Super User Daily: 2026-05-31
Two stories shaped the day. First, Opus 4.8 + Dynamic Workflows turned Claude Code into a fan-out engine — users routinely spun up 100+ parallel subagents and watched their token meter melt, with single tasks burning $30 in extra credits or 4.7M tokens in two hours. Second, the cost backlash from Microsoft, Uber, and an unnamed $500M-a-month Claude customer collided with Jevons paradox: cheaper inference made everyone use more, not less. Underneath those two narratives, the most interesting cases were people pointing Claude Code at things that have nothing to do with code — installing Google Play on a Doubao phone over USB, generating a music video ad for a luggage brand in 30 minutes, building a personal CFO over a small factory's Tally bookkeeping data, or scanning local businesses without websites and pitching them automatically while you sleep.
@MMMusol [OpenClaw]
https://x.com/MMMusol/status/2060276116615729374
Detailed thesis on how the Reddit-WSB-banned trader Serenity caught the Raspberry Pi (RPI) move in February. The trade was driven by OpenClaw demand: as people started buying tens to hundreds of Raspberry Pi boards to run parallel AI agent swarms (instead of expensive Mac Minis), Serenity correctly translated the same demand shock that's a rounding error for Apple into a 4x earnings beat for a 540M-euro company. March 31 financial report came in at 58% revenue growth versus Wall Street's 14% expectation, stock peaked +175%. Worth reading for the framework: same demand, different company size, different financial impact.
@bcherny [Claude Code]
https://x.com/bcherny/status/2060390852619272526
Salesforce published a detailed engineering writeup on going agentic with Claude Code. Two numbers jumped out: a migration originally scoped at 231 days shipped in 13 days, and a single PR delivered 21 endpoints at 100% test coverage. This is the first credible enterprise-scale data point on Dynamic Workflows in production, not on a benchmark.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2060168087165685799
Deep technical breakdown of Bun creator Jarred Sumner's port of Bun runtime from Zig to Rust using Dynamic Workflows: roughly 750K lines of Rust, hundreds of parallel agents with two independent reviewers per file, 11 days from first commit to merge, original test suite passing at 99.8%. Includes a tour of how the workflow runtime works under the hood — JS orchestration scripts running in a Node vm sandbox, deterministic caching via sha256(script + prompt + opts), 180s stall detection, 16-agent local concurrency cap, 1000 agent calls per workflow. The 99.8% number is the load-bearing one: it shows the dual-reviewer pattern can hold quality on a real codebase-scale migration.
@wshuyi [Claude Code]
https://x.com/wshuyi/status/2060286090448376026
Used Dynamic Workflows for a data analysis task. The first stage alone spawned 108 subagents in parallel. His Claude Code Max 5x plan (5-hour quota) drained in 20 minutes. Switched to Extra Credits and burned another $30 to finish. Notes the data analysis result was excellent — but the cost shape is real and immediate.
@yoshio_nocode [Claude Code]
https://x.com/yoshio_nocode/status/2060280843441147915
The /deep-research command spawned 101 agents to research the latest overseas AI business success cases in parallel, fact-checked them, and returned a unified report. Single query consumed 3.4 million tokens. Concludes you can't run it casually but it's the right call for serious research.
@Jeremybtc [Claude Code]
https://x.com/Jeremybtc/status/2060322740901659000
Gave Claude Code ultracode a fairly simple read-only codebase audit task. It ran for over 2 hours, activated 139 different subagents, consumed roughly 4.7M tokens. The "read-only audit" framing was the surprise — even non-write workflows now consume agent-team-scale tokens.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2060391222611747305
Arize AI's CPO demoed live during a conversation: built a working PM agent that pulls GitHub issues, scores by priority, generates a daily triage report, runs on cron, then instrumented tracing in a single command and had Claude write the first eval — the whole self-improving agent took about an hour from "I have an idea" to "it's running my daily triage." Bigger framing: the PM workflow that used to gate every feature request through a sprint cycle now runs morning-to-evening, prototype to production in a day.
