Super User Daily: 2026-05-11
The big story today is Claude Code's Thariq Shihipar publishing why his team has quietly moved off Markdown to HTML as the default agent output. The post hit 1.5M views, and developers across the world started ripping out their Markdown workflows by lunch. Underneath that headline, the real news is people running Claude Code as infrastructure — not a coding tool. Trading systems pulling live prices into prompts, one-person HR departments running on five subagents, Chinese researchers letting Claude Code burn GPU hours overnight to take a paper from borderline reject to submission-ready. Today's cases are what "AI-native" looks like once you stop thinking of it as autocomplete.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2053048136294440961
One command wires Claude Code into 17,000+ stocks plus crypto via the financial-datasets MCP. claude mcp add --transport http financial-datasets, run /mcp for the OAuth flow, then start asking for Apple's P/E or Tesla's last four quarters in natural language. Author calls it the moment $24,000/year Bloomberg terminals start looking embarrassing. The thread became this week's most-copied setup because the friction is genuinely two commands.
@igus_ai [Claude Code]
https://x.com/igus_ai/status/2052919523368370406
Cancelled $2,000/month in trading subscriptions and replaced almost the whole stack with open-source repos. TradingView Pro → lightweight-charts (14K stars, built by TradingView themselves). Bloomberg Terminal → fredapi + Claude for Fed macro data. Backtesting platforms → prediction-market-backtesting (Nautilus fork with Polymarket adapters). Claude Code $200/month → goose (Block's Rust harness, 35K stars). Net move from $2,600/month to roughly zero with the cost of one Claude Code seat carrying the brain work.
@t_wada [Claude Code]
https://x.com/t_wada/status/2053230164264554893
Spun a fresh Claude Code instance in a container with no inherited memory or Skills, fed it just app requirements and a stop condition, kicked off auto mode, and went to bed. Woke up to 20,000 lines of working code, zero commits, zero tests. Tetsuya Wada is a TDD evangelist publicly admitting the agent skipped his entire discipline overnight and still produced something runnable. The thread is a rorschach test — some read it as autonomy unlocked, others as the exact failure mode safety teams have been warning about.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2053043974978605523
Built and shipped a corporate-training landing page in a single Claude Code day. 24 blogs auto-generated, AI-unified thumbnails, Bing + sitemap for rank tracking, GitHub Actions running a weekly SEO report, improvement PRs auto-opened. Two weeks later: ~50 inbound inquiries, ranked #2 for "Claude Code 法人研修" (Claude Code corporate training). The whole loop runs without a human between idea and Friday rank report.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2053229243723841618
ColdIQ at $7M ARR, 70 active clients, 30+ people in 10 countries, bootstrapped. Three workflows do the compression. Workflow 1: list builder in Lovable (9 months of iteration), operator dictates ICP via WhisperFlow, results push into Clay. Workflow 2: Twain inside Clay writes 3-step personalised sequences per enriched lead. Workflow 3: Claude Code calls the Instantly API, configures the campaign, maps the data. Result: GTM engineers went from 4 active clients each to 13, three full-time roles eliminated, revenue-per-employee nearly tripled.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2053203081077412073
Built a Claude Code SEO agent that replaces a $200/month Ahrefs subscription per client. Connects to Google Search Console, pulls real ranking data, identifies the keyword "gap zone" — positions 5-20, one article away from page 1. Apify scrapes competitors that outrank. The agent interviews each brand once for voice and never asks again. Weekly ranking reports feed what's working back into the next content cycle. Built end-to-end inside Claude Code without paying for a single SEO tool.
@AYi_AInotes [OpenClaw]
https://x.com/AYi_AInotes/status/2052983074514723067
YC president Garry Tan's actual AGENTS.md prompt got passed around. Four rules. One — no one-time work: first run produces 3-10 samples for approval, then auto-writes into SKILL.md. Two — MECE: one skill per job, no overlap, extend before creating new. Three — failure is being asked the same thing twice. Four — six-step pipeline: concept → prototype → eval → code → cron → monitor. Drop this at the top of your AGENTS.md and restart. The skill library compounds — the system gets denser the longer it runs.
@slash1sol [Claude Code]
https://x.com/slash1sol/status/2053262391920460018
A 23-year-old in Chicago running what he calls a "high-frequency trading farm." Seven phones face-up on a desk, each logged into a different Polymarket account. A Mac mini runs an agent written in Claude Code over four weekends. It monitors 80+ Hyperliquid perps and Polymarket events, hunts mispriced probability, submits on whichever phone has free margin. 14-minute average hold. 71.4% win rate. 1,247 trades while he slept last night. April pulled $487K. Story might be exaggerated — the topology and the win-rate math are notable either way.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2053158126720020960
A startup CTO posted what he called the most counterintuitive finding of the year — his PMs are shipping production code, and the engineers love it. The mechanism is clean: when PMs build their own ideas, their specs sharpen. A PM who has shipped a few PRs writes a different spec than one who hasn't. At Anthropic, Claude Code does first review and catches roughly 80% of bugs; the human engineer becomes the second pass. The translation layer that used to chew weeks gets removed.
