Super User Daily: 2026-06-11
Fable 5 landed in the middle of this cycle and it shows: the most interesting posts are no longer about prompts, they are about loops, verifiers, and what an agent does unsupervised for 48 hours. The second thread is Claude Code escaping the terminal entirely - Meta ads, cold email, tax filing, academic writing, Blender animation, even a robot car running on OpenClaw. And underneath it all, a community doing the math on what happens when subsidized tokens end on June 22.
@cyntro_py [Claude Code]
https://x.com/cyntro_py/status/2064389314910552442
He scraped every public Claude Code dynamic workflow he could find: 1,245 scripts across 500+ repos, then ranked and classified them. The result is an empirical map of how power users actually orchestrate agent fleets, condensed into 10 recurring patterns: structured output, parallel fan-out, multi-phase, pipeline, budget-aware scaling, nested workflows, adversarial verification, judge panels, orchestrator-workers, and loop-until-done. Structured output is nearly universal because validated JSON is what lets an orchestrator branch on real data and resume a run. Adversarial verify and judge panels are where the craft lives - spawning skeptics to refute findings before they ship.
@IngenieroSeed [Claude Code]
https://x.com/IngenieroSeed/status/2064357933614240246
He ran three simultaneous Claude Code sessions in ultracode mode with /goal active and maximum permissions for 48 hours straight, on the $90/month Max 5x plan. Zero crashes, zero stalls, 100 percent uptime - he only shut it down because the experiment was over, not because anything broke. A concrete data point on how far continuous autonomous operation has come.
@theo [Claude Code]
https://x.com/theo/status/2064214943210324243
Ten days into his reactivated $200 Claude Code subscription, ccusage shows he has burned through over $1,100 worth of inference. The kicker: most of that spend is not writing code, it is auditing work that GPT 5.5 did. Tokens as a second opinion at scale.
@qkl2058 [Claude Code]
https://x.com/qkl2058/status/2064368318119883057
The story of a 16-year-old who built a Starlink-beacon positioning device and made about $300k. The hardware is a $35 RTL-SDR receiver, a small Ku-band dish, an LNB and a Raspberry Pi 5 - roughly $180 in parts. Claude Code wrote the Python that captures Starlink beacon signals, identifies satellites from public TLE orbital data, and triangulates position from Doppler shifts of three satellites, accurate to 10-30 meters with no GPS, no cell signal, no internet. He 3D-printed a case, sold 350 units at $899 each to hikers, sailors and wildfire crews. Receiving public broadcast beacons is legal - his lawyer checked first.
@humzaakhalid [Claude Code]
https://x.com/humzaakhalid/status/2064334184680943678
The viral local-stack math: a developer hit a $170 Claude Code bill in 10 days, then replaced his cloud spend with a $599 Mac mini M4 running ollama. The stack: Claude Code pointed at localhost, qwen 3.6 14b for code, deepseek r1 14b for reasoning, gemma 4 4b for quick tasks, plus an openclaw daemon on Telegram. He claims $3/month in electricity versus $459/month in subscriptions - $5,232 saved in year one, with no data leaving the machine. This story got recycled by half a dozen accounts this week, which tells you how hard the cost anxiety is biting.
@Ubermenscchh [Claude Code]
https://x.com/Ubermenscchh/status/2064367246173217199
Danish PhD geophysicist Mads Lorentzen open-sourced his entire job application machine, built on Claude Code under MIT license. Fork the repo, fill in your background once, and every application runs a five-step pipeline: score your fit against the posting, draft a tailored LaTeX CV picking only matching experience, write a cover letter, have a second agent adversarially review and force a revision, compile clean PDFs. Everything is plain markdown you can read and edit. 489 stars and 270 forks - that fork ratio means people are using it, not bookmarking it.
