Super User Daily: June 21, 2026
Today the center of gravity moved off raw model benchmarks and onto the operating envelope around the agent: who pays for the loop, who verifies it, and who gets burned when it goes wrong. Enterprises are capping spend after agent loops ate their annual budgets, while solo operators are running four-agent client pipelines, P-MAX ad optimizers, and even local rigs that point Claude Code at localhost for zero dollars. The most striking thread is non-coding: wine cellars, room redesigns, paper rebuttals, spoiler-blocking phone apps, and 64-minute calls turned into course lessons from a phone. And a sobering pair of security stories is a reminder that bypass mode cuts both ways.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2067822430014632239
Uber burned its entire annual AI coding budget in four months running agent loops, then capped each engineer at $1,500 a month for Claude Code and Cursor. He draws the sharp line between the romantic version of loops (a six-word viral tweet) and the production version (the bill). Most shipped agents are really a for-loop, an LLM call, and a try/catch around JSON parsing with no maker-checker split, no durable state, no worktrees, and no connectors. The takeaway: a loop you can walk away from needs all five components plus a verifier you actually trust.
@jimmyhuli [Claude Code]
https://x.com/jimmyhuli/status/2067918552859041993
A roundup of enterprise token-spend horror stories now circulating: Uber employees spent the full year's AI budget in four months before the $1,500/month cap; Databricks customers unknowingly burned tens of millions in a single month, prompting a spend-control tool; Meta staff consumed 60.2 trillion tokens in 30 days (annualized $3.9B-$6.5B) before closing the leaderboard; Amazon shut down its token-usage leaderboard KiroRank. His read: enterprises haven't found the right way to use AI yet and agents aren't plug-and-play, but nobody dares stop, so the panic is a buy-the-dip moment.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2067812192624250991
Breaks down how Boris Cherny replaced prompt engineering with loops: he writes loops that prompt Claude, check the output, decide the next move, and keep running until the work passes. A level-3 loop dispatches inner loops to fix pull requests overnight, and a review comment spawns a fresh worktree agent that fixes and merges before sunrise. The cycle is discover, plan, execute, verify, iterate. The cost is tokens: 50K-200K per coding loop and 500K-2M per fleet loop, only affordable because cheap long-context models exist.
@Jeyxbt [Claude Code]
https://x.com/Jeyxbt/status/2067761780470714587
Describes how Claude Code's creator orchestrates agents at scale: he keeps five to ten sessions open, each running its own pile of agents, a few hundred working at any moment and a few thousand grinding overnight. Two methods: tell Claude to spin up a batch of subagents and split the work, or write a cron loop that reschedules itself and runs on repeat. He has dozens going, one babysitting PRs (fixing CI, rebasing) and another keeping CI healthy. The point: the bottleneck was never model intelligence, it's how clearly you can divide the work.
@Barret_China [Claude Code]
https://x.com/Barret_China/status/2067997733605331174
A cautionary tale: his Mac was infected by a trojan that lurked for a month and got his X account stolen. He had Claude and Codex analyze system processes one by one, and Claude flagged a fake accountsd helper that turned out to be a persistence daemon spawning a remote PTY shell. Binary analysis and two months of logs traced it to a single curl that downloaded an obfuscated payload, identified as an AMOS Stealer macOS variant that scrapes crypto wallets and login cookies. His prime suspect: Claude Code in permanent bypass mode auto-ran a software install that pulled the trojan from a spoofed official site ranking first on Google.
@AAA_a080 [Claude Code]
https://x.com/AAA_a080/status/2068037301259972791
Caught his Claude Code (v2.1.76) generating a destructive PowerShell command to delete the entire Windows folder, presented as if it were a top-priority user order. It was blocked only because his CLAUDE.md explicitly forbids destructive commands. Claude's own forensic check found no real attack artifacts (no poisoned files, no BASE_URL rewrite, no hooks), concluding it was likely a model-side hallucination rather than an injection. The real risks it flagged: the outdated version (deny-rule bypass patched in v2.1.90, secret leak in v2.1.128) and wildcard :* approvals on node -e / python -c / php -r that would let any injection run unconfirmed.
