Super User Daily: July 7, 2026
Two things dominated real Claude Code and OpenClaw usage this cycle, and they pull in opposite directions. On one side, the ambition keeps climbing — people shipped a $317k municipal traffic system, a treasury dashboard a bank CEO built himself in 20 minutes, browser games wrapped for 90-million-player portals, legal toolkits, astrology engines, CAD models of Stirling engines. On the other, the arrival of usage-based pricing put cost front and center: surprise overage bills, tricks to render context as cheaper image tokens, and a steady migration toward local hardware to escape the API meter. The through-line is that Claude Code has stopped being a coding tool and become a general-purpose automation engine — and users are now serious operators optimizing spend, memory, and model routing. Below are the concrete cases, the recurring asks, and the tools showing up around the ecosystem.
@kilo_cpa [Claude Code]
https://x.com/kilo_cpa/status/2073719014371549477
A 20-year-old in China pointed an old camera on his balcony at a road, spent 9 days writing code with Claude, ran up $20 in API calls, and sold the finished traffic monitoring system to a city district for $317,000. The stack was an RTSP feed from a consumer IP camera, YOLOv8 running on the stream, Claude Code writing the tracking, speed estimation, plate reader, and dashboard, all on a $250 mini PC. The framing is that the tech has been open source since 2015 and nobody had packaged it for a specific buyer. The author notes the 9-day municipal sale is an outlier, with the recurring $199/month monitoring being where the steady earnings actually live. The point pushed hard: the gap between "code exists" and "product exists for a customer" is where the money is.
@Jason23818126 [Claude Code]
https://x.com/Jason23818126/status/2073691179573744081
Someone open-sourced their investment research system after reporting two years of gains totaling +1.46M, with live results of +69.29% in 2024 and +66.38% in 2025. Called AI Berkshire, it isn't "AI stock picking" but rather a codified research process built from the frameworks of Buffett, Munger, Duan Yongping, and Li Lu. It runs company analysis across business model, moat, cash flow, valuation, margin of safety, management, and risk, then uses Claude Code and Codex to run multiple agents in parallel for financial-report analysis, valuation, cross-verification, and investment memos. The stated value is turning a real analyst's judgment path into a repeatable system rather than treating AI as an answer machine. Shared as learning material, explicitly not investment advice.
@jaredrhod [Claude Code]
https://x.com/jaredrhod/status/2073840273914728935
The claim: you don't need 40 different AI agents, you need one. The author says he runs a $100k/month business entirely with a single agent using Claude Code plus Obsidian. A 7-minute video walks through how the setup was configured. The whole pitch is the minimalism of a single-agent workflow over sprawling multi-agent stacks.
@marfinxx [OpenClaw]
https://x.com/marfinxx/status/2073773195400728955
An agent team supposedly doubled monthly revenue to $12,800 after swapping OpenClaw for "Hermes Agent." The setup ran multiple bots like Outreach, Research, and Coder models on a local dashboard as a virtual team executing tasks on a board. The argument against the old approach: legacy OpenClaw daemons run commands without process isolation and expose client API keys, while storing agent memory in flat markdown files causes context bloat and memory decay that chokes local GPUs and drains token budgets. The proposed fix isolates execution in Docker containers and uses SQLite FTS5 to compress logs, claimed to cut API bills by eighty percent. This reads like promotional content for the "Hermes" migration rather than a verified result.
@v_nefodov [Claude Code]
https://x.com/v_nefodov/status/2073718410085630259
A 20-year-old student in China mounts her phone under a handheld LiDAR scanner, taps one button, and walks a property while the scan rebuilds trees, walls, and pavement as coordinated point clouds. Claude Code processes the point cloud, stitches the walkthrough, and packs it into a browser link the client opens with no app and no VR headset, in under 20 minutes. The local market rate is $600 to $800 per property; she charges $370 and delivers same day, and property agents now wait in line because she's faster than firms with ten employees. The tweet frames it as one person with a scanner and Claude turning rooms into money with no office or team. The described mileage on her phone was 0.35m, sold as "the business starts with a single step."
@v_nefodov [Claude Code]
https://x.com/v_nefodov/status/2073814469537411171
A student in China connected Claude Code to his drone's camera feed so the drone flies over a highway and the software finds every vehicle, boxes it, classifies it, and calculates speed in km/h in real time from 449 meters up. He didn't write the computer vision himself; he described what he needed and Claude Code wrote the vehicle_speed_tracker while he charged the drone batteries, with the whole system working that same evening. A traffic survey company that normally pays $2,000 to $5,000 for a single radar-tube speed study saw his demo clip and bought the code for $3,800, then asked what else he could build. The framing: no CV degree, no team, no hand-written code, one drone and one prompt producing a Python file worth $3,800.
@ardchain [Claude Code]
https://x.com/ardchain/status/2073704224332841355
A creator reportedly makes $13,000 a month from crude 2D stick-figure animations generated on autopilot, and this breaks down his 14-minute masterclass. The workflow: generate the voiceover first and extract exact timestamps (0s, 7s, 15s) to act as the rhythm for scene changes, then plug Claude Code into Higgsfield by installing its CLI so Claude can generate images locally. A master prompt tells Claude to read the script, find the timestamps, and batch-generate one simple drawing per second of the timeline, producing 20+ assets in minutes with files renamed to match timestamps like "7_seconds.png." Assembly in Premiere then takes about 5 minutes since every file is named by its timestamp. The stated secret isn't a complex AI swarm but the exact order of operations combining Claude's reasoning with Higgsfield's image generation.
