightly at 3am. Fable's diagnosis was brutal: the automation map was 80% wrong, boot files ate 73% of the 60K context budget with two contradictions, ~1.4GB of dead weight including a 654MB stale backup inside the vault, and a plaintext key in TOOLS.md. It rebuilt this in two stages: one generated ledger for all 53 cron jobs rendered nightly and purged boot files to zero live state, then split the vault into a knowledge layer and a machine layer with 20 exclusion patterns, consolidating 827 files with a nightly mechanical checker for every rule. The job cost 31M input tokens (94% cached) and 438k output across 338 turns with one orchestrator and 14 subagents, and his point is that a nightly gardener beats one-time cleanups that just rot back.

Couldn't understand why Claude Code burned through so many tokens per interaction until he actually looked at the full raw output of its requests. His punchline: every single time it's writing him a whole essay. This is less a build and more a pointed observation about the tool's verbosity driving token consumption, with a screenshot as evidence.

Highlights "Jarvis," an Iron-Man-inspired Claude Code setup built by Icare1337 for bug bounty work. It's a self-managing multi-agent system that has substantially sped up his bug bounty workflow. The one hard rule they emphasize is that every finding still requires human validation before it counts. Posted by the YesWeHack account as a spotlight on a community hacker's agent rig.

Open-sourced his AI agent skill collection "rnskill," currently five skills in three categories. Writing: rn-renhua strips the "AI flavor" from Chinese AI-tech articles, deleting filler like "not A but B" and "actually/essentially" while keeping the judgments and test details, which he runs over every tweet and review before posting. Video: rn-motion-director is a director layer that turns topics into motion concepts and forbids PPT-style videos where every frame must actually move, rn-dark-saas-video is a full style blueprint for dark SaaS product shorts with 8 scene templates and 9 hard rules, and rn-bw-text-opener does black-background white-text typewriter openers with a Python script that auto-computes each character's timing and sound-effect timestamps. QC: rn-replica-qc does closed-loop alignment when replicating reference videos with three fidelity levels, frame-by-frame breakdown, and PSNR/SSIM computation. It installs in one line and works across Codex, Claude Code, and Trae Solo, each skill shipping a SKILL.md plus openai.yaml.

A complaint rather than a build: he calls Claude Code stopping midway through a task one of the worst UX problems, wishing it would just finish what it's doing like Codex does. He's left stuck with a partially completed task. He adds he doesn't even want to buy a second account since Fable is about to be gone. Straight user frustration about the agent bailing before completion.

Spotlights someone who fully offloaded an ad agency's prospecting to Claude Code. You feed it a single niche (like skincare or hair salons) and it scans every brand currently running in the Meta Ad Library, then scores only the ones whose creative is "weak" despite spending on ads (too few ads, stale, static-image only). It pulls that brand's actual ads, critiques the weaknesses from a creative director's viewpoint, and identifies the email and official site to output an evidence-backed proposal list with contacts. The framing is sharp: agencies always assumed their bottleneck was production quality, when the real chokepoint was the research to find prospects who have budget and need, and Claude Code now does that upstream part.

Reflects on a project he'd launched months earlier called Juste Prix, a piece of software he'd developed with the help of Claude Code. It drew a lot of reactions both positive and negative, and he's decided to shut it down, not because it was AI-built but because it exceeded his ability to understand some essential steps of its development. His takeaway is that this kind of software creation isn't for him under these conditions, yet the experience taught him a lot, mainly that AI now lets you carry a project to completion even without all the technical skills up front. More an honest post-mortem on the limits of vibe-coding something past your own comprehension than a success story.

Pushes back on "AI-made stuff looks cheap" by showing a corporate website he built almost entirely with AI. Generation and coding were nearly all handed off to AI; his own work was the design brainstorming and giving "fix this here" instructions. His pipeline: about 3 hours brainstorming design and tone with Claude, about 1 hour on the hero background video (Higgsfield to ffmpeg), a few minutes of design-to-implementation with Claude Code, and about 2 hours of revision instructions that Claude Code applied. His thesis is you can delegate the bulk but humans must own the design and polish, and he's raised this to delivery quality and is actually taking paid client work with it.

Shared a creative video piece where images were made in MidJourney 8.1 and animated in Seedance 2.0. The distinctive part is that it was prompted using "Night Shift," his own custom Claude Code system built specifically for creators. So Claude Code functions as the prompting/orchestration brain driving his image and animation pipeline. Presented as a personal "itch to scratch" showcase of that custom setup.