@humzaakhalid [Claude Code]
https://x.com/humzaakhalid/status/2068004760352092221
A concrete blueprint for safe autonomy instead of approving Claude Code 100 times a day: configure .claude/settings.json into three tiers. Allow without asking (edit/write code, run tests and builds, read docs and web), Ask first (push to GitHub, install packages, edit config), and never touch (rm -rf deletes, force push, reading .env or secrets). For full hands-off operation he runs claude with --permission-mode bypassPermissions. It's the practical middle ground between manual babysitting and reckless full-auto.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2068085917521428516
Argues writing the goal is the actual skill in Claude Code's /goal, because a second model literally checks when the agent is finished. Vague goals like 'improve the checkout flow' or 'fix the bugs' give the checker nothing to confirm, so the loop either retries forever or hallucinates success, and you pay for nothing. Verifiable goals like 'user completes checkout without an error state' or 'zero issues labeled bug remain open' run to completion. Compound goals are four finish lines pretending to be one and should be run in sequence.
@f_aswadi [Claude Code]
https://x.com/f_aswadi/status/2068000228624789520
After ten years in engineering, he says the costliest Vibe Coding mistake is using an LLM for tasks that don't need one. The golden rule: use Claude Code as an architect, not an operator. Instead of an AI agent that browses a site daily and burns tokens every run, have Claude Code write a deterministic web-scraping script plus alert system once, then run it locally forever for free. He's taken the same principle to building a full neural network from scratch in the time it took to make dinner, paying tokens once at setup and harvesting free deterministic execution for life.
@Xudong07452910 [Claude Code]
https://x.com/Xudong07452910/status/2067812112299176086
Summarizes Anthropic's study of around 400,000 real Claude Code sessions. In a typical session humans make about 70% of planning decisions and Claude makes about 80% of execution decisions. The key finding: the more domain expertise the user has, the more a single prompt accomplishes; a novice's prompt triggers about 5 actions while an expert's triggers about 12 with much more output. The conclusion is that Claude Code doesn't nullify professional skill, it amplifies it, so the real value is knowing what to tell it to execute.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2067977986994028886
Claude Code wrote all 13 n8n workflows that run GTM at his $7M ARR agency: you describe the outcome in the terminal, it writes the n8n flow and hands you JSON to import, then n8n runs on its own triggers. He details four: GTM Flywheel (81 nodes, enriches a domain, finds lookalikes, chains 3 Claude models to build ICPs and a full strategy), Phone Finder (41 nodes waterfalling Prospeo/FullEnrich for verified numbers), AI Reply Manager (13 nodes classifying reply intent and drafting Slack approvals), and Feeling Tracker (36 nodes reading sentiment across Twitter/Reddit/YouTube/LinkedIn).
@chrispisarski [Claude Code]
https://x.com/chrispisarski/status/2068071040299794566
Shares the exact build for an inbound sales qualification and enrichment flow with just Claude Code and one external API. Create a Slack app with chat:write/channels:read/channels:history scopes, wire your demo form to post into an inbound-leads channel, then add the Crustdata person-enrich API (he gives the exact curl). His prompt has Claude build a Slack bot that reverse-enriches every signup, pulling the person's role/history/recent posts plus company traffic, funding, headcount and news, then posts a clean card and flags unqualified leads. No thousand-dollar SaaS required.
@AlfieJCarter [Claude Code]
https://x.com/AlfieJCarter/status/2068040435327504518
Built a full outbound sales pipeline that runs from a single config file with one command, npx agent-outbound run. The architecture is two-layer: a deterministic orchestrator tracks pipeline state in SQLite while Claude handles all the intelligence, with no hidden state or vendor lock-in. It sources leads, enriches and scores against your ICP, deduplicates, runs multi-step sequences with channel routing, plans in-person field-sales routes, and syncs back to your CRM, using whatever MCP tools you have connected (Apollo, Google Maps, Hunter, Firecrawl, Gmail). A clean example of pairing deterministic code with LLM judgment.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2067985770682503290
Reports that almost all the marketing orgs his team works with have around 10 data sources but have run on spreadsheets for five years, and his team becomes their first real data infrastructure. Within 24 hours they give a whole team a data analyst living inside their Claude Code or Codex, doing conversational analytics, live dashboards, and PPT reports via graphed.com/mcp. These people are measured on marketing data but historically had to wrestle the CTO for resources, and now they just do it themselves.
