Super User Daily: 2026-06-01
Saturday's data is dense with people running Claude Code as a business, not as an autocomplete. The pattern that keeps coming back: stop watching the terminal, set up the loop, let it ship while you sleep. A Spanish dev built a 7-agent landing-page agency that bills 47 clients a month from one MacBook. An Argentine lawyer turned Claude Code into a 14-branch legal system with 20 MCP connectors hitting BORA, SAIJ, CSJN. A miHoYo team lost $2M in a single night to runaway agent loops. The same product is making and burning money depending on who's holding it. Opus 4.8 and Dynamic Workflows landed in the middle of all this, and the people who built guardrails are pulling away from the people still pasting into one chat box.
@robiartec [Claude Code]
https://x.com/robiartec/status/2060800193029935464
A Chinese dev built a 7-agent system on Claude Sonnet 4.6 that sells landing pages to local businesses. The orchestrator runs Scout, Diagnoser, Builder, Filmer, Pitcher, Checker plus a Mobile agent on his iPhone, all coordinated through Claude Code Router and sharing state via the filesystem. About 3M tokens a day, ~$480/mo in API, 47 clients at $400 each, monthly revenue around $18,800. The agents only wake him up if a deal is above $3K or response rate drops below 12%.
@abogadoaboitiz [Claude Code]
https://x.com/abogadoaboitiz/status/2060514511136641247
An Argentine lawyer turned a single text file into a 14-branch legal system inside Claude Code. Each profile loads the relevant code, the markers for that branch, and triggers alerts. He wired 20 MCP connectors to BORA, SAIJ, CSJN, JusBaires, PJN, PTN, TFN, plus macOS automation for portals with no API. The system handles labor telegrams post-reform, contencioso administrativo across 13 provinces, and refuses to invent rulings — every gap is flagged with what's missing to resolve it.
@MushtaqBilalPhD [Claude Code]
https://x.com/MushtaqBilalPhD/status/2060748490272395453
An academic dropped the bibliography of his published paper on Danish fairy-tale colonial reception into a folder — sources in Danish, English, and Bangla — opened Claude Code, used Opus 4.8 on Max with two theoretical models, and asked it to write an original argument citing across all three languages. Hit rate limits in an hour, burned ~$12 in extra credits, then produced a Word draft with cross-language connections that he says is publishable with minor revisions.
@bridgemindai [Claude Code]
https://x.com/bridgemindai/status/2060693626712141885
Pushed 1,628,985,347 tokens through Claude Code in two days using Opus 4.8 in ultracode mode, 6-12 agents running in parallel live on stream. Equivalent API spend ~$1,500. He was paying a flat Max plan fee. Called it the arbitrage nobody's talking about — the gap between metered API cost and subscription cost when you actually run the agent loop hot.
@Atenov_D [OpenClaw]
https://x.com/Atenov_D/status/2060659782390804780
"Felix" is an OpenClaw agent running on a Mac Mini with two Claude subs, $400/mo total. The owner sends voice notes through Discord, Felix transcribes and executes. Felix has three memory layers (daily notes, full KB rebuild at 2am, 30-minute heartbeat to restart dead tasks) and hired two other agents (Irisbot for support, Remybot for sales) which it reprograms nightly. Month one: $80,000 reported. Owner doesn't touch a keyboard.
@LandseerEnga [Claude Code]
https://x.com/LandseerEnga/status/2060861825638666594
Pointed Claude Code at a live iOS device on revyl and said "test everything." It read the app, recognized four sub-apps (rides, delivery, services, account), split itself into 4 agents on 4 parallel cloud devices, and drove the UI like a human against expected screens. No xcuitest scripts, no element IDs, nothing pinned to brittle selectors. The instruction was three words.
