Super User Daily: August 18, 2026
The most instructive thing that happened yesterday was Anthropic going dark. Sunday evening the auth system, the API, Claude Code and Cowork all went red at once, the fourth multi-hour outage of the month, and the loudest posts of the day were people counting their dependency in dollars. Underneath the outage noise, the actual work kept moving in one direction: independent verification. A Japanese firm runs six named subagents to scan 490 subsidy programs a day, and the only irreplaceable one is the verifier that never sees the other agents' answers. A CLAUDE.md going around now ships a harsh critic that rejects by default. One user's nightly eval of Opus versus GPT came down to which model stops re-litigating its own conclusions. And the OpenClaw exodus continued in public: early adopters who bought hardware for it in February formally uninstalled it, while the personal-agent crowd moved to Grok Bot and told everyone why. The best cases were barely about code at all: a patent law office, a wedding-video startup with $4.18M in five months, window blinds, train-fare expense reports, and a fortune teller.
@tavst_yu_suk_o [Claude Code]
https://x.com/tavst_yu_suk_o/status/2088840022795722840
A Japanese company publishes daily subsidy-program updates, and the research desk is six named Claude Code subagents. On the day recorded, they scanned roughly 490 new government programs, shortlisted 5, adopted 4, in 22 minutes of agent runtime and 181 tool calls, with the human doing only approval and posting. The load-bearing agent is Ken, an independent verifier running on a different model that re-derives all 14 required fields from the source PDFs without ever seeing the other agents' reports. That day Ken found material errors in all three programs it checked: a 2-billion-yen investment floor, only about three grants to be awarded, generic AI excluded. The owner's principle is the transferable part: never show a checker the previous answer, because an AI shown an answer starts confirming instead of verifying.
@shmidtqq [Claude Code]
https://x.com/shmidtqq/status/2089087738330948085
A six-person Chinese startup turns ordinary phone footage into flythrough-able volumetric scenes with 4D Gaussian splatting: $4,180,000 in five months at $29 a month, off an initial $9,400 in rented GPUs. The entire training pipeline was built in Claude Code in 11 days against an original plan of half a year. The customers are the punchline: they pitched film studios, but 70% of revenue comes from wedding videographers in Chengdu who need a sweeping camera move around the couple, and Taobao sneaker sellers who need 360-degree product spins without a studio. Their biggest client, a wedding salon chain, pays $18,000 a month and asks for exactly one fix: stop the groom's glasses from doubling during the fly-around.
@entame_benrishi [Claude Code]
https://x.com/entame_benrishi/status/2088790818962481447
A Japanese patent attorney office reports what agents did to its cost structure since introducing them in March: case throughput more than doubled, administrative cost now under 10% of revenue, and the firm is restructuring toward five attorneys per single admin staffer. The recruiting page announcing all this was itself built with Claude Code wired directly to their Xserver hosting, so the agent edits and publishes files on the live server. First page took about three hours including setup; the next ones take about an hour.
@levelsio [OpenClaw]
https://x.com/levelsio/status/2089082507060793581
Pieter Levels runs his whole house through Alexa into Home Assistant, with Claude wired in to make devices behave. Two window shutters installed after the house purchase ran on a different system with their own RF remote, so every day meant putting all blinds down and then hunting for the extra remote. He asked the AI, it told him to buy a Broadlink RF bridge, he connected it to Home Assistant, and Claude set it up and learned the radio signal from a few button presses. The orphan shutters now live on the same ALL BLINDS DOWN button as everything else.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2088988766220517771
Claude Code with Opus 5 generated a full corporate planning workbook: 19 tabs, 36 months, a three-year P&L, with clean formulas, a data layer, and consistent cell references. His estimate for doing it by hand is a week. The model note is worth keeping: he finds Opus 5 better than Fable 5 specifically for Excel work, another data point that the frontier model is not automatically the right tool per task.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2089020242303557771
An outbound agency priced every step of a campaign instead of running everything on a frontier model. ICP qualification is a plain Python script at $0, suppression and send-window preflight are Claude Code hooks at $0, signal feeds run on Haiku, account research and per-send QA on Sonnet, and only first-draft copy and strategy earn Opus or Fable, because a weak draft there travels into every account. A person looks at about 50 accounts at the front and 50 sends at the back of a 500-account campaign; the other 400 flow through untouched. His routing rule generalizes: price a step by how open-ended it is and how far a bad output travels before someone catches it.