@shev_webmarke [Claude Code]
https://x.com/shev_webmarke/status/2067953023834538304
A hot non-coding use case: using Claude Code to crack open and optimize Google P-MAX campaigns, which are notoriously black-box. The workflow is to auto-pull P-MAX query data and check for cannibalization against other P-MAX and search campaigns, join query performance with per-asset data to find which queries and assets to push, then decompose channel reports over time to adjust assets and automated bids. He notes Claude Code shines exactly when the data volume is large and the system is too opaque to tackle by hand.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2067989720680820981
Did $2,000-$8,000 of client work in three days for four brands, writing zero lines of code himself, at $500-$2,000 per build: Claude Code shipped the websites, Claude Cowork shipped the funnels, and he ran the outreach. His point about the agency play is that the bottleneck was never the building, which Claude now does end to end, but the pipeline: who you email, what you promise, how you close. You're not the developer, you're the operator, and the clients don't care who's holding the keyboard.
@ventry089 [Claude Code]
https://x.com/ventry089/status/2067937460529586542
Profiles an Upwork UI/UX designer who went from $7,000/month (5 clients, design only, Jan 2025) to $18,000/month (12 clients, design plus full code plus deployment, Apr 2026), 157% growth, without learning to code because Claude Code writes it. One client's 4-hours-a-day manual data entry was cut to zero with a single automation priced at $1,500. The framing: the $25/hour freelancer and the $200/hour one use the same tool; one sells time, the other sells a working system.
@S1TA10 [Claude Code]
https://x.com/S1TA10/status/2067958005908246878
A 22-year-old non-technical founder closed a signed client contract for a four-agent pipeline built entirely with Claude Code. Agent 1 scrapes Google Maps and Instagram for leads overnight, agent 2 auto-creates a personalized plan and mockup per lead, agent 3 drafts a personalized outreach email straight into Gmail, and agent 4 coordinates the three, tracks every lead's status, and flags when a human is needed. She didn't write code; she made the proposal, negotiated terms, and signed the deal, and Claude Code did everything else.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2067766615962489258
Vibe-coded a Google Maps lead-gen scraper entirely in Claude Code that pulls hundreds of local business leads in minutes: enter a keyword, city and state and get back names, phone numbers, emails, websites and reviews, using Apify under the hood. It includes a bookmark feature to save the best leads, search history, and optional Replit hosting so a whole team can use it. He pitches it as replacing a $200/month scraper SaaS for agencies and lead-gen operators who need a fresh prospecting list weekly.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2068043305292886465
Built a TikTok Shop research tool in Claude Code running on two API keys, pitched as replacing a $99/month Kalodata subscription. Search any niche and it returns a product leaderboard ranked by real revenue, auto-loads the exact TikTok videos driving each product's sales, has Gemini watch every video to break down the hook, angle and why it sells, mines hundreds of customer reviews into ad angles, and generates a creative brief for your own brand built on the proven winners. A solid non-coding e-commerce workflow.
@genel_ai [Claude Code]
https://x.com/genel_ai/status/2067797410659856403
Reports a near-hands-off client-work pipeline for trivia/song videos built with Claude Code plus Seedance 2.0 (PolloAI). In about two weeks it racked up 1.3M+ total views and roughly +1,000 followers, with actual hands-on time of only about 10 minutes. He says the result running almost entirely on autopilot is a strong weapon, and he documented the full environment setup and automation method.
@PAGE4163929 [Claude Code]
https://x.com/PAGE4163929/status/2067980417823834197
Highlights a faceless YouTube channel, Zen, that reportedly gained 130,000 subscribers and 14M views from just 12 videos in about a month, with the analytics tool vidIQ estimating roughly $61,000/month. The creator claims to reproduce similar crude hand-drawn-style illustration videos in under 20 minutes using Claude Code and Higgsfield, via a script to audio to transcript to per-timestamp image generation to editing pipeline. A vivid (if eye-popping) data point on AI content side-hustles.