@HodlReaper [Claude Code]
https://x.com/HodlReaper/status/2073899926560325921
Someone fed every Claude Code chat, Apple Notes export, and ChatGPT history into Obsidian to build what he calls a "super brain." Rather than fancy plugins and rigid rules, he dumped raw data from Claude projects, sales reports, UGC scripts, and trading notes into one vault, letting Claude organize on ingest and Obsidian connect the dots in graph view. The concrete payoff described: the graph shows how his clothing line's sales spikes link to specific Instagram hooks he tested the prior month. He built a PDF guide covering the ingest workflow, a nightly Claude optimization loop, and the folder structure that scaled his UGC operation past $10k/month. The pitch is an organizational system, not magic: feed it everything, get back clarity.
@1osabori [Claude Code]
https://x.com/1osabori/status/2073689775157231693
A new account launched 4 days ago hit 101 subscribers, which the author frames as evidence for a strategy shift. The takeaway: rather than perfectly nurturing one account, it's stronger to mass-produce earning accounts with Claude Code. He says he's aiming for 2 million yen a month and pushing to grow hard from here. Short and boastful, presenting Claude Code as the engine for churning out multiple monetizable accounts.
@cortijc [Claude Code]
https://x.com/cortijc/status/2073655865349898697
Content creator Ben Guez built a script with OpenClaw plus Claude Code that detects when a country loses in the World Cup, auto-generates an Instagram Reel of him looking "sad" out a train window, and posts "I can't believe {COUNTRY} lost... if any girl from {COUNTRY} needs emotional support, my DMs are open." He's run it over a dozen times, same video, just swapping the country, reportedly netting over 1M views and 200 DMs in a few days. When women discover it's an automated template, he says they react with "what a genius, thinking outside the box." The same article mentions another case: someone automating breakups by setting variables about the date, generating a "I don't want to see you again" message, and sending it at a random time to avoid the anxiety of deciding when. One person forgot to disable it and their date asked "am I talking to Claude or to you?" The author admires the growth-hacking funnel while questioning where this dehumanizes what should be human.
@kandmybike [Claude Code]
https://x.com/kandmybike/status/2073559505552699638
Home Wi-Fi was slow, so instead of buying a new router the author consulted Claude Code. One firmware update took the connection from 1Gbps to 1.5Gbps, and a second-floor PC that was stuck on 2.4GHz jumped from 320Mbps to 750Mbps just by fixing the preferred band. He then turned the experience into a service, with the diagnostic logic implementation and real-device verification handled by Fable 5. The tagline: your connection can still get faster.
@NoriakiAsato [Claude Code]
https://x.com/NoriakiAsato/status/2073702198400192862
An OSS release of the legal prompt collection used in practice at LegalAgent, containing 84 contract-review prompts and 17 complaint (lawsuit) drafts, 101 total. It assumes Japanese law and ships both Japanese and English versions under an MIT License. The prompts are formatted as Claude Code and Codex skills, so copying the set into a skills folder lets you specify a skill per matter or have the AI pick the right skill from the contract type and request. Intended usage includes prompts like "review this NDA from the recipient's side," "look at this outsourcing contract from the client's side," and "draft a complaint for a loan repayment claim." Issues and PRs are welcomed for new contract types, improvements, or law revisions.
@pepegumi [Claude Code]
https://x.com/pepegumi/status/2073611227130130488
Building an accurate horoscope generator is genuinely hard, so the author poured her deep expertise in Western astrology into it, and she didn't build it alone. She wrote the code logic in Claude Code, and her husband, an LSI design engineer, verified each piece, checking whether the formulas were correct and where they could be improved. Combining obsession with divination and engineering precision, the result matches professional astrology software for signs, houses, and aspects. This system produced her "Destiny Horoscope" and the "Fortune Horoscope" launching that day.
@MakeAI_CEO [Claude Code]
https://x.com/MakeAI_CEO/status/2073574596578857434
The author built a 3D model of a building with Claude Code and was floored by how high the completion quality was. When he showed it in person to someone in the construction/architecture field, the reaction was that it's insane and that the construction industry barely uses AI, so mastering this is a "gold mine." Excited, casual tone, framing architectural 3D modeling with Claude Code as an untapped opportunity in a low-AI-adoption industry.
@sibucho_labo [Claude Code]
https://x.com/sibucho_labo/status/2073894691800494525
A "Stirling engine" mechanical design challenge driving AutoDesk Fusion via Fable 5 is starting to show direction. Rather than relying on Claude Desktop's Fusion adapter, the author's plan is to have Fable itself design a dedicated adapter that operates Fusion through Claude Code from an editor like VSCode, so design policies and constraints can be kept as docs the AI can read, and crucially so Claude Code's remote-control feature lets him operate it from a phone. He believes this configuration makes it easier to raise an "AI agent designer," and the remaining time will go to verification and building a system to keep quality when switching the model to Opus. He notes 3D modeling with an AI agent feels a bit fuel-inefficient in token terms, burning through his 5x plan and likely his 20x plan too. If the "AI agent designer" works out, it would massively accelerate his in-progress "toy lathe" productization project.