@ZhenZhu200 [Claude Code]
https://x.com/ZhenZhu200/status/2088937046983487758
A product engineer running three projects rebalanced from two Claude subscriptions plus one Codex to the reverse, and his reasoning is a controlled experiment. With context, skills, and a scored rubric all provided, Opus 5 pushed one instance from 60 to 90 points overnight; GPT 5.6 pushed four, because Opus kept convincing itself it had found bugs, tearing down finished work and redoing it while the clock burned. His division of labor: Claude still wins brainstorming and strategy, but that is maybe 10% of a day, and for long-horizon execution he wants the model that never stops believing in the goal.
@gregce10 [Claude Code]
https://x.com/gregce10/status/2089098490228195743
After 18 months of hosting his Claude Code and Codex shells inside Cursor and losing every session whenever it leaked memory and crashed, he built and open-sourced Tortie: a calm session multiplexer for Mac where every agent session is named, durable, backed by a private tmux server, and recoverable, with a VSCode-grade explorer and search around it. The motivation section is the honest part: the super-apps from OpenAI, Anthropic and Cursor all want to be the agent window, and he wanted something boring that hosts whichever CLI he feels like paying for that month.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2089001978781368374
He open-sourced the workflow behind $2M-grade AI video productions as seven skills that run the pipeline end to end for Seedance 2.5: setup, studio scaffolding, scene-by-scene breakdown with a 22-field card per shot, a locked reference board with ban lists, exhaustive asset passports, a stress test that only marks an asset locked at 10-out-of-10 repeatability, and a shot-prompt skill that refuses to run until everything in frame is locked. Works in Claude Code, Codex, Hermes or any harness. The design insight is discipline as code: the pipeline hard-blocks generation until consistency is proven.
@Stefan_3D_AI [Claude Code]
https://x.com/Stefan_3D_AI/status/2089037153888334312
A complete Souls-like game level, a boss and a player character in 72 hours, with a disclaimer he wrote himself: no one-shot, nobody typed a wish and walked away. Every asset was AI-generated, then assembled, textured, lit and integrated by hand in Unreal Engine 5 on a Souls template wired up with Claude Code and the UE MCP. It reads as the realistic template for AI game work right now: the model multiplies someone who already knows 3D and gamedev, it does not replace the knowing.
@Wakuwakukucx [Claude Code]
https://x.com/Wakuwakukucx/status/2088804332783435958
The most-viewed Claude Code post of the day, 11 million impressions: a fictional TV program's opening sequence generated with Claude Code driving Remotion. The reason this combination keeps producing broadcast-looking output is structural, and a reply put it best: Remotion turns video into React code, so a coding agent is a motion designer with deterministic output. He says anything short of high-end 3D animation is now makeable this way.
@bourneliu66 [Claude Code]
https://x.com/bourneliu66/status/2089064061078298842
New ritual: every time Claude Code finishes writing code, five colleagues hold a review meeting. The panel is GPT-5.6 Sol Max, ChatGPT Pro, Claude Opus 5, GLM 5.3 and Grok 4.6 xhigh. Cross-vendor review keeps showing up because different models carry different blind spots, and the writer model passing its own work is exactly the failure everyone has stopped trusting.