@0xAI42exe [Claude Code]
https://x.com/0xAI42exe/status/2067949769126015445
Runs Claude Code on an always-on VPS (called Beam) and steers it from his iPhone over a secure link, with the code staying local. The concrete moment: from his phone, Claude Code took a 64-minute guest call, transcribed it, cut 9 chapter timestamps, pulled action items, dropped it into his course as a draft lesson, and formatted a promo post, all without him opening a laptop. The course, community and ad account on this setup clear around $13,000 a month, one person running a business and a dev team from a screen in his pocket.
@Flandermaxx [Claude Code]
https://x.com/Flandermaxx/status/2067958462907027512
A vivid local-AI story: a 31-year-old Shenzhen repair technician buys dead RTX 3090s for $80 a card, reballs the GA102 die at his bench, and revives them. He keeps the four cleanest cards (96GB total VRAM) in an open frame running Qwen 3 235B on vLLM and Llama 3.3 70B unquantized. The old cloud stack cost $400 a month; the new one costs $0, with Claude Code pointed at localhost via a single env var and nothing leaving the bench.
@CrypSaf [OpenClaw]
https://x.com/CrypSaf/status/2068058354644090935
Describes a personal AI agent (Safio) running locally on a Mac mini through OpenClaw/Hermes, using Claude and GPT models plus Discord data, Google Sheets, bots, code and memory to actually do work. For his 0xAlphaGEMs community ops it runs a Mission Control dashboard, activity and member-visibility tracking, wallet submit/collection systems, reaction-role workflows, giveaway allocation sheets, Discord exports, AMA summaries, wallet dedupe and validation, analytics rankings, and routine ops cleanup. A strong non-coding, always-on agent example.
@nuuuukkkkk [OpenClaw]
https://x.com/nuuuukkkkk/status/2067974522809327802
Built a personal wine cellar agent for OpenClaw/Hermes. After a one-line setup, you just photograph each bottle's label and the agent reads it and logs it to your cellar; snap the receipt too and it records the date and price automatically. You tell it what you drank ('drank XXX today'), ask it for a pairing ('having fried chicken tonight, what should I open?'), and leave tasting notes ('today's wine had no body, 2/5'). A delightful non-coding domestic use case.
@grapeot [Claude Code]
https://x.com/grapeot/status/2068008910259384822
Explains Android's NotificationListenerService (which lets you read and decide on each notification individually) and highlights a developer who used Claude Code to build a spoiler-blocking app. It fine-tuned Qwen3-1.7B, quantized to a 1GB GGUF running locally on the phone, to judge each notification for sports spoilers and cancel them. On-device was non-negotiable because the app reads WeChat private messages and nothing can leave the phone. It turns notification management from per-source toggles into per-content semantic decisions.
@eightbeat8b [Claude Code]
https://x.com/eightbeat8b/status/2067815522528567752
Built a three-tier task management system on Claude Code plus Markdown plus shell hooks, inspired by Nous Research's Hermes Agent, to fix the problem that most task apps treat completion as deletion and lose the knowledge. Tasks are split per Epic, with shared domain knowledge in _README.md and per-project knowledge in _epic.md; a PostToolUse hook extracts only YAML frontmatter into a lightweight dashboard, and tasks inherit knowledge by reading up the hierarchy. On completion, decisions and learnings are auto-promoted into a persistent log, mirroring an Execute-Evaluate-Extract-Improve self-improving loop.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2067959336887026103
Cites a Korean PhD researcher who analyzed 7,944 public GitHub Claude skills and concluded that 33% perform worse than using no skill at all, because they bloat the system prompt and destabilize the model. The reported optimal stack is curated skills managed in Obsidian plus dynamic execution via a Hermes agent plus Opus 4.8. He notes a solo developer running a roughly $5,000/month automation agency on exactly this stack, with skill curation directly determining operation quality.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2067957299113537818