@masayan_ai_hack [Claude Code]
https://x.com/masayan_ai_hack/status/2060872395033088303
Ran Claude Code dynamic workflows with 39 sub-agents in parallel and produced 470 short videos in 48 minutes. Used ~40% of a Max 5x plan. The post is almost an aside — the casual scale is the point. One person, half an hour, almost five hundred videos.
@sairahul1 [Claude Code]
https://x.com/sairahul1/status/2060647386784137539
Someone vibe-coded an outbound prospecting tool in 2 weeks. You type a business type and city, it scrapes Google Maps for 30+ data fields, fetches verified emails and socials from each website, reads up to 50 reviews per business with AI to extract pain points, cross-references those with your offer, and writes personalized cold emails. Leads land in a GPS-mapped CRM with route optimization. Entire build inside Claude Code.
@draprints [Claude Code]
https://x.com/draprints/status/2060706800807317961
"Replaced an entire outbound agency with one Claude Code build." Client fills out a form → 20 minutes later a full campaign is live across thousands of verified leads. The system scrapes Google Maps across 200+ US cities for the ICP, runs Apollo for decision-makers by title, mines a third DB for emails, dedupes across all three, real-time-verifies every address, then writes 4 spintaxed copy variants per step and pushes through pre-warmed mailboxes. Five verticals running simultaneously.
@wolfejosh [Claude Code]
https://x.com/wolfejosh/status/2060837766603128869
Used Claude Code to ingest his heart rate, sleep, workout/step data and vitals from his watch, then correlate the spikes against his meetings and calendar. The annotated chart shows what triggered stress — which meeting, which call, which board. Not health-tech-product-built-by-VC-bros, just a dev with a CSV, a calendar export, and a prompt.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2060730707283169327
A workflow leaked on Reddit: use Claude Code's /schedule to run persistent cloud tasks, connect Brave Search MCP for real-time intel, mount your Obsidian vault via filesystem MCP, write a CLAUDE.md with focus areas and "do NOT want" filters, then tell Claude "6am every weekday, search developments, save brief to vault." Claude converts the English into a cron job. Machine doesn't need to be on. The author calls it "your personal analyst."
@Chain_GPT [Claude Code]
https://x.com/Chain_GPT/status/2060690435681772003
Built "Market Mood Ring" — a live NFT that re-paints itself based on crypto market sentiment — in one afternoon using the ChainGPT skill for Claude Code. The skill wrote the server, called the news API, ran sentiment scoring through ChainGPT's LLM, generated the matching artwork, and self-repaired when something broke. The builder mainly supplied the idea and final polish.
@AnatoliKopadze [Claude Code]
https://x.com/AnatoliKopadze/status/2060732744439644579
Spotify's recap: one background Claude Code agent merged 1,000+ PRs a month and cut migration time by 90%. Head of Claude Code's frame from the talk: agents don't fail because they're dumb, they fail because instructions are vague. Write the routine, describe the outcome, let it cook. Stop prompting back and forth — let Claude prompt itself.
@DataChaz [Claude Code]
https://x.com/DataChaz/status/2060645758295228423
Anthropic shipped Dynamic Workflows. The proof point already exists: Jarred Sumner ported Bun from Zig to Rust — 750,000 lines, 11 days, 99.8% test pass rate — using hundreds of Claude agents working in parallel with two reviewer agents per file. Work that historically takes a team a quarter, compressed into a week and a half. Activation: include "workflow" in your prompt or set /effort ultracode.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2060753351978184956
Arize's CPO walked into Claude Code with an empty directory and in under 45 minutes had a PM agent pulling 40 GitHub discussions, 60 issues, 8 releases, scoring every one by priority, then evaluating its own accuracy and feeding corrections back on a cron. Four asks in a terminal: build, instrument, suggest an eval, run the loop. The agent caught bugs being systematically underscored on its first pass — drift that normally takes weeks of manual backlog review to surface.