Had Fable 5 (via Claude Code) build a clone of the recently buzzed-about drawing/hide-and-seek game from a single detailed prompt. His spec was demanding: a browser-only hide-and-seek game running on pure Three.js where hiders eyedropper the surrounding colors and paint their bodies pixel by pixel, the seeker's view blacks out for the first 30 seconds, real-time joining via room codes, two rounds with role swaps, and a Minecraft-like town built from detailed textured voxels rather than flat colors. Claude Code just said "here you go." He notes there are several things to nitpick, but everything in the original instruction was implemented with no omissions, and for a one-shot result Fable 5's level of completeness is impressive.

Reports that a fortune teller (Spica-sensei) packed a top pro's know-how into Claude Code to build homemade digital content, which has already sold 600 copies. The framing is that AI is again changing the shape of fortune-telling, with quality content now going straight from the fortune teller into users' hands. He's essentially witnessing a practitioner productize their expertise via Claude Code and sell it directly. More of a celebratory observation than a technical build log.

Hit his Claude Code quota with the code only half written, then handed the work off to Codex seamlessly through mempal. He describes the transition as buttery smooth. The point is using mempal as a memory/context bridge so a different agent can pick up exactly where Claude Code left off when limits run out. Includes a demo link showing the handoff.

Relays a dev who runs a six-figure agency on Claude Code arguing Fable 5's cheap era ends within days, so he crammed his top use cases into a three-minute video. First, he cloned a $15/month dictation app in an afternoon: deep research pulled apart how it works and Fable 5 rebuilt it locally, free, private, and his to modify. The one people sleep on is number two: he feeds his last 50 Claude Code sessions back into the model and asks what he's doing wrong, turning the answer into skills, then automations, which forms the backbone of his agency. The finale includes a browser game that looks studio-made, existing because this model handles projects Opus couldn't finish. His overall point is that rented compute gets repriced but what you build with it stays yours.

After needling Kuba (pesvklobouku) about how he'd tackle "historical weather" as an agentic app in Claude Code, he decided he had to actually build it himself. Once it was working, he wrote up a summary of how you can construct such a thing with AI agents. He links to the writeup, flagged as premium content. A case of turning a bit of banter into a real agentic Claude Code build plus a how-to.

Asked Fable to refactor and improve a sim-racing skill his OpenClaw uses for telemetry and race results, and it missed several core things that were obviously outdated or broken. When he asked for a second pass it still missed a couple more, and started burning through his credits. His verdict is that it lacked judgment and common sense on what was admittedly a very specific and personal task. He uses this to push back on people calling it AGI, finding that label ridiculous given the result.

He built a cinematic real estate website and packaged the reusable Skill he used into a public GitHub repo. The Skill steers Claude Code on the visual direction, scene composition, scroll motion, and video for the site. He's inviting people to try it and show him what they build with it. It's a nice example of turning a one-off build into a shareable, opinionated Skill rather than just posting the result.

Two weeks in, he posted a blunt list of what he considers dealbreakers with Claude Code. It's memory-heavy so his machine slows to a crawl and blocks other work, it can go off the rails during long sessions so he pairs it with ChatGPT to supervise, and a weak spec means the computer freezes (and there's no phone option). He also notes you can't brainstorm loosely the way you can with ChatGPT, so you need real programming knowledge, and that the ~3,000 yen plan wasn't enough so he moved up to the 15,000 yen tier. His verdict: absolutely not for AI beginners.

He flagged that Qwen3.6-27B overthinks badly, to the point he had to disable reasoning entirely when running it on a DGX Spark because it was otherwise unusable with OpenClaw. He's praising a project aimed at improving reasoning efficiency and pointing people to its work and evals. It's a real usability complaint from actually running the model locally through OpenClaw, not a benchmark take.

He relayed that a Google engineer used Claude Code plus Fable 5 to port the classic 2003 PC game Command & Conquer to native iOS in "a few hours." The claim is the speed of a full platform port done through the model. Take the "few hours" framing with some skepticism, but the point is a native iOS build of a well-known game via Claude Code.