@jbarbier [Claude Code]
https://x.com/jbarbier/status/2089004502632325545
Two new rules landed in his open-source CLAUDE.md. First, fan-out plus harsh critic: every feature gets 2-3 competing variants from parallel builder subagents, a separate critic judges blind side by side and rejects by default, bug fixes get an attacker trying to re-break them, docs get a cold reader who fails them at the first confusion. Second, self-rating: after each task the agent scores itself 1-10 and answers whether it is proud of the work; if not, it names the gap and loops, and inflating the score to exit is explicitly banned. Free to copy, works across Claude Code, Codex, Cursor and Gemini.
@suna_gaku [Claude Code]
https://x.com/suna_gaku/status/2088888260638777749
Five months of accumulated CLAUDE.md, skills and rules were handed back to Claude itself for an audit against Opus 5, and not just a read-through: it checked usage history for whether each skill actually gets invoked, whether its outputs get read downstream, and whether current built-in features replace the mechanism entirely. Result: 40 skills cut to 6, always-loaded config down 72%. The principle generalizes to every setup thread this week: instructions written for last year's model are constraints on this year's.
@thekuchh [Claude Code]
https://x.com/thekuchh/status/2088950922978496829
He canceled a $200-a-month plan after a skill called caveman cut his output tokens by roughly 65-75%. The diagnosis: Claude writes essays to say what one tight line could say, and you pay for every hedge and every paragraph of nothing. The skill forces short answers by default while keeping code and errors exact, with lite, full and ultra intensities, and pairs with a local proxy that shrinks what the agent reads before it hits the model. Output tokens are the expensive ones; this is one of the simplest cost levers around.
@kemu_aii [Claude Code]
https://x.com/kemu_aii/status/2089094930425934037
Her token economics: people who use Claude Code with Fable 5 doing everything are wasting it. Fable holds only judgment and design as the commander; research goes to Sonnet, implementation to Codex, mechanical work to Haiku. Since switching, long sessions stopped drifting and Fable consumption dropped visibly. The same tiered-routing shape showed up in three unrelated posts yesterday, which is usually the sign a practice is hardening into a norm.
@mkwizrd [Claude Code]
https://x.com/mkwizrd/status/2089013712959189120
Day 52 of building an ecommerce brand: he spent the day talking to Claude Code through AquaVoice, training it on ad copy, and by evening it produced five drafts of which two were ready to launch in the ad account unedited. His plan is to keep training it until iterating new ad formats is fast by default, while the brand pushes toward $100k/day spend. Voice-driven agent sessions keep showing up in non-engineering work, where the bottleneck is describing taste, not typing code.
@R1H0_0 [Claude Code]
https://x.com/R1H0_0/status/2088983658099249511
Small, complete, and the kind of thing that spreads through an office: she connected Google Calendar to Claude Code and automated her train-fare expense claims. The calendar already knows where you went and when; the agent turns that into the expense filing nobody enjoys doing.
@pennyhunter_sub [Claude Code]
https://x.com/pennyhunter_sub/status/2089135340598817029
He handed Opus 5 a product video and the existing paper instruction manual and asked for a manual video with narration matched to the footage. It came back in a day, with the playback speed adjusted to 1.2x where the pacing needed it and self-composed background music. His framing: the weekend-course level of video editing skill is now something you subscribe to rather than learn.
@bkdgiffug [Claude Code]
https://x.com/bkdgiffug/status/2088948636012978457
Someone packaged Chinese four-pillar astrology into a Claude Code plugin, bazi-skill. Type three characters, it asks name and birth time, lays out the chart, reads the five elements and structure, and returns concrete advice on career, relationships and health, with a real body of traditional source material fed in rather than improvisation. Install is a clone into the skills directory. A thousand-year-old trade, absorbed as a plugin.
@Lummox_eth [Claude Code]
https://x.com/Lummox_eth/status/2088951287408714127
A student trained Claude Opus to play Flappy Bird and pushed it past 1,400 points. It started with dozens of birds mistiming and clipping pipes within seconds, then runs stretched 100, 500, 1,000-plus, until it was clearing pipes indefinitely with Claude Code making every move. Useless economically, perfect as a visible demonstration of what an iteration loop does to a task with instant feedback.
