Super User Daily: 2026-05-02
Yesterday's Claude Code and OpenClaw timeline was less about feature releases and more about people pointing their agents at messy, real-world problems and watching the dollars (or the housing court fines) come out the other end. Tenants weaponizing the NYC housing code. A laid-off engineer who let an agent eat 740 job postings before applying to one. Polymarket bots built in two days on $68 of seed capital. The vibe-coding-for-money post is gone, replaced by tenant lawyer agents, on-chain sleuths, 21-character agency teams, and a 16-year-old in Czechia paying for his Porsche off Discord bots. Underneath all of it: an ongoing fight about whether Anthropic is silently surcharging anyone whose git history mentions "openclaw."
@Argona0x [Claude Code]
https://x.com/Argona0x/status/2049921143067288016
Bushwick tenant gets a $400 rent hike, opens Claude Code, says "you're my tenant lawyer now." Forty-eight hours later there's a $4 VPS scraping the HPD violation database for his building's BIN, cross-referencing the DOB complaint history back to 2019, pulling all 847 sections of the NYC housing maintenance code, drafting 14 separate 311 complaints with the exact subchapter cites, and filing them on a staggered schedule so it doesn't trip the spam heuristic. The agent figured out on its own that class C hazards (lead, no heat, no hot water) get a 24-hour HPD dispatch. Day 19: $2,800 in fines, building hit with a class C designation, landlord legally barred from raising rent. The text from the landlord asking if he wants to buy the building came in around the same time the agent flagged two more violations on the floor above.
@heygurisingh [Claude Code]
https://x.com/heygurisingh/status/2049912647722860840
Guy gets laid off, builds career-ops on Claude Code, evaluates 740 job offers with it, lands a Head of Applied AI role, then open-sources the whole thing. One slash command per JD URL gives back an A-F evaluation, an ATS-optimized PDF tailored to that exact role, salary research, interview prep, and a tracker entry. The system explicitly refuses to recommend applying to anything scoring below 4.0/5. It also runs an interview Story Bank that accumulates STAR+Reflection stories until you have a master answer for every behavioral question. 14 skill modes, 45+ companies pre-loaded, 19 ATS search queries, batch mode that runs 10+ evaluations through Claude sub-agents in parallel. 8.2k stars in days. The actual pitch is the philosophy — it's a filter, not a spray-and-pray.
@0xkvro [Claude Code]
https://x.com/0xkvro/status/2049949658495987983
Stranger at Changi airport gate C14 explains a Polymarket strategy on the back of a receipt before two officers escort him onto a deportation flight. The thesis: 25% of Polymarket volume is fake but fake volume has graph structure. Wallets with 10+ shared trades become edges, density-filtered components surface 7 clusters of coordinated wallets, and you fade the retracement 6-10 minutes after their pump. The author flies home, opens Claude Code, spawns 4 parallel agents (edge scanner, fingerprint detector, sync tracker, fade trigger). $1,000 seed, +$21,300 in 34 days, Sharpe 2.9, 74% WR. Total infrastructure: Claude $20 + Hetzner $5 + networkx free. Whether you trust the framing or not, the workflow itself — graph theory plus parallel agents on a $25/mo stack — is the part that's actually new.
@0x_Discover [Claude Code]
https://x.com/0x_Discover/status/2049765508300353943
19-year-old Japanese student spends two days letting Claude Code write the strategy for a Polymarket arbitrage bot while OpenClaw handles execution and Binance sync for short-term BTC. iPad as second screen. One overnight run: 3 trades opened, 2 closed, BTC short on a 15m timeframe goes 0.31 → 0.79 for +$6,732. Reference wallet shows 54k trades and ~$1.7M cumulative profit on pure arbitrage. Started with $68 of seed, copied trades, then graduated into custom monitoring and execution. Risk management is interesting too — auto shutdown on liquidity anomalies, manual confirmation for emergency exits, worst observed drawdown ~3%.
@AzFlin [Claude Code]
https://x.com/AzFlin/status/2049837215182885127
Claude Code with two tools (a paid RPC and the free Etherscan API) ends up being a competent on-chain sleuth. The use case: tracing the initial liquidity add for a custom Uniswap v4 launch where there was no launchpad to fall back on. Manual etherscan would have taken hours. The clever piece is that Claude can recursively walk every connected smart contract, pull all source code (including the v4 pool manager and hook contracts) through the Etherscan API, drop it neatly into a folder, and then answer architectural questions about the system. Useful enough that the author keeps a dedicated repo with API keys in .env and reopens Claude Code there whenever a new launch needs investigating.