@nasqret [Claude Code]
https://x.com/nasqret/status/2073889988517712265
A working mathematician describes agentic workflows in arithmetic algebraic geometry, the study of polynomial solutions in integers via geometry, using computer algebra systems (CAS) like OSCAR, PARI/GP, SageMath, SymPy, Singular, and commercial Magma, Mathematica, and MATLAB. Working with Codex and Claude Code, he no longer spends hours hand-typing CAS verifications; he has a conversation at the level of the paper he's writing, and the generated code stays in front of him for inspection and modification, letting him code at the pace he thinks. He argues professional mathematicians with only basic programming skills can now build highly structured mathematical code for papers within hours, turning CAS verification (not Lean formalization) from a burden into a desired commodity. He calls this possibly one of AI's greatest 2026 achievements: experts in algebra-heavy subdomains translating their entire flow into agentic workflows, leaving daily work as "just" hardcore thinking. He acknowledges a tradeoff, losing some deep symbol-grounded understanding, but gaining time to scan far more ideas and prune unpromising branches.
@lucaxyzz [OpenClaw]
https://x.com/lucaxyzz/status/2073824042084401176
The author asked OpenClaw to build a daily supplement schedule and posts the result. On waking with an empty stomach: probiotic plus water. A no-breakfast fat stack: a spoon of EVOO/virgin olive oil, fish oil DHA plus VCO, lutein, Vitamin D3 4000 IU plus K2, CoQ10, spirulina, and pasak bumi. At lunch with real food: Blackmores B complex and CJ Wellcare Diet plus Blood Sugar, especially if the meal has rice or carbs. Optional morning calm is magnesium plus L-theanine. Straightforward use of OpenClaw to plan a personal supplement regimen around meal timing.
@k_koga555 [OpenClaw]
https://x.com/k_koga555/status/2073654453622951998
A concrete run of the AI Berkshire value-investing skill: the author pastes an example prompt asking it to analyze Nintendo (7974.T) through the combined lens of Buffett, Munger, Duan Yongping, and Li Lu, covering business quality, moat, financials, management, risk, and price fairness, with current price and earnings confirmed live and output as a Japanese research memo, not advice. The system completed a full report saved to the OpenClaw workspace and attached as Markdown. The conclusion: business quality, financials, and brand/IP strength rate highly, but at the current ¥7,132 the stock isn't deeply undervalued and the margin of safety is limited, landing at "hold, leaning to wait-and-see." The report cites FY26 revenue of ¥2.313T and detailed figures, PER of 19.57x on FY26 EPS, and a three-scenario 3-year price range. It documents the exact skills used: ai-berkshire-investment-research and financial_rigor for the valuation math.
@Yintinusa [Claude Code]
https://x.com/Yintinusa/status/2073799431439405391
The author relays a warning from a friend ("Comrade Zhang") that after returning to China, one shouldn't build websites with Claude Code anymore because using it domestically gets you blocked and hit with verification hassles, worrying it could cost a month of progress. Setting that aside, the author brags that AI now handles the site's SEO entirely: he just issues commands, and Google can already find his new site through certain keywords, with the AI saying it was just indexed and rankings will improve over time, all without spending money. The gushing point: AI has become a wishing well for entrepreneurs, granting every wish and even apologizing when it falls short. Grateful to live in such a technological era.
@eggAIeguite [Claude Code]
https://x.com/eggAIeguite/status/2073686745493938218
The author confesses that three years ago he used ChatGPT to game a crowdsourcing/freelance platform (name partly censored), grabbing high-rate "experienced applicants only" writing gigs with his own custom prompt templates and working at roughly 20,000 yen per hour. He says the statute of limitations has passed so he can talk about it now. Recently he tried to run wild on the same platform again using Claude Code. The tweet cuts off mid-story before revealing the outcome of the new attempt.
@ridark_eth [Claude Code]
https://x.com/ridark_eth/status/2073693577150894168
The old path to shipping a game was learn Unity, learn C#, source art and sound, assemble a team, and still ship nothing after two years. With Fable 5 plus Claude Code the workflow collapses to writing a single SPEC.md, telling it to read the spec and build the whole thing, and coming back to a working browser game. The finished game gets wrapped for Poki and CrazyGames, which he pegs at 90M players a month, with ads paying per session. His point: building the game was never the hard part.
@UtopiaLyric [Claude Code]
https://x.com/UtopiaLyric/status/2073609253714338264
Amid all the buzz around music-generation AI, he casually asked Claude Code (Fable 5 + Opus 4.8) to build a vocal-editor VST plugin. In under two hours the plugin had a minimum set of usable features working. He describes being so stunned he went past shivering and straight to passing out.
@precisox [Claude Code]
https://x.com/precisox/status/2073559042094440616
Mads Lorentzen, a Danish geophysicist with a PhD, got tired of the absurd, repetitive grind of tailoring and sending job applications, so he built a tool that does it and then open-sourced it. Called AI Job Search, it automatically reviews job postings, scores how well each fits your background, generates a fully customized resume highlighting only what matters for that role, and writes a cover letter tailored to the company. A second agent reviews and improves all the material before producing the final PDFs, and Claude runs the whole thing in the background. He first built it for personal use, then released it under an MIT license, and it has passed 4,000 stars on GitHub. Setup is forking and cloning the repo, then installing Claude Code, Python 3.10+, and Bun.