@shedntcare_ [OpenClaw]
https://x.com/shedntcare_/status/2060301611206119689
Built an OpenClaw workflow that runs 6 AI agents nonstop. It scans local businesses with missing or outdated sites, extracts services/reviews/photos, generates a custom website demo, sends personalized outreach with a live preview, includes a payment link, and follows up to close the deal. End-to-end prospecting plus fulfillment plus sales, no manual lead gen, no design work, no sales team.
@shivsakhuja [Claude Code]
https://x.com/shivsakhuja/status/2060238059132899655
Made a complete music video ad for Away luggage entirely in Claude Code in about 30 minutes, terminal only. The flow: Claude Code scraped real product photos from Away's site, generated a Suno song prompt, generated character + world stills with gpt-image-2/nano-banana via fal, locked them for consistency, drafted a scene-by-scene storyboard synced to the song's timing, generated keyframes per scene, ran Veo 3.1 / Seedance for keyframe-to-video, used ffmpeg to stitch and review frame-by-frame, then burned word-by-word captions synced to lyrics. The agent is the whole video editing suite.
@nicos_ai [Claude Code]
https://x.com/nicos_ai/status/2060412136321335716
Gave Claude Code one prompt to find AI engineering roles in SF, full-time, on-site, posted in the last month, exclude senior+ titles, open each listing to confirm still hiring, return 10 different companies with direct apply links. Used ego lite so it logged in with the real LinkedIn session, worked in an isolated Space while the user kept their tabs, came back with a clean verified shortlist. Real LinkedIn session beats every cookie-script demo.
@shobhitic [Claude Code]
https://x.com/shobhitic/status/2060348070030872699
A couple of factory owners he knows in Jaipur use ChatGPT and credit it with growing their export business — mostly to write messages. He set up Claude Code over their Tally bookkeeping data and produced cashflow statements, which blew their minds. Argues the bottleneck on AI use cases isn't tooling, it's imagination.
@op7418 [Claude Code]
https://x.com/op7418/status/2060367382884094026
Had been failing for ages to install the Google framework on his Doubao phone via Google Play. Today opened Claude Code, enabled USB debugging, told CC to handle it — CC downloaded the installer, installed it, debugged everything automatically. Not coding, just a domestic tool wrangling a stuck Android workflow.
@nijokestu [Claude Code]
https://x.com/nijokestu/status/2060253639009030258
Tested Pika MCP Founder Kit Skills, asked Claude Code to build a complete brand for a fitness app from scratch in one session. Output: full 15-page brand guidelines, logo files plus 13 custom icons, fonts, color palette, design tokens, AI prompts for tweets/emails/landing pages, cover hero photography. One prompt, no designer, no agency brief. The cheat code for technical founders who can build the product but die on marketing.
@CEO_loves_tech [Claude Code]
https://x.com/CEO_loves_tech/status/2060215886481588484
His happiest setup: rent a decent VPS, drop Codex + Claude Code + Hermes Agent on it, send instructions from MacBook or iPhone 24/365. Even running a local LLM, infrastructure ends up ~¥10K/month (excluding LLM cost). Asks why anyone bothers buying a Mac Mini.
@ruanyf [OpenClaw]
https://x.com/ruanyf/status/2060176579695915335
The OpenClaw founder published his monthly token usage: 603 billion tokens. At the standard rate that would cost $1.3M. He doesn't actually pay — he's an OpenAI employee with unlimited internal access — but the number is a measuring stick. If a company gave one engineer unlimited frontier model access, the bill would be ~9M RMB/month = ~100M RMB/year for one person. Even routing to cheap open-source domestic models at 1/30 to 1/50 the price still lands at 2-3M RMB/year per engineer. AI programming, fully unleashed, is more expensive than a human programmer.
@ren_aivest [Claude Code]
https://x.com/ren_aivest/status/2060273992867758165
Reports running 192 Claude Code instances simultaneously, with model auto-switching: 8 agents spawned 192 sub-agents, then condensed back to 1 master, for 201 total CC instances in one job. Two days earlier he was reporting his "new feature test" had gone wildly past expected scope.
@fta7 [Claude Code]
https://x.com/fta7/status/2060312733438853424
A creative writing workflow using Claude Code: type romaji on a keyboard (he can't read his own raw output), have an Obsidian-displayed file get continuously appended with the AI's Japanese translation, scrolling alongside his typing. The benefit is that not being able to re-read your own draft kills the "is this word right" tangents that stall writing. Long-form, messy, dense romaji input produces better AI translation than carefully-typed short Japanese, because context is richer.