@nicbstme [Claude Code]
https://x.com/nicbstme/status/2060754371156361275
His personal-life orchestrator: Codex or Claude Code as the orchestrator with gogcli for Gmail/Drive/Calendar/Docs/Sheets, wacli for WhatsApp, imsg for iMessage and SMS, browser automation for Chrome, AppleScript for macOS apps, and filesystem tools for local files. Says he's nervous about running an OpenClaw-style 24/7 agent on his personal life given current security attacks, so he stays in control. Centralized everything (medical results, kid info, car) in Google Drive so the agent has one source.
@ericosiu [Claude Code]
https://x.com/ericosiu/status/2060787867673379088
Ten "highest-leverage" Claude Code dynamic-workflow prompts for revenue generation, each spawning specialist subagents across pipeline, retention, upsells, SEO, paid acquisition. The Revenue Command Center one alone: inspect CRM exports, call transcripts, email threads, analytics, proposals, ad data, internal docs; produce 25 ranked opportunities with upside, owner, next action, 7-day execution plan; then implement safe automations and dashboards in parallel.
@nicbstme [Claude Code]
https://x.com/nicbstme/status/2060566811150798868
Every Friday he uses Claude Code with skills to triage and answer WhatsApp, SMS, and email. Tried the same workflow with Opus 4.8 — couldn't replicate what he was getting from 4.6. Switched to the Codex app: significantly faster and better quality. Says it's not the harness, it's GPT-5.5. Honest comparison from someone running the workflow weekly, not benchmarking.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2060773259746935232
Side-by-side Dynamic Workflows comparison. Codex: same task burned 175K tokens in 15 minutes, 99% of his weekly Pro quota still left. Claude Code: 700K tokens in 40 minutes, 86% left on a Team plan. Claude Code hit rate limits mid-task and had to pause. Same task, 4× the tokens, slower wall-clock. He still loves Claude — but the usability gap on dynamic workflows is real.
@shao__meng [Claude Code]
https://x.com/shao__meng/status/2060554568849907945
Salesforce engineering's writeup on Agentic adoption. The 33 APIs / 231 person-days → 13 days case is real but specific to "rule-formable + auto-verifiable" tasks. The bigger lever: rules-as-code (markdown rules + reference implementations + PR feedback written back into rules) compounds precision instead of re-prompting from scratch. They consolidated everyone on Claude Code, removed the token cap, and turned PR comments into a feedback loop into the rule library.
@KingBootoshi [Claude Code]
https://x.com/KingBootoshi/status/2060626640884171132
Stopped writing things into docs and CLAUDE.md files — instead pointing agents at their own chat history. Every Claude Code / Codex log is a chronological story of what he's been working on. Now telling agents to grep their own history before starting new work. Says it's the best memory system he has on him right now and asks how others are indexing this.
@dotey [Claude Code]
https://x.com/dotey/status/2060865571244183797
On running multiple agents together — Codex, Claude Code, Cursor, GitHub Copilot, each strong at different things. Matt's Sandcastle uses TypeScript scripts to orchestrate workflows across all of them inside a VM. Too geeky for everyday use but right for extreme cases — cyber gu-shocking: write a tech spec, have each agent produce a version, then score and refine each other's.
@guansi [Claude Code]
https://x.com/guansi/status/2060716603214368825
Multi-stack local agent setup. Hermes runs two agents: GPT 5.5 for hard dev, DeepSeek V4 Pro for meeting minutes and docs. He has DeepSeek compress context for GPT 5.5 because GPT 5.5 alone blows context constantly. Hooked Claude Code in so Hermes can call it for code edits, then compares output against GPT 5.5 direct edits. Says he's stopped caring which model is strongest and started caring how to split tasks, control context, and pick the cheapest worker per job.
@lepadphone [Claude Code]
https://x.com/lepadphone/status/2060793663240946047
Opus 4.8 critique from a working user: excellent for code review and security auditing, terrible for actual project development. Defaults to heavy over-negation — labels reasonable ideas "fundamentally flawed" or "dead on arrival" — instead of iterating with you. Strong on raw intelligence, weak on collaboration. Not a launch take, a multi-day usage take.