While comparing NativePHP against React Native and Flutter for mobile, he expected LLMs to handle the two established frameworks far better since NativePHP is a newcomer. What he found instead is that NativePHP leans on Laravel Boost and ships Skills that reference its docs, and that's enough for LLMs to work with it competently. He made a YouTube video specifically on building a NativePHP mobile app using Claude Code Skills, plus a full one-hour course comparing all three frameworks.

He built a system that turns any product into an end-to-end AI ad in whatever style you want, from claymation to fake doctor-patient to interview formats. You feed in the product info and benefits, then human strategists plus Claude plus Arcads produce the videos, built across Claude Code and Arcads. He claims $0 upfront generation cost and ads that went viral within two weeks, and he's making full AI UGC videos in about three minutes. His pitch is that AI alone doesn't produce viral scripts, the workflow and strategy do, and he's packaging the 7-step process, prompts, and tool stack for people who repost and reply.

His AI affiliate account is doing well at 200,000 yen a month, so he's spinning up a second account and having Claude Code (running the top-tier Fable 5 model) handle everything from concept creation onward. He's basically delegating the whole account-building process to the model and betting it grows like the first one. It's a straightforward "this worked once, now I'm cloning it with the AI doing the thinking" play.

He laid out a detailed token-economy strategy for Fable 5. The core moves: drop the Effort Level to Low or Medium for most coding since it gets close to the high settings for far fewer tokens, and reserve Max for architecture and expensive-to-get-wrong decisions. He advocates splitting planning from execution (Fable maps architecture and risky files, then Sonnet or Opus does the actual coding), using Advisor Mode where Fable only answers hard escalations, using Ponytail to cut Claude Code's verbosity by roughly 20%, and running Opus Deep Research first to gather fresh docs before handing the package to Fable. The whole pattern is Fable for judgment, cheaper models for repetition.

He shipped a full tarot web app called Barot in a single day, built end-to-end with Claude Code. The details he's proud of are the taste-level touches: a breathing ritual, 3D card flips, and handwritten verdicts, all of which came out of the conversation with the model. His argument is that people underestimate how good AI is at aesthetic judgment, not just code. It's live with free readings.

Replying in a thread, he argued that with Claude Code configured properly and real programming skill (his bar: you need to be a senior who knows the codebase), it's a serious productivity boost. He now solves in hours what used to take days, which frees him to refine other tasks. His framing is explicitly that it's a force multiplier for experienced engineers, not a shortcut for people who can't code.

He recapped an interview between Claude Code creator Boris Cherny and DoorDash co-founder Andy Fang. The hook is that Fang had stopped shipping production code years ago as DoorDash scaled, and Claude Code pulled him back in, to the point he set himself a rule not to write code manually and was shipping in five different languages to production while running multiple agent sessions at once. The standout company-level number: a migration that used to take four engineers a full quarter was done by one person in three weeks. He's pointing to the free ~25-minute video.

He set up Claude Code with the Racoon Marketplace running 24/7 and woke up to find it had autonomously bought a game account while he slept. The setup watches 1,000+ listings across 8+ games, snaps up underpriced ones in under 60 seconds, and takes plain-English instructions for what he wants. He says it cut roughly 10 hours a week of manual scrolling down to basically zero, and his only job now is checking what it flagged and deciding buy or skip.

He's releasing his full "Kuryu-style" ad-operations AI agent system, built on Meta ads plus Claude Code plus GPT-Image2, which he says automates 90% of ad operations. It's structured as nine AI-agent "departments" running end to end: command center, knowledge management, research, strategy, LP production, banner creation, operations, analysis, and reporting, and he claims non-engineers can build it by following the steps. He reports sales and profit tripled and ROAS reached 600% after deploying it. He's selling it via Brain with a stack of purchaser bonuses and asking followers what perks they want.

Comparing tools purely on behavior for game development, he concluded that for someone who can already code and build a game in an engine, GPT-5.4 or 5.5 CodeX with Rider is currently the better choice. He says Claude Code with Sonnet 5 is fine too, but that Opus and other models tend to start doing things beyond the spec, get themselves stuck, and waste time debugging problems they created. It's a hands-on judgment that the higher-reasoning models overreach on game-dev tasks.

He argues API-relay providers face an impossible trinity of high quality, cheap, and stable. His team burns hundreds of millions of tokens a day and only uses Claude and GPT, and his takeaway is that for production you prioritize quality and stability, while for basic tasks and OpenClaw-style workloads you optimize for cheap. So they purchased two tiers of provider service for the two use cases, with a 4-to-5x price gap between them. It's a practical procurement split from someone operating at real scale.