@tomcrawshaw01 [Claude Code]
https://x.com/tomcrawshaw01/status/2060373159162126811
Last month's Claude Code usage: $2,123.53 of metered value over 2.38B tokens across 611 sessions. He paid $100 (Max plan, flat fee). The gap is the whole opportunity — a 21x usage-to-price gap that won't last forever, but right now is the cheapest serious AI leverage available.
@mardehaym [Claude Code]
https://x.com/mardehaym/status/2060394825606729994
Runs a 2-person AI-augmented engineering team at $200/dev/month. 330 PRs shipped in 6 months, 90% AI-generated. Every token traced to a ticket, every ticket traced to shipped code. The argument against Tokenmaxxing-by-leaderboard isn't "spend less" — it's "make every token answer to a shipped artifact."
@linglingfa [Claude Code/OpenClaw]
https://x.com/linglingfa/status/2060268840496566767
After two years cycling through Cursor, Antigravity, Claude Code, Codex, OpenClaw — settled on: one Codex on the local machine, one Codex CLI in the cloud, two OpenClaw on separate VPSes. Four "digital employees" he has to manage daily. Honest about the friction: they're not proactive enough (and when they are, he's scared), so most of his time goes into planning their work and writing specific requirements.
@cjzafir [Claude Code]
https://x.com/cjzafir/status/2060417431324311977
Workflow that's been hitting 10/10: Codex 5.5 Extra High for the plan, Opus 4.8 Extra for the execution. Codex plans brilliantly but overcomplicates execution; Opus 4.8 plans loosely but executes without lazy investigation. The trick is asking Codex to write a detailed execution prompt for Claude Code — Codex puts in extra effort and Opus doesn't waste tokens on silly things. Running 3-4 sessions at once and still hasn't touched 50% of his 5-hour limit.
@dunik_7 [Claude Code]
https://x.com/dunik_7/status/2060457079760945631
Argues the actual upgrade buried in Opus 4.8 is effort control plus the cheaper /fast mode plus dynamic workflows — used right, they cut Claude Code bills ~50% without touching output quality. Recipe: Low effort on the 60% of prompts that are "format this" or "what does this return," High on daily coding, Max on the 10% that's real architecture, Fast mode (now 3x cheaper) for big mechanical refactors. His math: ~$400/month at defaults → ~$200/month routed correctly.
@ericciarla [Claude Code]
https://x.com/ericciarla/status/2060375920968696312
Wired up Firecrawl /monitor with Claude Code in under 60 seconds to email him whenever any artist he likes goes on tour in his city. The agent stays asleep until the page actually changes. Useful demo of the event-source-for-agents primitive.
@Voxyz_ai [Claude Code/Hermes/OpenClaw]
https://x.com/Voxyz_ai/status/2060425284566778010
The pattern: his cron loops on Hermes and OpenClaw kept forgetting each other because each fresh session had no clue what the last one did. He was hand-carrying state between them. So he stopped — each loop now leaves the next a handoff note: where things stand, what to do, when to call him. Inbox triage, content drafts, support replies, morning digest, all pass the same note. A few weeks in, the AI crew relays itself for days.
@Abdulaziz_Hmadi [OpenClaw]
https://x.com/Abdulaziz_Hmadi/status/2060465709466206267
Says 90% of his work producing applications and digital services runs through OpenClaw because it lets him use GPT, Claude, Gemini and dozens of other models from a single window, with all projects and conversations in one place — instead of separate apps like Cursor. Runs locally on Mac and Windows. Treats OpenClaw as a stage you have to learn if you use AI seriously.
@DanKornas [Claude Code]
https://x.com/DanKornas/status/2060311678084931967
Released book-to-skill, a Claude Code skill that turns a technical book or document into a structured skill. Generates a SKILL.md, per-chapter markdown files, glossary, patterns, and cheatsheet from PDF/EPUB/DOCX/TXT/Markdown/HTML/MOBI. On-demand chapter loading instead of stuffing the whole book into context. Use case: query a book from inside Claude Code with /your-book-slug commands.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2060370121013494053
Built a Claude Code Usage VS Code extension that gives token counts, costs, sessions, and projects directly inside VS Code. Reads local JSONL transcripts, no API calls. The "Claude subscribers get a progress bar; this gives you the full picture" framing nails the gap — official UI shows you the speedometer, not the fuel gauge.