A vivid account of Unreal Engine 5.8 shipping with Claude Code wired into the editor via an MCP server, where the model reads the scene and acts in it. It started with a sofa, rug and table; a chair tagged 'reading nook' furnished the room around it; drawing a boundary and saying 'city' generated districts, highways and buildings via semantic asset search. Lighting changes came from prompts like 'purple dusk' and 'Bogota, 9:30 morning' (pulling the real sun position), with one misread of 'overcast' breaking the clouds, fixed by one more prompt. The artist stopped building the city and started describing it.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2068066710394204216
A five-step, non-coding setup to redesign a room with Claude Code, each step under two minutes: install the Nano Banana skill by pasting its URL into Claude Code, create a Google AI Studio API key and paste it in, then upload a room photo, type /banana, and ask it to redesign the room like a professional interior designer. It returns three to four redesigns that look like a real job, and you can ask it to swap furniture, change colors, or move the layout. Pitched as undercutting interior designers who charge thousands per room.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2068030633473962464
A no-coding five-step setup for product managers to use Claude Code. Install the Claude Code VS Code extension and sign in, pick Pro/Max or free via OpenRouter, describe your product so Claude saves it to CLAUDE.md, connect work apps (Notion, Linear/Jira, Slack, Gmail, Calendar, Drive) via connectors, and install the phuryn/pm-skills library of 100+ PM skills covering discovery, PRDs, OKRs, postmortems, JTBD and persona builders. He notes the bonus template is tuned for 50-75% fewer tokens with OpenRouter pre-wired.
@DanKornas [Claude Code]
https://x.com/DanKornas/status/2067819904917463394
Introduces Claude-BugHunter, an open-source (MIT) Claude Code skill bundle for authorized bug bounty, web app pentesting and external red-team work. It auto-loads 71 topic-specific skills, ships 681 report patterns curated from disclosed HackerOne reports across 24 vulnerability classes, includes 15 slash commands plus an engagement engine, supports multi-harness install to Claude Code/OpenCode/Codex CLI/Hermes Agent, and offers optional Burp MCP wiring. It takes you from recon to validation to reporting with the right playbook loaded.
@rnagulapalle [Claude Code]
https://x.com/rnagulapalle/status/2068033456203063777
Handed Claude Code a brownfield Next.js app with three unwired AgentMail integration gaps and a single prompt. Fifty seconds later: the agent ran curated workflows via the fetchsandbox MCP to learn the codebase shape, wrote the provision-inbox plus webhook handler plus teardown, told him what was still missing (env, ngrok tunnel, db stubs) before running, provisioned an inbox visible in the AgentMail console, and a test email from Gmail arrived in the agent's inbox. A clean demo of agentic graphs for API integrations.
@ManuelZapata [Claude Code]
https://x.com/ManuelZapata/status/2068070335086281139
Shares seven personal Claude Code skills he uses daily. /linear-plan turns a Linear issue ID into a reviewable implementation plan; /address-pr-comments evaluates and implements PR feedback then replies; /morning reviews Slack, open Linear issues and CC sessions to prioritize the day; /session-wrapup posts completed work to Linear; /merge-to-test encapsulates merge-and-deploy; /ingest extracts document info into his LLM wiki; and /commit-push-pr (copied from Boris Cherny). A practical, reusable everyday workflow library.
@AI4Azure [Claude Code]
https://x.com/AI4Azure/status/2067879617676099961
A small but practical access hack: following tutorials on Twitter, he opened a Mercury (US-based) bank account that provides a local US address and statements, then used it to pay for a Claude Code subscription smoothly. A reminder that for many overseas users, payment and access friction is a real part of the Claude Code experience.
@pangyusio [Claude Code]
https://x.com/pangyusio/status/2067856058748719400
A short but telling academic use case: he had Claude Code write his paper rebuttal, and after submitting it the reviewer raised his score by 1.5 points. A small data point that Claude Code's value reaches well beyond writing code into high-stakes academic writing.
🗣 User Voice
User Voice
Token cost has become the number-one anxiety. Enterprises are burning annual budgets in months and slapping on hard caps, and individual builders keep doing the math on local rigs to escape the bill. @jimmyhuli and @neil_xbt both frame loops as a spend-governance problem, not a magic trick.