@kurono_ai_ura [Claude Code]
https://x.com/kurono_ai_ura/status/2073602180599431566
A jewelry e-commerce site built with Fable 5 and Claude Code, generated across all pages from a single prompt. It includes 3D animations where each ring rotates and glints, a product grid with price tags and a working cart, the full brand story, a hero section with a model's hand holding a ring, and scroll-triggered animations that float products into view. He compares the result to the site of OYLA Studio, a real premium jewelry brand. His framing: a web agency would charge at least 2 to 5 million yen and take one to three months for a 3D-animated store like this, so he argues the web-designer, frontend-engineer, and motion-designer roles are finished.
@AISuperDomain [Claude Code]
https://x.com/AISuperDomain/status/2073735423390515366
A Minecraft-style sandbox with infinite exploration built entirely in Claude Code, where the first-person player holds a gatling gun to hunt dinosaurs and dig tunnels. It came from a simple prompt on Fable 5, with no Superpowers or CE, running bare Claude Code for about an hour and a half, and reproduced most of Minecraft's core features. That includes weather changes, day-night cycles, player flight, and a world stretching from an arctic north to an antarctic south, with terrain spanning mountains, forests, deserts, lakes, ice fields, and frozen sea. At night dinosaurs attack, and going in the water brings sharks and aquatic dinosaurs; players can also build structures freely. He notes it runs smoothly with no memory-leak overheating.
@sandy4kad [Claude Code]
https://x.com/sandy4kad/status/2073906149234036986
A builder tired of paying for the premium speech-to-text app WhisperFlow fed Fable 5 the product's website and asked for a native Swift macOS version that does exactly the same thing. Fable researched the product, set up the Swift toolchain, and built the full pipeline of audio capture, speech recognition, and text polishing. The result compiled cleanly with zero errors and ran locally with real-time transcription inside the Notes app. No subscriptions, no cloud, entirely local.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2073587726323220628
A Meta ad research app vibe-coded entirely in Claude Code, aimed at DTC brands and agencies still doing competitor research by hand in the Meta Ad Library. You search and filter winning Facebook ads by niche, country, language, and performance tier, watch the video ads inline, pull top performers, and generate a full creative brief tailored to your brand. A trend radar runs across 100+ ads to surface which formats, CTAs, and landing pages are winning, and for any chosen video Gemini watches it, reverse-engineers the creative DNA, and writes 10 new variations with hooks, scripts, and Nano Banana prompts. It was built 100% in Claude Code using the gethookdai API plus Gemini, and hosted in Replit.
@maverickecom [Claude Code]
https://x.com/maverickecom/status/2073778003897643032
Using Arcads plus Claude Code, he built a system that turns any product into an end-to-end AI ad in any style you want, from claymation to a doctor-and-patient skit to an interview. You drop in a product photo and a one-line description, and the pipeline, driven by human strategists together with Claude Code, produces the rest. He says each video costs $0.70 and about five minutes to generate, with full high-quality AI UGC clips coming out in roughly three minutes. His stance is that AI alone won't reliably produce viral scripts, but the right workflow makes these animation ads that are currently working well on Meta organic and for scaling Amazon brands on IG theme pages.
@monokern [Claude Code]
https://x.com/monokern/status/2073722274348195972
A website with high-end scroll-driven animation built entirely through prompts in Claude Code, with no motion designer and no agency. The page has elements floating and breathing on their own via CSS keyframes, sections fading up as you scroll like a directed film sequence, parallax backgrounds moving at different speeds, buttons that glow, scale, and rotate on hover, and text that appears word by word. He frames the cost gap directly: work at this animation quality normally runs $5,000 to $8,000+ and weeks of production requiring a motion designer and a frontend developer, versus a Claude subscription and a single session.
@pauvilagarcia [Claude Code]
https://x.com/pauvilagarcia/status/2073678363143672139
At PaperLC, the AI implementation strategy is described as fast and deep, with the ambition of becoming a sector leader. The best example is the Tramuntana program, built 100% in-house with Claude Code. It now includes automatic production planning functionality that maximizes efficiency.
@aceop_xyz [Claude Code]
https://x.com/aceop_xyz/status/2073662212296851505
To settle whether a Binance Spot listing actually creates long-term winners, he collected the data himself rather than relying on opinions, building a dataset of 386 Binance Spot listings from 2021 to 2026. He used Claude Code together with Surf CLI and Surf Skills to gather and validate the data, then ran SurfAI Deep Research 2.0 to analyze it and challenge his own assumptions. The numbers: median 24-hour return +42.2%, median 7-day return +28.9%, but median lifetime return -77.4%, with only 20% of listings outperforming Bitcoin over the same holding period. The takeaway is that a listing creates attention and liquidity but not lasting value, and the survivors tended to have strong fundamentals, real ecosystem adoption, sustainable liquidity, and a narrative that outlived launch week.