@nosp321 [Claude Code]
https://x.com/nosp321/status/2060236390504309023
Pipeline for faceless YouTube data-ranking videos (countries, facts, comparisons) using Claude Code + Remotion for motion graphics + ElevenLabs for voice. Each video 5-10 minutes of work. Channels in this niche are already pulling 400-700K subscribers. Operating cost: ~$150-200/month. Not "press a button" easy, but the system seriously speeds up production.
@0xDeliriumm [Claude Code]
https://x.com/0xDeliriumm/status/2060466314607530085
Use case for the "Obsidian Mine" skill: connects Claude Code directly to Obsidian vaults so Claude updates them in real time and pulls from them in real time. Becomes a bridge between your knowledge base and your AI — Claude reads your vault before starting and already knows your context, projects, and history. Stops being a "fancy folder system."
@kevtunis [Claude Code]
https://x.com/kevtunis/status/2060422536228397420
Production Run #1 of a daily news pipeline: pulled signals from Scoble's Aligned News network, automatically selected stories, generated a report, and produced a video — all through Claude Code. The real test is what comes next: running every day without intervention.
@MinLiBuilds [Claude Code]
https://x.com/MinLiBuilds/status/2060282533506601323
Built a Claude Code subtitle skill, then used it to caption Boris Cherny's latest talk on "AI directing AI" in 8 languages — English, Japanese, Spanish, Arabic, Turkish, French, Korean, Chinese. A working illustration of the talk's own thesis: one skill, many languages, one operator.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2060239259873788229
Eight-hour build of a personal Notion clone — just the parts the builder actually uses, stripped of everything else, running on his own machine. Whole thing runs on two tools: moonchild for design (brief through final visual), Cursor + Codex for code. The actual trick: 8 markdown files committed before a single line of code — principles, design library, flow, spec, style, architecture. Cursor reads style.md and architecture.md directly via MCP, so it doesn't have to be told what the fab looks like. One hour for the backend.
@gengdaJ [Claude Code/OpenClaw]
https://x.com/gengdaJ/status/2060295835230073300
Pitch on offline B2B AI consulting: line up Obsidian + Claude Code + Codex + OpenClaw + WeChat data parsing + workflow automation + digital-twin building, and the big-account work follows. Concrete workflow he uses for B2B meetings: record (Tongyi Tingwu on PC, Yuanbao on phone, Feishu Recording Bean for hardware), use Step ASR for transcription (much cheaper than Doubao), then Codex summarizes the key points into Obsidian for analysis. The pattern: cheap good ASR + structured note system + AI assistants = sellable B2B service.
@stevekrouse [Claude Code]
https://x.com/stevekrouse/status/2060410355679961299
Had to fax the IRS this morning. Claude Code got him 90% of the way there — even filled out the PDF — but couldn't actually send the fax. So he had Claude Code find a fax API, wrap it in a Tempo proxy, and deploy on Val Town. Two hours end-to-end, while working on other stuff. Now there's agentfax.
@hyuki [Claude Code]
https://x.com/hyuki/status/2060474910691229937
The hardware setup: 24-inch vertical display with Ghostty fullscreen, 2-4 tabs of Claude Code running in parallel. Likens it to Shogi multi-board play. Has built a personal "AI assistant Poppy-chan" via Claude Code that featured in his newsletter.
@obscaries [Claude Code]
https://x.com/obscaries/status/2060217982421549193
Released Claude-BugHunter — a structured bug bounty workflow pack built for Claude Code. The agentic-security-research category was barely a category six months ago. Worth tracking what the playbook looks like as it gets standardized.
@SydneyDaddy1 [Claude Code]
https://x.com/SydneyDaddy1/status/2060211532898947240
Opus 4.8 called via API in Chinese answers "I am Qwen" when asked "what model are you." Doesn't happen in Claude chat or Claude Code; only via raw API. He tested French, Japanese, Spanish — same self-identification pattern. Pre-4.8 Opus models mostly didn't do this. Sonnet 4.6 sometimes claims to be DeepSeek. The training-data fingerprint is now showing up at API level.