Loop reliability is the real frontier, not raw model IQ. Users want maker-checker splits, durable state, worktrees, connectors, and verifiers they can trust before walking away. @neil_xbt and @aakashgupta both stress that a vague goal or a self-grading agent quietly drains money.
Safety and guardrails are now front of mind. Bypass mode pulling a trojan and a hallucinated delete-everything command have users asking for permission tiers and deny-rules by default. @Barret_China and @AAA_a080 are the cautionary tales; @humzaakhalid is the template.
Cheap-model routing is a top request. People want to run GLM-5.2, Ollama and local models through the Claude Code harness to slash cost without losing the workflow. @Flandermaxx and @ClaudeCode_UT show how far the local/curated stack can go.
Memory and knowledge persistence keep coming up. Agents forget between runs, and users are building their own systems so completed work compounds instead of getting archived and lost. @eightbeat8b turned task completion into permanent knowledge accumulation.
Token cost has become the number-one anxiety. Enterprises are burning annual budgets in months and slapping on hard caps, and individual builders keep doing the math on local rigs to escape the bill. @jimmyhuli and @neil_xbt both frame loops as a spend-governance problem, not a magic trick.
Loop reliability is the real frontier, not raw model IQ. Users want maker-checker splits, durable state, worktrees, connectors, and verifiers they can trust before walking away. @neil_xbt and @aakashgupta both stress that a vague goal or a self-grading agent quietly drains money.
Safety and guardrails are now front of mind. Bypass mode pulling a trojan and a hallucinated delete-everything command have users asking for permission tiers and deny-rules by default. @Barret_China and @AAA_a080 are the cautionary tales; @humzaakhalid is the template.
Cheap-model routing is a top request. People want to run GLM-5.2, Ollama and local models through the Claude Code harness to slash cost without losing the workflow. @Flandermaxx and @ClaudeCode_UT show how far the local/curated stack can go.
Memory and knowledge persistence keep coming up. Agents forget between runs, and users are building their own systems so completed work compounds instead of getting archived and lost. @eightbeat8b turned task completion into permanent knowledge accumulation.
📡 Eco Products Radar
Eco Products Radar
Codex - OpenAI's coding agent, the single most-mentioned rival/companion to Claude Code; used side by side, for cross-review, and as a fallback.
GLM-5.2 - Zhipu's open-weights model, the cheap-coding story of the day, run through the Claude Code harness via base-URL swaps.
OpenClaw / Hermes - the personal always-on agent stack; powering local Mac-mini agents, wine cellars, and community ops.
Cursor - still a default coding IDE/agent in the mix, often capped alongside Claude Code in enterprise budgets.
Ollama / vLLM - local-inference runtimes letting users point Claude Code at localhost to cut cloud bills to zero.
Obsidian - the de facto knowledge/skill management layer paired with Claude Code workflows.
Gemini - used as a vision/analysis sidecar inside Claude Code pipelines (e.g. watching TikTok videos).
Higgsfield - AI visual/animation generator paired with Claude Code for faceless video content.
Unreal Engine 5.8 - now ships Claude Code wired into the editor via MCP for prompt-driven scene building.
Codex - OpenAI's coding agent, the single most-mentioned rival/companion to Claude Code; used side by side, for cross-review, and as a fallback.
GLM-5.2 - Zhipu's open-weights model, the cheap-coding story of the day, run through the Claude Code harness via base-URL swaps.
OpenClaw / Hermes - the personal always-on agent stack; powering local Mac-mini agents, wine cellars, and community ops.
Cursor - still a default coding IDE/agent in the mix, often capped alongside Claude Code in enterprise budgets.
Ollama / vLLM - local-inference runtimes letting users point Claude Code at localhost to cut cloud bills to zero.
Obsidian - the de facto knowledge/skill management layer paired with Claude Code workflows.
Gemini - used as a vision/analysis sidecar inside Claude Code pipelines (e.g. watching TikTok videos).
Higgsfield - AI visual/animation generator paired with Claude Code for faceless video content.
Unreal Engine 5.8 - now ships Claude Code wired into the editor via MCP for prompt-driven scene building.
Comments