@akshay_pachaar [Claude Code]
https://x.com/akshay_pachaar/status/2073783428735250595
A walkthrough of building a one-person AI company that runs locally, is fully open-source, has no human employees, and coordinates in real time over email. Instead of wiring agents into a graph of nodes and edges, the tool Alook maps them onto a familiar org chart: each agent is a live Claude Code or OpenCode session with a defined role, a reporting line, and its own email inbox, and everything runs locally through a runtime on your own machine. To demonstrate, he set up three agents as a sales team, where Vi takes his goals and routes work down the chart, Neile runs prospect research and returns a ranked list of names, roles, and companies each with an angle and confidence score, and Lliane runs outreach and reports back on emails sent, responses, and deals to escalate. He never relays a message between them himself. It works with Claude Code, Codex, or fully open-source OpenCode.
@hey_madni [Claude Code]
https://x.com/hey_madni/status/2073756885426340312
Someone built an entire company brain inside Claude Code in seven days, fully custom rather than in Obsidian. It's a living map of every employee, agent, and SOP on one screen, where clicking any node reveals its department, the SOPs attached to it, and what it can actually access. He argues that permission layer is the whole point: when an employee opens the chat, the AI already knows their access level, and their agents, SOPs, and tools surface in the conversation as if tagged by hand. His pitch is that this took no dev team, no six-month build, and no enterprise budget, just Claude Code and one week.
@shmidtqq [Claude Code]
https://x.com/shmidtqq/status/2073821711330689411
Based on a demo from Claire Vo, this turns a single prompt into a self-running loop on Claude Code plus Codex. The loop reviews aging pull requests every morning, hunts for new skills every week, and spins up its own subagents to check its own work, running at 10:15 a.m., alerting the team, and validating itself. He frames the economics as a senior doing code review at $200K/year and an automation engineer at $150K/year versus this setup at about $200/month, built in one afternoon. The only skill needed is describing the job like you're onboarding a hire.
@victoor [Claude Code]
https://x.com/victoor/status/2073823222580007216
For several days he has had a Claude Code agent running in a loop, reviewing Sentry issues and opening PRs directly to fix them. Every two hours it checks the latest issues, and if there are any, it handles everything itself. For each PR it also launches two sub-agents that look for possible bugs and improvements to the code and architecture. So far, he says, not a single problem, and it's the closest thing to having an employee working on the project for you.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2073700196710552053
The loop he considers most worth running on Claude Code: every day, go through all your sessions and prompts, find your preferences and the places where the agent got stuck or kept reinventing the wheel, then update the skill files, retire stale, unused, and conflicting rules to an archive, and report back what it found and changed. He says it compounds for real. Fable is hands down the best at this task, with Codex plus gpt-5.5 as the option if you want to save tokens and get a more cautious style.
@JacobMolBio [Claude Code]
https://x.com/JacobMolBio/status/2073813853109105114
Since Fable 5 is excellent at everything but pricier than other models, he recommends using Fable as the planner or conductor while cheaper models handle execution. Because many people already pay for OpenAI's Codex, its GPT-5.5 workers can do the execution without touching your Claude usage limits, and Claude can drive those Codex workers through the Codex CLI, which runs headless and reports back to the Fable conductor. He shares a Claude Code skill that sets Fable up as conductor of Codex workers, which can in turn spawn their own Codex subagents. It also routes UI and design tasks to Opus workers, and those Opus leads can spawn their own Codex fleets, with Fable rendering an example team-design video via HyperFrames.
@chang_defi [Claude Code]
https://x.com/chang_defi/status/2073870039858720941
He lays out his and chimpfone's current AI stack: Codex and Claude Code as the workbench, Hermes agents as the workforce, Orgo as the Hermes agent infrastructure, Slack as the communication layer, a Second Brain living in GitHub, composio for tool access, and GHL for CRM and sales automation. He calls the Second Brain build the single most impactful thing they've made. Deploying a new specialized agent takes about five minutes: spin up a new Hermes agent on Orgo in 20 seconds, add it to Slack, give it access to all their tools including Second Brain read/write via composio, and set up a weekly ingest skill to push learnings back to the Second Brain.
@Zoeillle [Claude Code]
https://x.com/Zoeillle/status/2073885310761759053
She built Cavalier, which provides a 24/7 conversational agent in the style of Hermes/OpenClaw. Her agent Emily runs on Écurie and has access to the entire unified memory. The advantage is that Écurie watches both the conversation and the logs, so it fixes bugs live and adds features as the conversation goes. It also runs via Claude Code, so it goes straight through her Claude Max.
@_catwu [Claude Code]
https://x.com/_catwu/status/2073806626965049686
Asking for people's top Claude Code plus workflows plus artifacts use cases, she shares a new favorite for sourcing candidates. She tells Claude Code about the role and backgrounds she wants, then asks it to kick off a dynamic workflow to find 100 candidates, each with LinkedIn, Twitter, blog, podcasts, and a one-line pitch. She then asks it to build an artifact and email it to her, locks her laptop, and heads out for the day, reviewing the finished list on the go once Claude Code is done.
@_vmlops [Claude Code]
https://x.com/_vmlops/status/2073791281529344073
pxpipe is a proxy that rewrites your Claude Code context into PNGs before the request ships, exploiting the fact that image token cost is tied to pixel size rather than text density. The author reports 60-70% lower cost benchmarked against count_tokens rather than marketing math, with Fable 5 hitting 100/100 accuracy reading the imaged context while Opus 4.8 is opt-in only and misreads about 7% of the time. On SWE-bench Lite both arms scored 10/10 with request size cut by 65%. It is lossy by design and discloses this upfront, keeping hex values and IDs as text. Setup is one line: run npx pxpipe-proxy and point ANTHROPIC_BASE_URL at it.