@LawrenceW_Zen [Claude Code]
https://x.com/LawrenceW_Zen/status/2088861693292249570
Full walkthrough of android-remote-control-mcp: the phone itself runs the MCP server, so an agent operates any Android app like a human, no root, no cable, no computer running ADB relay. It reads the screen, taps buttons, swipes, types, opens apps, takes photos, reads notifications, 57 tools total; a full UI-tree read is about 4,000 characters, roughly a thousand tokens, far leaner than XML dumps. He documented both the real-phone path and the emulator path, including the Android 13+ restricted-settings trap for sideloaded accessibility services.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2089027423774048326
By default agents cannot read social media; login walls and bot blocks stop them cold. His /social-fetch skill takes any URL from X, LinkedIn, Instagram, TikTok, Reddit or HN and returns structured data, author, text, engagement, replies, trying free APIs first, escalating to browser automation when walls go up, and falling back to Wayback for dead posts. Part of an 18-skill free open-source pack for founders and operators, installable as a Claude Code plugin marketplace.
@Huahuazo [Claude Code]
https://x.com/Huahuazo/status/2088838409947074875
Hyperresearch is a Claude Code plugin that runs a 16-step adversarial research pipeline: it decomposes your question, fans out across 250+ sources in parallel, then four critics attack the draft simultaneously, and they are only allowed local surgical edits, never a full rewrite, before it emits a fact-checked report with complete citations. Everything fetched lands in a permanent searchable knowledge base so the next run reuses it. Three sizes: 30-40 minutes for daily questions, about 2 hours for the full 16 steps, and a dissertation mode producing 25,000-80,000 words over 300-450 sources.
@chroniki_ai [Claude Code]
https://x.com/chroniki_ai/status/2088907144188064230
The context-rot fix making the rounds in Japan: instead of stuffing one session until early instructions fade, hand Claude Code a single design file defining seven roles on a three-shift relay, scout, architect, engineer, reviewer, deployer, reporter, orchestrator, where each role finishes and passes the baton instead of everyone working at once. Non-engineers can run it, and her practical advice is to start with the scout role alone, because vague role definitions are what make these systems wander.
@_DQQ_1990 [Claude Code]
https://x.com/_DQQ_1990/status/2088797171684606009
Home in China on vacation, he didn't dare open Claude Code for fear of a ban and knew pure Codex would eat his $200 weekly quota in a night. So he had Codex hand off to DeepSeek V4 Pro in a homemade loop for long-running tasks and burned 2 billion tokens in three days for about 140 yuan, roughly $20, with GPT 5.6 Sol at the top for planning and final review. His estimate for the same workload on Opus: $5,000 or more. His conclusion was not that quality matched, but that V4 Pro sits between Opus and Sonnet while being fast and absurdly cheap, and even a 10x price rise wouldn't change his decision.
@nagataprogram [Claude Code]
https://x.com/nagataprogram/status/2088917447243894822
A freelance web builder's honest accounting of the web is dead discourse: Claude Code made his coding about 5x faster, and he watches contract rates fall accordingly, because nobody keeps paying the old price for the newly cheap part. What still sells is judgment: for a recruiting-site pitch he audited four local competitors, found zero actually designed for hiring and search results owned by job boards, repositioned the site as the destination for brand-name searches, and delivered a 10-page proposal and quote two days after the first meeting against a one-week promise. His two AI-proof weapons: same-day replies and the guts not to decline.
@jasonoliver [OpenClaw]
https://x.com/jasonoliver/status/2088841374485422299
Months of running an OpenClaw agent changed how he uses the internet: personalized research, automated tracking, and an always-on companion that reads all the subreddits and watches all the YouTube he doesn't have time for, then synthesizes it. His sober caveat came in a second post: he did the compliance exercise for what it would take a VC firm to run one properly, which makes it funny and alarming to watch people blindly hand passwords to anonymous free computer-use agents. His verdict on the space: crowded, nobody has gotten it right yet, and whoever does unlocks the agentic internet for normal people.