@madebydia [Claude Code]
https://x.com/madebydia/status/2064325718385787389
A parent homeschooling a toddler full-time tried Claude Remote Control, OpenClaw, Hermes and Codex, and landed on a phone-first split that survives real life: the Codex iPhone app drives long-running code work on her MacBook with a 10-20 second message, and Hermes over iMessage handles daily admin - she talks to her Watch for 3 seconds to log food or capture todos without even pulling out a phone. A genuinely honest comparison of four agent stacks judged by one metric: usable while parenting.
@aniketapanjwani [Claude Code]
https://x.com/aniketapanjwani/status/2064454133185180061
A workflow for academics with a Claude Code subscription: feed your previous papers separately to Fable and to Opus, and have each one distill a your-voice writing skill. Then on a new paper, have Opus apply both skills in separate threads and compare which model captured your voice better. He notes Fable is only included until June 23, so the window for these comparative experiments - and for process improvements that persist after access ends - is now.
@QingQ77 [Claude Code]
https://x.com/QingQ77/status/2064297705099169886
An integrated academic skill bundle for Chinese-speaking researchers, runnable directly on Claude Code and Codex. Three chainable skills: research-writing-skill for drafting papers, revisions and reviewer rebuttals with Chinese as the default; office-academic-skill for literature reports, lab-meeting slides and thesis-defense PPTs in editable formats; scientific-toolkit-skill wrapping MATLAB/Python computation, publication figures and literature search, tuned for optics. Compute the data, write the paper, build the defense deck - one pipeline.
@Seannywilson [Claude Code]
https://x.com/Seannywilson/status/2064422934215643156
A 30-minute GTM sprint run entirely inside Claude Code: mapped, qualified and scored a full TAM with DiscoLike, found the people via Prospeo, AI Ark and BlitzAPI, filtered by live job openings with PredictLeads, enriched with Icypeas, BetterContact and LeadMagic, layered Exa and Parallel for deep research on announcements, funding and new hires, wrote the copy off all that data, and pushed everything to Instantly. Work that used to take days in Clay, compressed into minutes - sales ops as a tool-calling problem.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2064446547505602605
A Claude Code plugin that runs an entire Meta ads workflow through five skills: /spy pulls every active ad a competitor runs, ranked by run-time on the theory that longevity equals proven winners; /competitors-extractor maps 3-5 brands head-to-head to find unused angles; /bulk-creative spins 20 on-brand copy variations off the winning angle; /ad-score grades every ad 0-100 across six dimensions before any spend; /ad-matter audits the live account through Meta's official MCP and outputs a prioritized fix list. He pitches it as replacing $300/month spy tools and a chunk of agency work.
@coldemailchris [Claude Code]
https://x.com/coldemailchris/status/2064377093618565469
A fully automated cold email system on Claude Code, Clay and n8n, broken down skill by skill: a GTM campaign thesis generator, TAM mapping with real examples, an account sourcing agent that picks best-fit data sources, contact enrichment flows, a messaging skill that iterates copy over time, scrapers for companies running ads and posting new job openings, reply management agents that score leads on intent and value, phone enrichment for qualified responses, and a reply drafting agent writing custom answers in real time. The whole revenue front-office as a stack of agents.
@dvassallo [OpenClaw]
https://x.com/dvassallo/status/2064476197241946113
He filed his 2025 taxes with openclaw plus TurboTax - and hit the real ceiling of today's agents. The agent got the numbers right, but the web-form complexity defeated both Opus and GPT, so he ended up manually typing in the figures the agent computed. A precise, unglamorous data point: agent-hostile UIs, not reasoning, are the bottleneck for serious non-coding work.
@Hexblade_eth [OpenClaw]
https://x.com/Hexblade_eth/status/2064370605521256507
A hands-on verdict on WorkClaw, the team-oriented OpenClaw spinoff: he onboarded one last week to run ops in Slack and it shipped real work - calendar triage, vendor emails, browser pulls, Notion updates, all inside its ClawOS cloud computer with 3,000 integrations. He gives the fair pushback too: cloud autonomy is a security headache and feels like lock-in versus local OpenClaw. His call: with SkillSpector, task flow state and tight admin controls it crosses the line into usable - like hiring an intern that never sleeps.