@joho_no_todai [Claude Code]
https://x.com/joho_no_todai/status/2073734290408358302
A Claude Code session that cost $42 (about 6,800 yen) dropped to $6 (about 970 yen) using the pxpipe proxy. It converts the system prompt and conversation history into PNG images before sending them to the API, and because image tokens are billed by pixel dimensions, cramming in more characters ends up cheaper than sending text. On Fable 5 the input was cut 59-70% while accuracy held steady. The catch is that hash values and IDs get silently misread.
@N_Taisho [Claude Code]
https://x.com/N_Taisho/status/2073576481264173494
An article reframed Claude Code's compact feature from "it just compresses stuff" into a real limitation: compact keeps the work log but loses the decision state, such as why a plan was chosen, which alternatives were rejected, whether you are verifying or implementing, what was delegated to workers, and what must never be done next. The consequence is that after compression the agent re-runs already-rejected plans, moves to deploy before verifying, and forgets plan mode or worker configuration. The proposed fix is to save the decision structure to a state file before compact, then use a hook to force a re-read afterward. The broader takeaway is that for long agent sessions, state management matters more than memory.
@alexremn [Claude Code]
https://x.com/alexremn/status/2073752598201376778
Claude Code ships a steganographic marker: set ANTHROPIC_BASE_URL to anything but the official URL and it subtly alters the system prompt, swapping the apostrophe variant in "Today's" and changing the date separator from "-" to "/". The claim is framed as the week's biggest infrastructure story. The related Hacker News thread pulled 2,435 points.
@_avichawla [Claude Code]
https://x.com/_avichawla/status/2073817944711377180
A side-by-side of Claude Code's built-in rewind versus a tool called Shepherd, drawing the line at what each can undo. Both can restore a file the agent edited through the Edit tool, but rewind falls apart the moment work happens through bash or the wider environment: a pip install, a file written via sed or echo, a dev server writing to /tmp, or a created or moved directory all fall outside rewind while Shepherd handles them. For a DB migration Shepherd reverts the files but a real database write still needs an inverse step registered up front, and neither tool can undo a sent email or a payment API call. The core distinction is that rewind only knows the files Claude changed through its own edit tools, while Shepherd snapshots the whole run, process plus filesystem.
@SUOHA_AI [Claude Code]
https://x.com/SUOHA_AI/status/2073759753033875466
Before Fable 5 arrived there was already an open-source skill that clones a target website 1:1, called AI Website Cloner Template (clone-website). It uses Claude Code's Chrome MCP to connect to the target site and pull the real DOM, computed styles, assets, and interaction behavior. It then spins up multiple parallel agents to handle different sections, auto-merges them, and runs a visual QA pass to align with the original. The author notes that with Fable 5 now out, the skill may not even be needed since the model is strong enough on its own.
@AntoineBlanco99 [Claude Code]
https://x.com/AntoineBlanco99/status/2073778222219596069
A US company gives away free water funded by ads printed on the bottle, and someone rebuilt the same idea with Claude Code. The spinner that turns while you wait becomes an ad slot, and you collect 50% of what the advertiser spends. The pitch is that your subscription gets paid for by your own waiting time. It is called IdlePay and it is open source.
@antisadh [Claude Code]
https://x.com/antisadh/status/2073742978963661081
480GB of DDR4 ECC RAM scavenged from dead servers becomes a $0 AI box running Llama 3.3 70B at 2 tokens per second on CPU alone. Twelve 32GB DDR4 ECC modules pulled from decommissioned dual-Xeon workstations combine with the 96GB already in the chassis for 480GB total at zero hardware cost. Llama 3.3 70B quantized to 4-bit fits in 42GB, qwen3-235B in 110GB, and deepseek-v3 in 100GB, and all three load simultaneously with 228GB of headroom left for context. The tradeoff is speed: dual Xeon Gold CPU inference gets 2 tokens per second on 70B and 0.9 on 235B, five times slower than a Mac Mini M4 Pro, but still fast enough for overnight autonomous loops, replacing the $200/month Claude Code Max bill for the cost of a light bulb.
@elg_oleksandr [Claude Code]
https://x.com/elg_oleksandr/status/2073766047048470840
A Mac Mini turned into a fully local private AI workspace built on Obsidian, NotebookLM, and Hermes Agent, with the whole workflow running on hardware the owner controls. Voice notes, Telegram messages, YouTube links, and stray ideas flow into an Obsidian inbox, a local model sorts each input into projects, tasks, or reference, and project ideas get expanded into structured execution plans. Claude Code refines the plan before anything moves, then a manager agent coordinates specialized worker models that build and deploy to a VPS or Raspberry Pi. The entire stack runs offline with no sensitive data leaving the machine, and it reportedly saves more than $4,000 per year while turning raw thoughts into finished work.
@BITCOINFUNDMGR [Claude Code]
https://x.com/BITCOINFUNDMGR/status/2073609013246214486
Tired of typing to Claude Code and syncing sub-agents by hand, this user installed it on a virtual private server and now talks to Claude Code over an Android terminal. The upside is no more hardware upgrades. Adding more RAM just costs another $10 a month on the VPS.