@AlexFinn [Claude Code]
https://x.com/AlexFinn/status/2060538955096097134
After 24 hours of testing Opus 4.8 nonstop, his actual workflow: give the same hard problem to both Opus 4.8 and Codex (GPT-5.5), have each build a plan, then hand each plan to the other agent. Opus 4.8 plans ended up better almost every time. So he uses Opus 4.8 to write plans for the hardest problems, then hands them to Codex to execute. Codex remains the daily driver. Total reversal from a month ago.
@cyber_cat7 [Claude Code]
https://x.com/cyber_cat7/status/2060611224673648714
Translation of a Chinese builder's Shopify dropship operation: 13 Claude Code agents spread across a 3×2 monitor grid and a vertical second monitor, each working a different layer of a store — catalog rewrite, homepage layout, cart and checkout per country, 30-email warmup sequence, banners, A/B tests, analytics setup. 13 stores running in parallel, daily token cost under $80 across all of them, sells finished stores at $800 each, 6-7 stores a day. A 14th agent runs a Polymarket arbitrage wallet on the side. One desk, no employees.
@stacyonchain [Claude Code]
https://x.com/stacyonchain/status/2060697734160556091
Replaces 5-10 paid productivity apps with one Obsidian vault and Claude Code. No subscriptions, full data ownership. The whole 18-month system — structure, plugins, templates, workflows — runs as a living self-updating second brain because Claude Code can read/write the vault directly. Cyril's setup, recommended as the cleanest he's seen.
@polydao [Claude Code]
https://x.com/polydao/status/2060810164068856130
Boris Cherny detail from the keynote: 80% of Anthropic engineers use Claude Code daily. Two specific tactics worth stealing: pipe massive GCP logs or git status into Claude Code via SDK with -p flag and parse the JSON output, and use git worktrees to run 5 parallel Claude sessions on the same repo at once. Also: # before a message permanently saves to CLAUDE.md. Ctrl+R shows the exact context window the model currently sees.
@andersonlimadev [Claude Code]
https://x.com/andersonlimadev/status/2060846728660545635
DeepSeek V4 Pro driving Claude Code cleaned every vulnerability in his project for $0.27 total — about R$1.35. The point isn't the savings, it's the proof that a serious agent loop on a non-Anthropic model now costs change. He's calling out devs who still won't try the alternate-provider route.
@MerlijnTrader [Claude Code]
https://x.com/MerlijnTrader/status/2060745548995420344
The macro tell: Microsoft cancelled Claude Code licenses by June 30. Uber exhausted its full 2026 AI budget in 4 months. H200 GPU rentals dropped 40% in 3 weeks. Uber's COO on AI value: "That link is not there yet." The buyers are pulling back at the same moment the buyers' bosses are doubling down on stock. One side is wrong.
@mubeitech [OpenClaw]
https://x.com/mubeitech/status/2060866497300361641
Translation of a Chinese postmortem on miHoYo burning $2M overnight to a multi-agent loop. Several dozen agents on a multi-agent system kept handing hallucinations to each other as ground truth, each turn growing the context by 16× by turn 30. No step cap, no cost ceiling, no loop detection. Compared to OpenAI's 4-agent ping-pong that burned $47K in 11 days zero output, and an OpenClaw experiment with 100 Codex agents that hit $1.3M in 30 days. The fix afterward: cost controls, context pruning, loop detection — but the lesson is that any system without those three is a time bomb.
@mstockton [Claude Code]
https://x.com/mstockton/status/2060741863221657883
Compound engineering in practice. Weekly GitHub Action runs Claude Code in non-interactive mode (-p) with custom skills that inspect his repos, look at recent changes, and generate three markdown files: architecture overview, recent code changes summary, cross-repo system view. Those markdown files deploy into a repo where a Claude plugin can pick them up later. Then an extra step turns the markdown into a self-contained HTML/CSS/JS site you can click around. The system distills context once so future agents have a clean map instead of re-deriving everything.