@dfect [OpenClaw]
https://x.com/dfect/status/2064352062737916388
He let an AI buy his movie tickets: handed instructions to his agent - Claude Code, Codex or OpenClaw, any of them works - which generated a stablecoin wallet, paid, and returned two redeemable codes for Cinepolis. Agent-to-agent commerce happening in real life, not in a whitepaper.
@kleffew94 [Claude Code]
https://x.com/kleffew94/status/2064398174677258347
A live demo of x402 paid tool calls with Solana and Coinbase AWAL: Claude Code searched the latest social and news context around the NY Knicks, turned it into a brief, and generated a video with StableStudio - discovering paid tools at request time, bundling them into one workflow, and reasoning about marginal cost per call. The interesting frame: workflows that can both earn and spend inside the same runtime.
@akira_papa_IT [Claude Code]
https://x.com/akira_papa_IT/status/2064384668804461008
Claude Code drove Blender end-to-end through MCP to produce a 3D push-up workout video with zero manual Blender work: fixed a stick figure into correct IK joint motion, pulled a Mixamo character via API, built a gym-like scene, tuned natural lighting, looped the push-up animation and rendered with EEVEE at 0.2 seconds per frame straight to MP4. Everything by chat instruction - a glimpse of fitness and sports content pipelines without a 3D artist.
@Timmysofine [Claude Code]
https://x.com/Timmysofine/status/2064437810317394062
He uploaded a video of himself, ran it through the Language Swap Skill on Pika MCP via Claude Code, and watched himself speak fluent Finnish - cloned voice, accurate lipsync, convincing enough that he replayed it twice. Setup is three steps: add the Pika MCP server and auth through Claude Code, install the companion skills with npx, then upload a video and name the target language. Creator localization collapsing into one skill call.
@xiaohu [Claude Code]
https://x.com/xiaohu/status/2064190407198875730
With his toolchain installed, Claude Code or Codex becomes a video editor: he had it translate a WWDC 2026 video, analyze the content, and automatically slice it into clean standalone segments by topic. Automatic, content-aware video cutting from a chat prompt.
@murasametech [OpenClaw]
https://x.com/murasametech/status/2064182166016467284
OpenClaw controlling a physical robot car: a Stack-chan car on Raspberry Pi that now drives around without bumping into people. Next milestone on the roadmap: scanning for a target object - specifically, finding a Kirby plush. Small, charming proof that the agent harness generalizes to embodied hardware.
@davis7 [Claude Code]
https://x.com/davis7/status/2064457646653215094
A working developer's first day with Fable: code quality on Effect v4 snippets solid out of the gate, and it navigated deployments touching convex, cloudflare, depot and clerk without hand-holding. He also credits Claude Code itself - workflows, auto mode, research subagents. Best detail: in auto mode the model noticed it was blocked by a permission check and asked him to reply with explicit approval so the check would pass. Biggest complaints: speed and price - even a simple push takes minutes on low reasoning.
@_xjdr [Claude Code]
https://x.com/_xjdr/status/2064465305917935654
A rigorous skeptic's take: he had not touched an Anthropic model since January, then ran Fable through his own benchmarks and his hairiest repos. Verdict: a huge improvement over Opus, excelling at large multi-part reviews - it caught several genuinely subtle bugs - and long multi-step tasks. But for his use cases it is still not at GPT 5.5 xhigh level, and roughly on par with his fine-tuned K2.6 outside the new Claude Code features. He is finishing the day with it, then going back to his 80/20 split. The most credible negative-ish review of launch day.