He profiled a $700 Beelink mini PC (grabbed below its $900 list price) that ships preloaded with OpenClaw and 24GB of RAM. The key detail is that the LLM runs entirely on the SSD via Qwen through llama.cpp, so not a single token leaves the box, with OpenClaw sitting on top 24/7 as the gateway he texts on Telegram like an intern. While he sleeps it sorts files, drafts emails, and runs programs, all from a drive that fits in one hand. The whole pitch is owning a private local AI employee with no API meter, no monthly bill, and no cloud reading his documents.

He reported that Sanofi saved millions of euros this year by replacing some Salesforce and ServiceNow tools with AI agents like Claude Code. The French pharma giant (75,000 employees, roughly $100B market cap) is part of a broader wave of large companies cutting classic enterprise-software subscriptions and internalizing those tools. He notes Salesforce stock is down over 50% from its December 2024 peak, and that Salesforce's CEO pushed back that you can't just "code on the fly" software as reliable and secure as a proper enterprise solution. He asks whether classic SaaS like Salesforce will survive.

He highlighted the Alberta, Canada provincial government case where Claude Code scanned 466 million lines of code in 20 hours. The framing is that an audit that would traditionally take about 6.5 years was run in parallel across roughly 50 agents. His takeaway is that companies sitting on large legacy codebases especially should pay attention to this. It's a scale-of-parallelism argument for using agents on massive audits.

He's assembling a single PC running three free models at once as proof that local hardware has caught up: DeepSeek R1 for math and reasoning, Llama 3.3 70B for everything else, and a vision model on top for images, all local and free. The economics are the point: about $9 a month in electricity versus the $459 a month he used to pay for Claude Code, ChatGPT Pro, and Cursor, which added up to $5,500 a year. He says the hardware costs less than one year of that bill and never sends another invoice. He ignored the renewal reminder that hit while he was finishing the build.

He claimed the president of Y Combinator replaced his entire engineering team with Claude Code, redoing in two months and with zero teammates what first took two years, 10 engineers, and $10 million. The workflow: 10 to 15 Claude sessions at once, up to 50 pull requests a day, and treating the AI like a real team with roles, review, and QA rather than a single chatbot. He notes the free toolkit built for this already has more GitHub stars than Ruby on Rails, and that in 21 minutes it runs a raw idea through YC-style office hours, adversarial review, live design, and automated testing. He's tying it to his own writing about running a one-person $10M+ company.

He cancelled seven AI subscriptions and built a local "Jarvis" that runs entirely in his room, visualized as a living graph of every note he's ever written lighting up as it reasons. The rig is one RTX 5090, two 5060 Tis, 64GB of local VRAM, and about $5 a month in electricity, with Claude Code running the whole thing on his own hardware. His math: he was paying $340 a month ($4,080 a year) across seven providers, six of which he hadn't opened in a month. He argues you don't need his $10k setup, since a used 3090 at $700 can run a 27B model locally and pays for itself in two months, with nothing he says ever leaving the room.

He claims a website build that used to cost $5,000 and a team of five designers now runs on one laptop at $40 a month and two hours per site. The pipeline: Claude Code builds the site from a reference folder in about 90 minutes, two design Skills stop it from repeating the last AI-generated site, and Vercel deploys it in 15 minutes. He reports four sites shipped this month for $9,200 in revenue. He says every Skill and prompt is in the linked article.

He wrote a long, skeptical analysis of Hassam Bhatti, a non-coder in Toronto who built "Launch Fast," a SaaS for Amazon sellers, in 48 hours and grew it to $30k MRR. His sharper point is that the real lever wasn't the tech but distribution: Bhatti already had access to the Legacy X coaching community full of sellers who wanted exactly this tool, and he traded equity rather than ad spend to get in, which is also the single-point-of-dependency risk. On the build itself, the documented stack includes Cursor, Vercel, Supabase, Resend, Apify, TypeScript/Next.js/React, and Cloudflare, with Claude Code and Codex used interchangeably by task and terminals managed via Superset, plus a self-built MCP server letting AI agents do product research and ad optimization. His conclusion is that as coding got cheap, the scarce, non-automatable asset is owning access to an existing crowd.