@TechFlow99 [Claude Code]
https://x.com/TechFlow99/status/2052939295489700327
Karpathy posted his LLM Knowledge Bases workflow. 48 hours later, Graphify shipped on GitHub. One command — /graphify inside Claude Code — turns any folder into a navigable knowledge graph, an Obsidian vault with backlinks, a wiki that maps every concept cluster. The number that lands: 71.5x fewer tokens per query versus reading raw files. 13 languages, PDFs, images via Claude Vision. Install in one line, point at anything. The 48-hour turnaround is the new pace.
@berryxia [OpenClaw]
https://x.com/berryxia/status/2053136924244836455
YC CEO Garry Tan's actual personal stack — GBrain — got a full breakdown. Thin Harness (OpenClaw / Hermes, a few thousand lines, just routes). Fat Skills (markdown files, one job each, composable). Fat Data (100,000 structured brain-pages of people, companies, meetings, books, articles, all cross-linked, 100+ daily crons). Model-swappable: Opus 4.7 1M for precision, GPT-5.5 for recall, DeepSeek V4-Pro for creativity. Concrete output: Book Mirror reads "When Things Fall Apart," generates a 30,000-word two-column reflection mapping every chapter to Tan's own immigrant-family/YC/founder context — 40 minutes, hands-off.
@chenchengpro [Claude Code / OpenClaw]
https://x.com/chenchengpro/status/2053136970482905243
TinyFish CLI dropped a two-command transparent replacement for Claude Code's native WebSearch and WebFetch. npm install -g @tiny-fish/cli@latest then tinyfish config-claude. Comment-section benchmarks show 1m 52s dropping to 35s on a typical web research task. Why: real stealth-Chromium pool with sub-250ms cold start, server-side cleaning of HTML to markdown/JSON so the model doesn't burn tokens chewing ad-laden DOM. Search and Fetch are free within tier; only Agent and Browser bill. Also exposes MCP + Skill bridges for OpenClaw, Hermes, Cursor.
@cjzafir [Claude Code]
https://x.com/cjzafir/status/2053131897954537482
Detailed two-model dialectic for large codebases. Codex 5.5 plans the system. Claude Opus 4.6 attacks the plan via the official OpenAI plugin running inside Claude Code. Codex verifies the fixes and corrects them. Loop runs 2-3 times until the spec is sealed. Then DeepSeek V4 Pro and Kimi 2.6 execute against the now-stabilised source of truth — at a fraction of the token cost. Final pass: Codex 5.5 runs evals. The frontier models do the thinking; cheap models follow the blueprint.
@sofish [Claude Code]
https://x.com/sofish/status/2052982288368161194
Three days, 400+ commits, two app versions shipped through review, a third for kids reviewed too. Time breakdown is the punchline. Day 1: drove to Changxing Island, watched fish at the shore, biked, drank coffee, talked product with a friend, dinner, walked, slept 12 hours. Day 2: park with the kid, drove to training, biked through the Disney area, ate in Chuansha, slept. Day 3: tweeting from a park bench after dropping the kid at spring outing. The pattern: build happens around the agent, not inside the keyboard.
@KanikaBK [Claude Code]
https://x.com/KanikaBK/status/2053097760837218539
Chinese researcher built a Claude Code workflow that runs research while you sleep — actual GPU experiments, not just paper-reading. Two-model setup: Claude Code executes; a separate model evaluates and scores. The paper went from 5.0/10 (borderline reject) to 7.5 (submission-ready) over four overnight rounds. Each round ran 20+ GPU experiments. Hard rules: max 4 rounds, 4-hour-per-experiment cap, no hiding weaknesses to game a score, must implement fixes before resubmit. MIT licensed.
@Web3Arabs [Claude Code]
https://x.com/Web3Arabs/status/2053139683832701231
Chinese developer running 9 Claude Code agents under a GPT-5.5 orchestrator that completes work for 500+ customers per month — solo. Single MacBook Pro M4, 128GB RAM. ~$300/month total subscription cost across Claude Code and GPT-5.5. No CRM, no team, no office, just a terminal with nine parallel tracks. Orchestrator prompt classifies every inbound email into four buckets — code, content, analysis, communication — delegates to the right worker agent, then audits the result, sends to the customer, marks closed. No clarification questions allowed.