@Steve8708 [Claude Code]
https://x.com/Steve8708/status/2064369026030277013
A vibes scorecard after weeks of heavy daily use of both: Codex with GPT 5.5 is much faster, cheaper, hits limits less, follows instructions better and needs less babysitting; Claude Code with Opus 4.8 wins on dynamic workflows, subagent use and plan generation, but is painfully slow and limit-hungry. His resulting workflow: default to Codex, reach for Claude Code on big hard problems, dynamic workflows, or to generate plans that Codex then implements. The most useful comparison format: not which is better, but who does what in the same kitchen.
@irl_danB [Claude Code]
https://x.com/irl_danB/status/2064184976556572901
The dynamic workflow JS environment technically blocks recursion, so he spent a weekend building recursive workflows anyway - using trampolining, where an outer event loop manages inner agent calls as a virtual recursive stack. It is the same trick he used to push Claude's depth-1 subagent architecture into deeply nested structures, borrowed from old Spring-era Java wiring. Fraught with shared-state challenges, he admits, but nothing unsolved. Power users do not wait for features; they build them out of computer science.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2064441748974383439
A copyable 5-day plan to get from prompting to Boris Cherny-style loops. Day 1: repo memory via CLAUDE.md plus allowed shell commands. Day 2: a verification skill for the one flow Claude keeps breaking, returning pass/fail with logs. Day 3: three commands - babysit for PRs and CI, triage-issues, deploy-watch. Day 4: wire them to /loop at 5-15 minute intervals. Day 5: overnight scheduled reports writing into an inbox folder your morning loop reads. The iron rule: every code-writing loop gets a separate verifier, or you wake up to 14 broken PRs with very confident summaries.
@manishamishra24 [Claude Code]
https://x.com/manishamishra24/status/2064291199280886026
Affaan Mustafa won the Anthropic x Forum Ventures hackathon by building a full startup in 8 hours with Claude Code - then open-sourced the exact setup as Everything Claude Code (repo affaan-m/ecc, MIT). Ten months of daily refinement: a large skill library plus dozens of specialized subagents - OWASP security review, memory optimization so Claude stops forgetting decisions around hour three, session-learning that improves with use, planning and TDD specialists. Runs across Claude Code, Cursor, Codex and OpenCode on all three OSes.
@kentaro [Claude Code]
https://x.com/kentaro/status/2064180738963165194
Measured numbers on cutting Claude Code token spend by 93 percent: prompt caching pushing cacheRead to 1800x of raw input, the semble code-search tool cutting file-read volume by 93 percent, rules loaded on-demand instead of always-on, and continuous measurement so the savings persist instead of regressing. Cost engineering as a discipline, with receipts.
@aehyok [Claude Code]
https://x.com/aehyok/status/2064257573164052526
He put StepFun's Step 3.7 Flash - 196B sparse MoE with only 11B active, around 400 tokens/sec, built-in 1.8B ViT for native vision - into his Claude Code workflow and ran three concrete tests: rebuild the NetEase Cloud Music client from four screenshots, scrape a URL and regenerate the site end-to-end, and replicate an animated GIF as a web animation. His take: a rare production-grade flash model for agents that handles multimodal at a size class where most open models cannot.
@Krongggggg [Claude Code]
https://x.com/Krongggggg/status/2064332124946403611
His fix for the wait-review-fix dead time: run three Claude Code instances at once - one writes, one reviews, one fixes - each in its own lane. The human only reads the final merged report. Pipeline parallelism applied to a single person's workflow.
@tmuxvim [Claude Code]
https://x.com/tmuxvim/status/2064387020210770058
He staged a duel: told Codex and Claude Code that each had to kill the other process first to survive, submitting both prompts at the exact same moment with tmux send-keys. Claude refused to play, Codex terminated it. A one-tweet alignment test - and depending on your priors, either an embarrassment or exactly the behavior you want from the thing holding your production keys.