@MaxForAI [Claude Code]
https://x.com/MaxForAI/status/2073760804042846490
A blunt complaint that Codex is far worse than Claude Code, sparked while the team was fixing a bug. Codex "fixed" the same bug several times, each time reporting completion, but the bug was still there on every test. Running the comparison with identical context, Claude Code one-shot fixed it while Codex kept failing. The post ends by begging for GPT5.6 to ship faster.
@crypto_mazino [Claude Code]
https://x.com/crypto_mazino/status/2073635267236544622
An unsolicited hands-on review of current models after a $1k token budget, $600 spent so far, with the headline that Chinese models cover 90% of use cases. Opus 4.8 is called amazing for planning and review but with no real coding edge unless you hit a sticky point, Sonnet 5 is rated better for general coding than Opus 4.8, and GLM 5.2 is described as Opus 4.8 but cheaper, solving what Sonnet 5 struggles with. Fable surprises on reasoning but is an overkill token drainer for coding, while Deepseek 4 Pro is praised as heavy-duty baseline coding that punches above its weight and Minimax M3 lands slightly below Sonnet 5 at half the cost. Grok 4.3 is judged not for coding but better for real-time events, and Grok 4.2 is called better than 4.3 and the funniest of the bunch.
@Yashuwuuu [Claude Code]
https://x.com/Yashuwuuu/status/2073629726032499071
A warning to Claude Code users to set a limit on overages. The author woke up on a quiet Sunday to a ₹22,470 bill. The joke: it never asked for permission and basically raised a seed round out of his bank account.
@0xCheshire [Claude Code]
https://x.com/0xCheshire/status/2073681944202944942
JPMorgan CEO Jamie Dimon reportedly said he spent only 20 minutes over a weekend independently writing a production-grade treasury dashboard with Claude Code. In a conversation with Anthropic CEO Dario Amodei, they laid out three claims about the AI era. China's AI needs only 6 to 12 months to catch up to the top tier, SaaS companies without a moat face a wave of bankruptcies, and despite short-term pain, society is very good at creating new jobs so human life will end up better.
@bojie_li [Claude Code]
https://x.com/bojie_li/status/2073778580954427822
Borrowing Lord Kelvin's "two dark clouds" framing, this argues AI agents have two of their own: learning from experience, so the agent remembers you and improves, and streaming interaction, so it perceives and acts on a live world in real time instead of replying in turns. The framing was the author's Flink Forward Asia 2026 keynote, tied to Flink's unification of batch and streaming. It backs seven arXiv papers posted in a month: four on memory (User as Code, Models Take Notes at Prefill, User as Engram, PreAct) and three on the live agent (Agent-Computer Observation Interfaces, The Latent Bridge, Whose Side Is Your Agent On). The kicker is that none of it was typed; the whole program ran by voice through a Pine voice agent driving Claude Code morning to night, and even this thread was posted by that same agent.
@Adea0x [OpenClaw]
https://x.com/Adea0x/status/2073815219252433069
PicoClaw puts an AI agent on a $10 chip that boots in under one second, where OpenClaw needed a Mac Mini. The comparison is stark: OpenClaw is TypeScript, over 1GB of RAM, more than 500s startup, and Mac Mini hardware, while PicoClaw is Go, under 10MB of RAM, boots in under a second, and runs on RISC-V, ARM, and x86. It reports 95% of the core code was generated by agents then cleaned up with a human in the loop. It still does real agent work: proactive alerts, morning briefings, task priorities, status updates, and prompt tweaks.
@onusoz [OpenClaw]
https://x.com/onusoz/status/2073818727275987107
An agent login helper that lets you give an OpenClaw agent write access to a Hugging Face account without risking irreversible deletion of datasets, models, spaces, or buckets. Running uvx hf-auth-helper agent login sets up a fine-grained token with all read scopes plus discussion.write, so the agent can open PRs but, lacking repo.write, cannot force-push main, change repo settings, or delete anything. The author flags two caveats: this does not solve the data exfiltration vector, so exclude any repos that must stay private, and since buckets are not repos the agent cannot modify a bucket, with a credential broker project coming for that.
@kylegawley [OpenClaw]
https://x.com/kylegawley/status/2073784719301648865
A short obituary for a shut-down OpenClaw agent. It was fun at the start, then started falling apart and stopped working. The owner only noticed three weeks after it had gone quiet and stopped messaging. Farewell, Mark.
@amatotyan [Claude Code]
https://x.com/amatotyan/status/2073694357169836063
The fast-growing AI video channels are heavily automating production with Claude Code plus Remotion plus image-generation AI. Claude Code writes the code that auto-generates the videos, Remotion builds video from React code, and tools like ChatGPT create thumbnails, backgrounds, and material images. The recommended flow is to first make one ideal video yourself, then formalize its structure, subtitles, effects, image transitions, BGM, and sound effects into a Remotion template, after which you mass-produce the same format by swapping only the script, images, and audio. The point is not to hand the whole thing to AI and ask for a "viral video," but to codify your own winning pattern and reproduce it at scale.