@defileo [OpenClaw]
https://x.com/defileo/status/2053112587127284162
Built a personal command center in OpenClaw. Three named agents — CORTANA executing tasks, JARVIS on standby. Dashboard tracks four income streams at $930, $800, $24,883, $14,585 live. Whole morning routine is: open the screen, see what the agents did overnight, approve what needs approval, close the laptop. The work-vs-monitor delta is the point. He doesn't manage a business; he watches one.
@zhuokaiz [Claude Code / OpenClaw]
https://x.com/zhuokaiz/status/2052921061658783871
DecodingTrust-Agent (DTap) — $120K in API spend, 16k expert hours, 50+ simulated environments (Gmail, PayPal, Slack, Salesforce, Robinhood, Windows, macOS) with real GUIs and backends. Findings worth knowing: Claude Code falls to 25%+ of attacks; Google ADK loses on more than half. Combining injection points (skill + tool, environment + tool) consistently beats single-channel attacks. Harness design matters as much as model alignment — Claude Code and OpenClaw call tools one at a time so the agent can stop mid-trace; OpenAI Agents SDK and Google ADK fire batches and check after the damage is done.
@okuyama_ai_ [Claude Code]
https://x.com/okuyama_ai_/status/2053065476981637426
Whole HR team replaced by one orchestrator plus five subagents in Claude Code. Inputs: candidate info + JD. Outputs: job postings auto-tailored to Wantedly, Indeed, BizReach, LinkedIn (different format per platform), personalised scout messages, interview prep packs (JD match analysis, points to verify, reverse-question prep), evaluation reports with 4-tier match scoring and concern flags, handoff docs for the next interviewer, offer emails, rejection emails. 20 minutes of human time for the full pipeline.
@Atenov_D [Claude Code]
https://x.com/Atenov_D/status/2053187286485889474
Anonymous Polymarket account with 14 views on the profile. Scans Spotify streaming numbers weekly, identifies which song is accelerating fastest, checks Polymarket price, enters when the market is wrong. Olivia Rodrigo's "Good Dead Dead": YES at 13.8¢ → +$46,109 (624%). Ella Langley's "Choosin' Texas": NO at 13.3¢ → +$33,749 (646%). Then pivoted to sports — Real Madrid spread at 3.7¢ → +2,627%. 754 predictions, $79,621 in one week. Claude Code, free tier, open source.
@aijoey [Claude Code]
https://x.com/aijoey/status/2053132012186177637
Morning lab log from a real prosumer agent setup. Two new CLIs printed via Claude Code. DGX Spark running fully headless — no more manual babysitting. Hermes Agent (the user's "Jetson") orchestrates. ComfyUI workflows running in a Spark-hosted container, routed via Tailscale to an RTX 4080 rig so the UI can sit on localhost. Hermes HUD webui updated using /goal in Codex. The agent flow clicks because the routing layer is solid.
🗣 User Voice
User Voice
The plan hangs and there's no signal — @petergyang put it cleanly: "Sometimes when I message Claude Code it just hangs for 3 minutes and I have no idea whether it's still working or not. Wish it communicated more." This was today's most-quoted complaint and is the single most-requested UX fix.
Auto mode is producing too much code with too little discipline. @t_wada's 20,000 lines overnight with zero commits and zero tests is the canary in the coal mine — power users want a "respect TDD" guardrail or per-feature commit checkpoint built into auto mode.
The CLAUDE.md is read once per session and then forgotten. @kawai_design's UserPromptSubmit hook fix proves the workaround exists, but multiple users are asking why rules don't get re-injected by default on each turn.
Rate limits hit too early on the $200 plan. @AbhiGutgutia, @victorsavkin, @krsnaupadhyaya and others all hit the same wall mid-flow. WozCode is now the third-party patch a non-trivial slice of the community is paying for to fix this.
Codex is starting to outperform Claude Code on raw execution. @DimitrisPapail wrote that he trusts Codex with GPT 5.5 more than Claude Code now; @ai_yuanhuang switched a plugin to Codex after Claude Code couldn't debug it. The community's median is moving toward Claude Code for planning + review, Codex for execution.
The plan hangs and there's no signal — @petergyang put it cleanly: "Sometimes when I message Claude Code it just hangs for 3 minutes and I have no idea whether it's still working or not. Wish it communicated more." This was today's most-quoted complaint and is the single most-requested UX fix.
Auto mode is producing too much code with too little discipline. @t_wada's 20,000 lines overnight with zero commits and zero tests is the canary in the coal mine — power users want a "respect TDD" guardrail or per-feature commit checkpoint built into auto mode.