@peter_szilagyi [Claude Code]
https://x.com/peter_szilagyi/status/2060400404647596467
Real-world friction report on the new dynamic workflows: Claude Code now interprets the word "workflow" in a prompt as a multi-agent orchestration trigger. He's getting tired of telling Claude "no, I really mean the word workflow." Anthropic added a kill switch in 2.1.157 — "Workflow keyword trigger" setting in /config — within a day, which says how universally annoying the auto-trigger was.
@daisuke_wakui [Claude Code]
https://x.com/daisuke_wakui/status/2060471404051705863
Reflection from a consultant: 21 parallel CC agents kills the "collect-info-then-analyze-for-money" consulting business. Future consulting work is to let AI handle overwhelming divergence and spend all the freed-up time on the messy human conversations needed to make the output land. Consulting moves from research-for-pay to translation-for-pay.
@NFTCPS [Claude Code]
https://x.com/NFTCPS/status/2060187041074430111
Pointer to a GitHub repo that lets users run Claude Code for free by redirecting traffic to 10 free domestic Chinese providers (DeepSeek, Kimi, etc.) — 5 minutes to configure. Around 20,000 developers already using it. The angle: cheap models filling in for the official Anthropic backend until the gap closes.
🗣 User Voice
User Voice
@peter_szilagyi — "Workflow" keyword auto-trigger is annoying when "workflow" actually means a normal workflow. Anthropic responded within a day with a config flag, but the pattern (helpful feature that surprises users) keeps repeating.
@wshuyi / @Jeremybtc / @yoshio_nocode — Dynamic Workflows are wildly powerful but burn quotas in minutes. Users want clearer pre-flight token estimates, hard budgets per workflow, and a "show me the plan before you spawn 100 agents" first-time gate that's actually load-bearing, not just informational.
@chapati23 — Asks for an LTS track for OpenClaw because every release breaks his agent. Generalizes: power users running production workloads want a stable channel separate from the rapid release channel.
@hkarthik — Handing every employee raw Claude Code is going to keep producing mixed results in companies with many non-technical users. Purpose-built shared agents with optimized prompts and centrally-managed token usage are the path forward; power users will stay on direct harnesses, but everyone else needs a paved path.
@dunik_7 / @mardehaym — Cost discipline is a real differentiator. Effort routing, fast mode, per-ticket token tracing, and a flat-fee Max plan can swing 2-3x value with the same workflows. The CFO-on-Claude question is now table stakes.
@MLBear2 / @MrAhmadAwais — Opus 4.8 has visible behavioral regressions in some setups: stopping mid-task, hallucinating other agents on the same repo, occasional rm -rf .git attempts on auto. Users want a way to pin a model version they trust until the next stable lands.
@peter_szilagyi — "Workflow" keyword auto-trigger is annoying when "workflow" actually means a normal workflow. Anthropic responded within a day with a config flag, but the pattern (helpful feature that surprises users) keeps repeating.
@wshuyi / @Jeremybtc / @yoshio_nocode — Dynamic Workflows are wildly powerful but burn quotas in minutes. Users want clearer pre-flight token estimates, hard budgets per workflow, and a "show me the plan before you spawn 100 agents" first-time gate that's actually load-bearing, not just informational.
@chapati23 — Asks for an LTS track for OpenClaw because every release breaks his agent. Generalizes: power users running production workloads want a stable channel separate from the rapid release channel.
@hkarthik — Handing every employee raw Claude Code is going to keep producing mixed results in companies with many non-technical users. Purpose-built shared agents with optimized prompts and centrally-managed token usage are the path forward; power users will stay on direct harnesses, but everyone else needs a paved path.
@dunik_7 / @mardehaym — Cost discipline is a real differentiator. Effort routing, fast mode, per-ticket token tracing, and a flat-fee Max plan can swing 2-3x value with the same workflows. The CFO-on-Claude question is now table stakes.