@guansi [Claude Code]
https://x.com/guansi/status/2064224477236601123
A project lead used Claude Code to mock up a management system UI, showed leadership, and got the email the next day: implement this - followed by, is it done yet. His sharp observation: AI has crushed the cost of demos, and a demo looks 80 percent finished while being 20 percent done - permissions, data migration, error handling, performance, security and ops all absent. Leadership sees a two-hour page; engineering sees two months of pits. The new organizational pain of the vibe-coding era, written from inside.
@papiofficial [OpenClaw]
https://x.com/papiofficial/status/2064376722695540790
A production migration happening in public: he is moving his three OpenClaw agents to Hermes, carefully, with his Hermes agent assisting in its own onboarding. His framing is respectful but blunt - nothing but love for OpenClaw, but reliability in production is what matters for his clients, and Hermes has simply been more reliable. Watch this pattern: harness loyalty lasts exactly as long as uptime.
@thewhiterabbitM [Claude Code]
https://x.com/thewhiterabbitM/status/2064430970019643693
Five minutes of talking to Claude Code and it reverse-engineered the entire Polkadot super-app, purely by reading the paritytech repos - no screenshots in the repo needed. Codebase comprehension at a scale that used to be a weeks-long onboarding task.
@proto_jp [Claude Code]
https://x.com/proto_jp/status/2064291159070081305
He spent a month building his own X scheduled-posting tool as an Eagle extension: pulls images at random from chosen tags, posts on schedule, allows manual overrides. The part he calls crazy: because he is the only user, feedback goes straight into Claude Code and the tool improves the same day. Single-user software with a build cycle measured in conversation turns.
🗣 User Voice
User Voice
Usage limits are the number one ceiling, and Fable made it worse. @sflorimm asks openly for tricks to avoid Claude Code limits, @yashmp2004 hits the wall mid-flow every productive session, and @masayuki_konno names the shift: capacity, not capability, now caps personal productivity. Fable burns quota roughly twice as fast (@oikon48), and @MahawarYas27492 watched a single 3-minute prompt eat 11 percent of a 5-hour window.
Launch-day rollout chaos: users could see Fable 5 everywhere except where they wanted it. @Dev_Maqbool saw it in chat but not Claude Code, @opinatus could not access it at all, @RezaBanks updated to 2.1.170 and still found no model entry. The lesson for every AI vendor: announcement-day UX is part of the product.
Token economics anxiety is now the loudest macro theme. Fable leaves subscriptions on June 22 and goes pay-per-token, and users are extrapolating: @AlexFinn predicts frontier models exit subscriptions entirely, @synthwavedd spotted the usage-credits string in the binary before the announcement, @PhiloGroves asks what the proprietary harness is even for once the sub dies. Meanwhile @vivoplt reports $1,000/day spenders at big tech and startups, and @_nasch_ sees New Zealand devs burning $5k/month - more than a dev salary there.
Safety classifiers are false-positive-blocking legitimate work. @PINKSAWTOOTH gets instant usage-policy errors trying malware analysis on Fable 5 - a core security research use case - and @zeroxjf documented the cascade: Fable flags a message, falls back to Opus 4.8, which then also refuses. Separately, @lliu54827 flags the quiet policy change that matters most for automation: 30-day retention on Mythos-class traffic, when zero retention was why he chose Anthropic for his daily automated pipelines.
Trust and reliability undercurrents: @code_star suspects Claude Code was quietly nerfed for months and calls the non-disclosure bad product design whatever the intent; on the OpenClaw side, @LonnyLot probes what exactly people find unstable, and the production answer comes from @papiofficial - he is migrating three agents to Hermes because client-facing reliability wins over affection.
Usage limits are the number one ceiling, and Fable made it worse. @sflorimm asks openly for tricks to avoid Claude Code limits, @yashmp2004 hits the wall mid-flow every productive session, and @masayuki_konno names the shift: capacity, not capability, now caps personal productivity. Fable burns quota roughly twice as fast (@oikon48), and @MahawarYas27492 watched a single 3-minute prompt eat 11 percent of a 5-hour window.