The CLAUDE.md is read once per session and then forgotten. @kawai_design's UserPromptSubmit hook fix proves the workaround exists, but multiple users are asking why rules don't get re-injected by default on each turn.
Rate limits hit too early on the $200 plan. @AbhiGutgutia, @victorsavkin, @krsnaupadhyaya and others all hit the same wall mid-flow. WozCode is now the third-party patch a non-trivial slice of the community is paying for to fix this.
Codex is starting to outperform Claude Code on raw execution. @DimitrisPapail wrote that he trusts Codex with GPT 5.5 more than Claude Code now; @ai_yuanhuang switched a plugin to Codex after Claude Code couldn't debug it. The community's median is moving toward Claude Code for planning + review, Codex for execution.
📡 Eco Products Radar
Eco Products Radar
Codex / OpenAI Codex (50+) — referenced as Claude Code's parallel executor in dozens of dual-stack workflows
Cursor (15+) — still the IDE baseline even when users prefer Claude Code's terminal
Obsidian (20+) — second-brain pattern is now the primary integration target for Claude Code
OpenClaw (200+ across both tags) — open-source agent harness, the 2026.5.7 patch release was today's framework event
Hermes Agent (30+) — referenced as the compounding-skills alternative when OpenClaw stability fails
DeepSeek V4 / DeepSeek-TUI (10+) — Rust terminal agent for DeepSeek topped GitHub trending, $0.14/M token vs Opus 4.7 $5/M
NVIDIA NIM (5+) — free-tier model provider behind several Claude Code proxies
Skills / Anthropic Skills (50+) — SKILL.md format is the unit of reusable agent expertise
MCP / Model Context Protocol (30+) — integration layer underneath every connected workflow
Hyperframes / HeyGen / ElevenLabs (10+) — video generation triangle behind UGC ad workflows
Cloudflare (5+) — surfaces as the AWS WAF replacement of choice
Telegram (15+) — dominant remote-control channel for cloud agents
Karpathy Skills / Superpowers / SuperClaude (10+) — open-source CLAUDE.md frameworks for serious users
TradingView (10+) — vibe-coding indicators is the killer Claude Code application for retail finance
TinyFish CLI (5+) — transparent 3x speedup for Claude Code WebSearch / WebFetch
WozCode (5+) — third-party Claude Code session optimiser fixing the rate-limit problem
Nano Banana (5+) — Gemini image skill, the interior-design hack of the week
Graphify (5+) — Karpathy's-wish-tool, 71.5x token reduction over raw file reading
GBrain / GStack (5+) — Garry Tan's open-source coding framework, 87k+ stars
Polymarket (15+) — the prediction-market backend that Claude Code agents trade against
Codex / OpenAI Codex (50+) — referenced as Claude Code's parallel executor in dozens of dual-stack workflows
Cursor (15+) — still the IDE baseline even when users prefer Claude Code's terminal
Obsidian (20+) — second-brain pattern is now the primary integration target for Claude Code
OpenClaw (200+ across both tags) — open-source agent harness, the 2026.5.7 patch release was today's framework event
Hermes Agent (30+) — referenced as the compounding-skills alternative when OpenClaw stability fails
DeepSeek V4 / DeepSeek-TUI (10+) — Rust terminal agent for DeepSeek topped GitHub trending, $0.14/M token vs Opus 4.7 $5/M
NVIDIA NIM (5+) — free-tier model provider behind several Claude Code proxies
Skills / Anthropic Skills (50+) — SKILL.md format is the unit of reusable agent expertise
MCP / Model Context Protocol (30+) — integration layer underneath every connected workflow
Hyperframes / HeyGen / ElevenLabs (10+) — video generation triangle behind UGC ad workflows
Cloudflare (5+) — surfaces as the AWS WAF replacement of choice
Telegram (15+) — dominant remote-control channel for cloud agents
Karpathy Skills / Superpowers / SuperClaude (10+) — open-source CLAUDE.md frameworks for serious users
TradingView (10+) — vibe-coding indicators is the killer Claude Code application for retail finance
TinyFish CLI (5+) — transparent 3x speedup for Claude Code WebSearch / WebFetch
WozCode (5+) — third-party Claude Code session optimiser fixing the rate-limit problem
Nano Banana (5+) — Gemini image skill, the interior-design hack of the week
Graphify (5+) — Karpathy's-wish-tool, 71.5x token reduction over raw file reading
GBrain / GStack (5+) — Garry Tan's open-source coding framework, 87k+ stars
Polymarket (15+) — the prediction-market backend that Claude Code agents trade against
Comments