@sumika45379 [Claude Code]
https://x.com/sumika45379/status/2060564357688521210
Connect Claude Code to the Shopify CLI and the workflow becomes: Claude reads every product in the store, rewrites all the descriptions, and pushes the new copy back live — all from the terminal, one instruction. The more products you have, the higher the leverage. Translation of a Japanese post that flags how many small Shopify operators are still doing this manually.
@pauhei_saunner [Claude Code]
https://x.com/pauhei_saunner/status/2060712010158158008
A non-engineer HR person built a recruiting scout tool in Claude Code. The hard part wasn't writing code — that was the easy day. The hard part was language-extracting his own judgment criteria: which sections of which platforms to read, how to set search filters, what makes a profile worth a message, where the "send/don't send" line is. Ten days on language extraction, then Claude Code built the working tool in under a day. Scout time dropped to roughly 20-25% of what it was.
@ai_xiaomu [Claude Code]
https://x.com/ai_xiaomu/status/2060714594583343518
Bare-bones CRM for freelancers using Obsidian and Claude Code: one markdown file per client with contact, scope, price, progress, payment log. Then ask Claude things like "which clients haven't received delivery," "what was the requirement change from Zhang San last week," "how much did I collect this month." Claude aggregates across all the markdown files. Lighter than any CRM software because there is no CRM software.
@IAmAaronWill [Claude Code]
https://x.com/IAmAaronWill/status/2060689597680537648
Replaced $6,247/month in SaaS subscriptions with AI-assembled alternatives. 42 swaps across CRM (Salesforce → Notion + Claude scoring), email (ConvertKit → Loops), automation (Zapier → self-hosted n8n), scheduling, copywriting, scraping, lead scoring, proposals, project management, customer support, design, meeting notes, invoicing. The lever is Claude Code generating proposals, scoring leads, building chatbots, replacing wrappers. $74,964 annualized.
@AiSparks12 [Claude Code]
https://x.com/AiSparks12/status/2060590761733652711
The Opus 4.8 effort-level breakdown people are missing. Six tiers: Low (fastest, lookups), Medium (summaries, edits), High (default, deep reasoning), XHigh (long async workflows), Max (research and strategy, full budget), Ultracode (max + dynamic workflows in parallel). Leaving everything on Max burns the limit in hours. Wrong level = wasted tokens. Match effort to task or pay for it.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2060594556778029541
What separates Claude Code users who get results from those who don't, in one sentence: the difference isn't prompting, it's whether the agent remembers anything. Three concrete actions: a file the agent reads on start and rewrites on exit, a single filter for what gets stored ("does this change next run's behavior?"), and a weekly background job that cleans memory while you sleep. Memoryless agents have a hard ceiling — they're the same on the 100th run as on the first.
@miroburn [Claude Code]
https://x.com/miroburn/status/2060728424897880435
A startup credit clearinghouse: 55 programs totaling $2.4M in credits curated, plus a new Claude Code / Codex skill that lets your agent mass-apply to matching programs using your own company context. Sounds dry until you realize this is precisely the kind of bureaucracy-language work Claude is good at, and the alternative is two interns and a spreadsheet.
@0xDeliriumm [Claude Code]
https://x.com/0xDeliriumm/status/2060783276684444116
Built a visual graph of every Claude and ChatGPT conversation he's ever had — every project, idea, conversation becomes a node, bigger circles for topics he talks about more, 1,000+ connections. When he starts a new Claude Code session, he points it at older projects and it pulls relevant history in seconds. Auto-updates after each project so the next session starts smarter than the last.