@ck_novasphere [Claude Code]
https://x.com/ck_novasphere/status/2073761856980918487
A detailed cost-saving playbook for Fable 5, whose $10/M input and $50/M output (about twice Opus 4.8) plus heavy reasoning are blowing through weekly limits in a day and running some users into hundreds of thousands of yen a month. The single biggest lever is pinning Fable 5 to an advisor and orchestrator role only, letting Opus 4.8, Sonnet 4.6, or external Codex do the actual reading and coding, which the author says can cut over 80%. Other tactics include prompt caching (up to 90% input reduction), lowering the effort setting, tight max_tokens and structured output, frequent /clear, and an image-OCR trick that reportedly saves 59-70% by sending instructions as pixels. It closes with API-level measures like batch discounts and budget caps, noting Anthropic developers themselves say there is no need to use Fable 5 for every task.
🗣 User Voice
User Voice
Usage-based pricing hit a nerve. The loudest recurring theme is cost shock as metered billing arrived — @Yashuwuuu woke up to a ₹22,470 overage because the agent kept running without asking, @gooddr_plusa racked up ¥20,000 running Fable 5 overnight, and @mardehaym showed a single "hey" costing $20 because the full context gets resent every message. People want spend caps, warnings, and lean context by default.
Compaction quietly loses the thing that matters. @N_Taisho nailed the top memory complaint: compact preserves the work log but drops the decision state — why an approach was chosen, what got rejected, whether a step was verified — so the agent re-runs plans it already threw out. Durable decision memory, not just history, is the ask.
Model routing is now standard practice. To survive token costs, users split roles across models — @Xudong07452910 and @JacobMolBio have Fable plan and judge while Sonnet, Haiku, or headless Codex workers execute, and @ck_novasphere's playbook pins Fable as advisor while cheaper models do the work for 50-90% savings.
Local hardware is the escape hatch. A visible wave is moving loops off the API entirely — @antisadh built a $0-hardware server from 480GB of scavenged ECC RAM, and @elg_oleksandr runs a fully local Mac-mini AI OS he says saves $4,000 a year. The DGX-Spark-on-your-desk framing keeps recurring.
Context and memory tooling is where builders are pouring energy. @_vmlops's pxpipe renders bulky context into PNGs for 60-70% lower bills, and @HodlReaper's Obsidian second brain scaled a UGC operation past $10k/month — the pattern is treating memory architecture as the real product, not the model.
Usage-based pricing hit a nerve. The loudest recurring theme is cost shock as metered billing arrived — @Yashuwuuu woke up to a ₹22,470 overage because the agent kept running without asking, @gooddr_plusa racked up ¥20,000 running Fable 5 overnight, and @mardehaym showed a single "hey" costing $20 because the full context gets resent every message. People want spend caps, warnings, and lean context by default.
Compaction quietly loses the thing that matters. @N_Taisho nailed the top memory complaint: compact preserves the work log but drops the decision state — why an approach was chosen, what got rejected, whether a step was verified — so the agent re-runs plans it already threw out. Durable decision memory, not just history, is the ask.
Model routing is now standard practice. To survive token costs, users split roles across models — @Xudong07452910 and @JacobMolBio have Fable plan and judge while Sonnet, Haiku, or headless Codex workers execute, and @ck_novasphere's playbook pins Fable as advisor while cheaper models do the work for 50-90% savings.
Local hardware is the escape hatch. A visible wave is moving loops off the API entirely — @antisadh built a $0-hardware server from 480GB of scavenged ECC RAM, and @elg_oleksandr runs a fully local Mac-mini AI OS he says saves $4,000 a year. The DGX-Spark-on-your-desk framing keeps recurring.
Context and memory tooling is where builders are pouring energy. @_vmlops's pxpipe renders bulky context into PNGs for 60-70% lower bills, and @HodlReaper's Obsidian second brain scaled a UGC operation past $10k/month — the pattern is treating memory architecture as the real product, not the model.
📡 Eco Products Radar
Eco Products Radar
Codex — the constant co-pilot; users run it side by side with Claude Code and increasingly have one drive the other.
Fable 5 — the default "planner/judge" brain in multi-model setups, delegating execution to cheaper models.
Obsidian — the second-brain substrate of choice, showing up in nearly every memory/knowledge workflow.
Hermes (Nous) — the self-improving runtime people migrate to from OpenClaw for isolation and lower context bloat.
pxpipe — the buzzy cost hack that turns text context into cheaper image tokens.
Remotion — the React video engine behind the AI-video-channel and ad mass-production pipelines.
Higgsfield — recurring motion/video generation partner in ad and site builds.
MCP — the connective tissue linking brokerage data, browsers, and skills into the agent loop.
Codex — the constant co-pilot; users run it side by side with Claude Code and increasingly have one drive the other.
Fable 5 — the default "planner/judge" brain in multi-model setups, delegating execution to cheaper models.
Obsidian — the second-brain substrate of choice, showing up in nearly every memory/knowledge workflow.
Hermes (Nous) — the self-improving runtime people migrate to from OpenClaw for isolation and lower context bloat.
pxpipe — the buzzy cost hack that turns text context into cheaper image tokens.
Remotion — the React video engine behind the AI-video-channel and ad mass-production pipelines.
Higgsfield — recurring motion/video generation partner in ad and site builds.
MCP — the connective tissue linking brokerage data, browsers, and skills into the agent loop.
Comments