@TC8880 [OpenClaw]
https://x.com/TC8880/status/2089065513284673958
One of the earliest OpenClaw adopters, who bought a gaming laptop for it in February, formally uninstalled it. The record of what it cost him: sleepless AI-FOMO nights, hours on skills, configs, and rescuing it from random disconnects and amnesia. Six months later, everything that made it magnetic has been absorbed into big-vendor products that are more stable and near-zero setup; his Codex now houses his assistant, designer, programmer, editor and analyst. His consolation is the clean summary of the whole cycle: he didn't miss anything, he just beta-tested the future six months early at his own expense.
@SparXPodcast [Claude Code]
https://x.com/SparXPodcast/status/2088981854968037783
Wispr Flow's founder describes his newest hire: no CS degree, never wrote production code, now automating the company's entire post-sales function. The qualification was hospitality-grade customer instinct plus being decently familiar with Claude Code. The founder's own output has flipped too: 99.9% of the company's code is AI-generated, and he argues his 18 years of programming are exactly why his prompts reach finished product 4-5x faster. What he hires for now is two skills: first-principles thinking and decomposing complex problems.
@HowardTheDuck48 [Claude Code]
https://x.com/HowardTheDuck48/status/2088818961572061317
Sixty years of programming, and he calls the last three months with Claude Code the most productive of his life, with the honest detail that half of that time went to getting Claude up to speed on how he wants things done. The onboarding cost is real and front-loaded; the payoff compounds after.
@bboym0dE [Claude Code]
https://x.com/bboym0dE/status/2088860395897541044
Claude Code has been building two of his projects continuously for 72 hours in loop plus gauntlet mode and is still going, with both projects aimed at the biggest pain points he has hit building apps with AI. He reports them noticeably improving every day, which is the quiet norm-shift buried in this post: continuous multi-day runs are becoming an unremarkable way to develop.
🗣 User Voice
User Voice
Reliability replaced capability as complaint number one. The August 16 outage ran roughly 22 hours across auth, API, Claude Code and Cowork, with a second degradation incident opened four minutes after the first (@buildwithhassan), and the status page still showed operational while nothing worked (@SKatalystAI). The tally that stuck: 164 incidents since January, and 90-day uptime of 99.34% on Claude Code, below the 99.9% enterprise bar (@apricotkinkuma).
Scope discipline is still broken. Ask Claude Code to fix one bug and it fixes 15 others, refactors two unrelated files, declares victory, and the one bug you asked about is still there (@totoche).
The AGENTS.md standoff is costing Anthropic goodwill with maintainers. Pull requests from Claude Code into the Elixir repo are worse because the agent won't read the project's AGENTS.md instructions (@josevalim), while mitsuhiko deleted his CLAUDE.md files entirely and found a one-line pointer telling Claude to read AGENTS.md works fine (@mitsuhiko) — which makes the refusal look like policy, not capability.
Lock-in is being named in public: subscriptions banned from third-party harnesses, no AGENTS.md support, lobbying framed as safety — one widely-read post assembled the list as a pattern of anticompetitive behavior rather than isolated choices (@lian75864).
The value math is tightening at both ends. On the $20 plan, one Opus 5 plan-and-build workflow barely fits inside the 5-hour window (@jzgarden), and at the frontier end, running agents 24/7 without a clear plan just manufactures review fatigue: the bottleneck is deciding what to build, not executing it (@pauliusztin_).
Reliability replaced capability as complaint number one. The August 16 outage ran roughly 22 hours across auth, API, Claude Code and Cowork, with a second degradation incident opened four minutes after the first (@buildwithhassan), and the status page still showed operational while nothing worked (@SKatalystAI). The tally that stuck: 164 incidents since January, and 90-day uptime of 99.34% on Claude Code, below the 99.9% enterprise bar (@apricotkinkuma).