@kawai_design [Claude Code]
https://x.com/kawai_design/status/2060843697013469351
A series of clean explainer cards for Claude Code basics by analogy: terminal = workbench, git = work journal. The git Vol.7 specifically says don't stop at "make the change" — ask Claude to inspect the diff, write the commit message explaining intent, and prep the PR. That's the practical version of the "Claude as work journal" model: every commit ends up readable by the next person in the room.
🗣 User Voice
User Voice
@lepadphone — "It shuts down ideas instead of building on them." Opus 4.8 is great for review but punishing in iteration. People want strong reasoning without the smackdown energy.
@daniel_mac8 — Dynamic Workflows on Claude Code burn 4× the tokens of equivalent runs in Codex and hit rate limits mid-task. Users want effort levels and Dynamic Workflows that don't blow the weekly budget on one prompt.
@godofprompt — "/goal is designed for even longer autonomous runs, and there are no built-in spending caps tied to goal completion." Users want spending guardrails baked into long-running modes, not as an afterthought.
@ClaudeCode_UT — "Memoryless agents have a hard ceiling." Users want memory that survives sessions without manual prompt engineering — file-based, simple, audited.
@dotey — "Claude Code's TUI was so successful that they ignored the GUI." Users want a desktop app that doesn't feel like an afterthought next to the terminal.
@kevinnbass — "The parallel tool use bug in Claude Code with Opus 4.8 has created a lot of hallucinations and made it dangerous to use with my codebase and database." Users want reliability on the basics before chasing more model intelligence.
@shrav_10 — Users hit monthly limits within 48 hours of Opus 4.8 dropping. The new model isn't cheap to actually run.
📡 Eco Products Radar
Eco Products Radar
Codex / Codex App / Codex CLI — the consistent comparison partner; people running both side by side
Hermes Agent (Nous Research) — referenced as the orchestrator running multiple Claudes underneath
OpenClaw — the "always-on" agent runtime, both celebrated and feared
Obsidian — the file-based memory store of choice when paired with Claude Code
DeepSeek V4 Pro — the price-arbitrage model people are routing through Claude Code via cc-switch
Cursor — still the IDE for file-level edits while Claude Code handles features
HyperFrames — the video-generation engine paired with Claude Code in real-estate and content pipelines
n8n — self-hosted automation that replaces Zapier in Claude Code stacks
OpenRouter — the routing layer for free / low-cost models in proxy setups
Step 3.5 (StepFun) — Chinese flat-fee model routed into Claude Code via config
ChainGPT — crypto news skill making one-afternoon NFT builds plausible
Grok Build (xAI) — newest CLI competitor showing up in side-by-side comparisons
Brave Search MCP — the real-time search backend powering /schedule morning briefs
revyl — cloud iOS device fleet that powers parallel mobile testing
Polymarket — the bet venue that keeps showing up in solo-agent finance stacks
Codex / Codex App / Codex CLI — the consistent comparison partner; people running both side by side
Hermes Agent (Nous Research) — referenced as the orchestrator running multiple Claudes underneath
OpenClaw — the "always-on" agent runtime, both celebrated and feared
Obsidian — the file-based memory store of choice when paired with Claude Code
DeepSeek V4 Pro — the price-arbitrage model people are routing through Claude Code via cc-switch
Cursor — still the IDE for file-level edits while Claude Code handles features
HyperFrames — the video-generation engine paired with Claude Code in real-estate and content pipelines
n8n — self-hosted automation that replaces Zapier in Claude Code stacks
OpenRouter — the routing layer for free / low-cost models in proxy setups
Step 3.5 (StepFun) — Chinese flat-fee model routed into Claude Code via config
ChainGPT — crypto news skill making one-afternoon NFT builds plausible
Grok Build (xAI) — newest CLI competitor showing up in side-by-side comparisons
Brave Search MCP — the real-time search backend powering /schedule morning briefs
revyl — cloud iOS device fleet that powers parallel mobile testing
Polymarket — the bet venue that keeps showing up in solo-agent finance stacks
Comments