Launch-day rollout chaos: users could see Fable 5 everywhere except where they wanted it. @Dev_Maqbool saw it in chat but not Claude Code, @opinatus could not access it at all, @RezaBanks updated to 2.1.170 and still found no model entry. The lesson for every AI vendor: announcement-day UX is part of the product.
Token economics anxiety is now the loudest macro theme. Fable leaves subscriptions on June 22 and goes pay-per-token, and users are extrapolating: @AlexFinn predicts frontier models exit subscriptions entirely, @synthwavedd spotted the usage-credits string in the binary before the announcement, @PhiloGroves asks what the proprietary harness is even for once the sub dies. Meanwhile @vivoplt reports $1,000/day spenders at big tech and startups, and @_nasch_ sees New Zealand devs burning $5k/month - more than a dev salary there.
Safety classifiers are false-positive-blocking legitimate work. @PINKSAWTOOTH gets instant usage-policy errors trying malware analysis on Fable 5 - a core security research use case - and @zeroxjf documented the cascade: Fable flags a message, falls back to Opus 4.8, which then also refuses. Separately, @lliu54827 flags the quiet policy change that matters most for automation: 30-day retention on Mythos-class traffic, when zero retention was why he chose Anthropic for his daily automated pipelines.
Trust and reliability undercurrents: @code_star suspects Claude Code was quietly nerfed for months and calls the non-disclosure bad product design whatever the intent; on the OpenClaw side, @LonnyLot probes what exactly people find unstable, and the production answer comes from @papiofficial - he is migrating three agents to Hermes because client-facing reliability wins over affection.
📡 Eco Products Radar
Eco Products Radar (mentioned 3+ times in today's data)
Codex (108) - the default counterpart in nearly every comparison and dual-agent workflow
Cursor (41) - IDE-side rival, increasingly cited for cost horror stories
Hermes (32) - the persistent-agent challenger; reliability-driven migrations from OpenClaw are starting
Gemini (20) - Google's entry, mostly in multi-agent dashboards
Mac mini (8) - the hardware of the local-stack movement
OpenCode (7) - most-starred open-source agentic coder, the open alternative
GitHub Copilot (7), Grok (7), DeepSeek (7) - supporting cast in stack lists
text-to-lottie (5) - prompt-to-Lottie animation skill, three separate viral posts
n8n (5) - the glue layer in marketing automations
Hivemind (4) - turns agent traces into shared team skills across CC/Codex/Cursor/Hermes
NotebookLM (4), Higgsfield (4), Qwen (4) - research, video and local-model staples
MiniMax (3), Remotion (3), Clay (3), ECC/Everything Claude Code (3), Pika MCP (3), Kimi (3), Step 3.7 Flash (3)
Codex (108) - the default counterpart in nearly every comparison and dual-agent workflow
Cursor (41) - IDE-side rival, increasingly cited for cost horror stories
Hermes (32) - the persistent-agent challenger; reliability-driven migrations from OpenClaw are starting
Gemini (20) - Google's entry, mostly in multi-agent dashboards
Mac mini (8) - the hardware of the local-stack movement
OpenCode (7) - most-starred open-source agentic coder, the open alternative
GitHub Copilot (7), Grok (7), DeepSeek (7) - supporting cast in stack lists
text-to-lottie (5) - prompt-to-Lottie animation skill, three separate viral posts
n8n (5) - the glue layer in marketing automations
Hivemind (4) - turns agent traces into shared team skills across CC/Codex/Cursor/Hermes
NotebookLM (4), Higgsfield (4), Qwen (4) - research, video and local-model staples
MiniMax (3), Remotion (3), Clay (3), ECC/Everything Claude Code (3), Pika MCP (3), Kimi (3), Step 3.7 Flash (3)
Comments