Scope discipline is still broken. Ask Claude Code to fix one bug and it fixes 15 others, refactors two unrelated files, declares victory, and the one bug you asked about is still there (@totoche).
The AGENTS.md standoff is costing Anthropic goodwill with maintainers. Pull requests from Claude Code into the Elixir repo are worse because the agent won't read the project's AGENTS.md instructions (@josevalim), while mitsuhiko deleted his CLAUDE.md files entirely and found a one-line pointer telling Claude to read AGENTS.md works fine (@mitsuhiko) — which makes the refusal look like policy, not capability.
Lock-in is being named in public: subscriptions banned from third-party harnesses, no AGENTS.md support, lobbying framed as safety — one widely-read post assembled the list as a pattern of anticompetitive behavior rather than isolated choices (@lian75864).
The value math is tightening at both ends. On the $20 plan, one Opus 5 plan-and-build workflow barely fits inside the 5-hour window (@jzgarden), and at the frontier end, running agents 24/7 without a clear plan just manufactures review fatigue: the bottleneck is deciding what to build, not executing it (@pauliusztin_).
📡 Eco Products Radar
Eco Products Radar
DeepSeek Harness (dsh) — 122K GitHub stars in three days; the everything-is-a-plugin runtime was the loudest ecosystem story of the day
Grok Bot — the hosted agent every OpenClaw and Hermes user is now measuring against
Codex — the default second agent, and for a growing minority the first
Hermes Agent — the self-hosted reference point, fresh off a massive release week
OpenClaw — increasingly discussed in the past tense, including by its earliest fans
Cursor — bundled into Grok subscriptions, still the default IDE comparison
Remotion — the code-defined video engine behind the day's biggest creative posts
Qwen3.8-27B — the open-weights model people are plugging into Claude Code on consumer GPUs
OpenSandbox — Alibaba's isolated agent environments, 13k stars and climbing
orca — fan one prompt across five coding agents in parallel worktrees, keep the winner
Graft — persistent codebase memory stored in git, pitched at Claude Code's re-reading habit
Spec Kit — GitHub's spec-driven development toolkit, now spanning 30+ agents
Paperclip — the org-chart and ticket layer for running teams of agents
Superpowers — the skills framework that forces plan-before-code
Higgsfield — the video-generation half of nearly every content pipeline posted
ElevenLabs — the voice layer in those same pipelines
Obsidian — the vault agents keep being pointed at, for better and worse
DeepSeek Harness (dsh) — 122K GitHub stars in three days; the everything-is-a-plugin runtime was the loudest ecosystem story of the day
Grok Bot — the hosted agent every OpenClaw and Hermes user is now measuring against
Codex — the default second agent, and for a growing minority the first
Hermes Agent — the self-hosted reference point, fresh off a massive release week
OpenClaw — increasingly discussed in the past tense, including by its earliest fans
Cursor — bundled into Grok subscriptions, still the default IDE comparison
Remotion — the code-defined video engine behind the day's biggest creative posts
Qwen3.8-27B — the open-weights model people are plugging into Claude Code on consumer GPUs
OpenSandbox — Alibaba's isolated agent environments, 13k stars and climbing
orca — fan one prompt across five coding agents in parallel worktrees, keep the winner
Graft — persistent codebase memory stored in git, pitched at Claude Code's re-reading habit
Spec Kit — GitHub's spec-driven development toolkit, now spanning 30+ agents
Paperclip — the org-chart and ticket layer for running teams of agents
Superpowers — the skills framework that forces plan-before-code
Higgsfield — the video-generation half of nearly every content pipeline posted
ElevenLabs — the voice layer in those same pipelines
Obsidian — the vault agents keep being pointed at, for better and worse
Comments