@MLBear2 / @MrAhmadAwais — Opus 4.8 has visible behavioral regressions in some setups: stopping mid-task, hallucinating other agents on the same repo, occasional rm -rf .git attempts on auto. Users want a way to pin a model version they trust until the next stable lands.
📡 Eco Products Radar
Eco Products Radar
Dynamic Workflows (Anthropic) — the headline ship. Up to ~100-1000 parallel subagents in one session, JS orchestration scripts, /effort ultracode trigger. Spawned dozens of in-production stories on day one.
ego lite — Chromium-based browser purpose-built for letting Claude Code / Codex / Cursor run in isolated Spaces with inherited Chrome state. Heaviest promo of the day; appeared 10+ times.
Step 3.7 Flash (Stepfun) — 198B MoE, ~11B active, 256K context, 400 TPS, Apache 2.0, native multimodal. Targets agent inner loops as a cheap and fast tier. Multiple deployments and OpenRouter routing announcements.
Firecrawl /monitor — agent-native change-detection webhooks. The "event source for agents" primitive. Featured in real workflows (artist-tour notifier, codebase-watch).
Obsidian — recurring substrate for personal-knowledge skills. Pairs with Claude Code via skills like Obsidian Mine, claude-mem, and direct vault reads.
Hermes Agent — the cross-runtime competitor. Velocity Release shipped this week with Kanban, MCP catalog, OpenHands orchestration. Appears alongside OpenClaw in many multi-agent setups.
Superpowers / book-to-skill / React Doctor / taste-skill / frontend-design — the skill ecosystem keeps cementing: structured engineering loops, technical-book ingestion, React anti-pattern scans, anti-AI-slop frontend constraints, all distributed as drop-in skills across Claude Code, Codex, Cursor, OpenCode.
ClaudeCodeLog — the unofficial changelog account is now load-bearing infrastructure for keeping up with CC's daily-shipping release cadence.
KiloCode / OpenCode / Hermes / Pi — the multi-harness world is now real. Step 3.7 Flash, Grok, Kimi, etc. ship with first-class compatibility lists across them.
claude-mem / Memanto — persistent memory across sessions remains an unsolved layer; multiple competing skills are racing to own it.
OpenRouter — the routing layer is everywhere; Series B announcement still echoing through every multi-model setup.
Dynamic Workflows (Anthropic) — the headline ship. Up to ~100-1000 parallel subagents in one session, JS orchestration scripts, /effort ultracode trigger. Spawned dozens of in-production stories on day one.
ego lite — Chromium-based browser purpose-built for letting Claude Code / Codex / Cursor run in isolated Spaces with inherited Chrome state. Heaviest promo of the day; appeared 10+ times.
Step 3.7 Flash (Stepfun) — 198B MoE, ~11B active, 256K context, 400 TPS, Apache 2.0, native multimodal. Targets agent inner loops as a cheap and fast tier. Multiple deployments and OpenRouter routing announcements.
Firecrawl /monitor — agent-native change-detection webhooks. The "event source for agents" primitive. Featured in real workflows (artist-tour notifier, codebase-watch).
Obsidian — recurring substrate for personal-knowledge skills. Pairs with Claude Code via skills like Obsidian Mine, claude-mem, and direct vault reads.
Hermes Agent — the cross-runtime competitor. Velocity Release shipped this week with Kanban, MCP catalog, OpenHands orchestration. Appears alongside OpenClaw in many multi-agent setups.
Superpowers / book-to-skill / React Doctor / taste-skill / frontend-design — the skill ecosystem keeps cementing: structured engineering loops, technical-book ingestion, React anti-pattern scans, anti-AI-slop frontend constraints, all distributed as drop-in skills across Claude Code, Codex, Cursor, OpenCode.
ClaudeCodeLog — the unofficial changelog account is now load-bearing infrastructure for keeping up with CC's daily-shipping release cadence.
KiloCode / OpenCode / Hermes / Pi — the multi-harness world is now real. Step 3.7 Flash, Grok, Kimi, etc. ship with first-class compatibility lists across them.
claude-mem / Memanto — persistent memory across sessions remains an unsolved layer; multiple competing skills are racing to own it.
OpenRouter — the routing layer is everywhere; Series B announcement still echoing through every multi-model setup.
Comments