@businessbarista [Claude Code]
https://x.com/businessbarista/status/2049919504277283071
The hiring problem is "I don't want a lame post for the Strategic Finance role." The solution: vibe-code a 20-level game where applicants find bugs in broken financial statements for a fake SaaS company. Built and refined in roughly an hour using Claude Code for the build, Claude Design for the redesign, and tight feedback loops via "dangerously skip permissions" + Opus 4.7. 2,000 plays in 48 hours, 615 caught at least one bug, 81 found all 20. He shipped a self-destruct mode (a fixed timer that physically cracks and drops the spreadsheet off-screen if you fail), a 21st boss bug behind the gate, and an AI detector that flags people pasting the screen into Claude. About half the winners are finance pros and half are AI users — both qualify, but in different ways.
@ashen_one [Claude Code]
https://x.com/ashen_one/status/2049894511145279801
One-hour walkthrough of building an iOS app that doesn't look like AI slop. The stack actually uses every recent toy: GPT Image 2.0 for mascots, Seedance 2.0 to animate them, Claude Design for the onboarding screens, Claude Code for the GitHub setup and multi-terminal simulator workflow, Supabase + Railway for backend, RevenueCat for paywall. The reason this matters more than another tutorial is the explicit attention to friction points — what to do when GPT Image 2 starts hallucinating biology diagrams instead of mascots, how to manage simultaneous Codex/Claude Code sessions in different terminals, what the right cadence of error fixing looks like when the agent is implementing your plan in real time.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2049868452987588619
A 45-second cultivation animation episode in two days, ¥75 (~$10), one person plus four AIs. DeepSeek V4 Pro (called from Claude Code) wrote the script and a fictional character bible. Codex with GPT Image 2 generated three-view character sheets to lock visual consistency. Then Seedance 2.0 handled the camera-heavy shots while Kling 3.0 took over for dialogue and faces (using first/last-frame locking). Doubao TTS did voiceover. Total time was actually about 8 hours, most of which was prompt iteration and recovering from "agent hallucinated a botany diagram, restart the session." The takeaway from the author: solo studios are real, but the new core skill is prompt engineering.
@browomo [Claude Code]
https://x.com/browomo/status/2049857376853848472
Chinese solo dev runs a website-audit-and-redesign agency for SMBs at $30K/month with no team, using a 7-agent stack on Claude Code Router. Daily token usage is around 3M, average API bill ~$450/month. The orchestrator's system prompt names six sub-agents (Scout for SMB leads, Researcher for site audits, Writer for cold outreach and proposals, Planner for milestone breakdowns, Builder for code, Reviewer for evals on every artifact) and forbids two of them from touching the same file. Human approval is only required when an invoice exceeds $5,000 or eval score drops below 0.85. The shared state lives in the file system, not in memory. Daily log entries read like a real ops report — "187 SMBs found, 14 with 2010-2014 websites... passing top 5 to researcher."
@Sprytixl [Claude Code]
https://x.com/Sprytixl/status/2049817031428739527
16-year-old in Czechia DMs a friend "bro let's try making Discord servers maybe they'll pay us," opens Claude Code, describes a moderation/economy/welcome bot, and just watches Claude write Python. $5 to list on Top.gg, first server connects that evening and pays $20 through PayPal, then four more in one night. Math at scale: 500 servers × $9/mo = $4,500 in passive income while he sleeps. The framing he chose — Midjourney makes $2.5M/mo and still has no website, just a Discord bot — is the whole point. Distribution channel selection matters more than the build.
@xiangxiang103 [Claude Code]
https://x.com/xiangxiang103/status/2049678737684525352
claude-tap turns Claude Code from a black box into something you can actually inspect — System Prompt, Tools list, Context, all visible after a one-time install. The use cases the author is running with it: trimming prompt to save tokens, optimizing tool-call logic, validating any custom model's real behavior under the harness, and digging into agent internals. Setup is genuinely zero-friction because Claude itself walks you through plugin install. If you've been using Claude Code for months without ever seeing what it actually sends to the model, this is the cheapest way to fix that.
@cocoindex_io [Claude Code]
https://x.com/cocoindex_io/status/2049916946284470412
Spektr is a self-hosted RAG-as-MCP backend that plugs straight into any MCP-compatible agent (Claude Code, Claude Desktop, etc.) as a first-class knowledge tool. The dual-path architecture matters more than the buzzwords: bulk ingestion from S3 or local FS uses CocoIndex with Docling OCR for PDF layout and Graphiti or GLiNER2 for the knowledge graph; live ingestion is a FastAPI endpoint that vector-indexes incoming text in ~200ms while building the graph as a background task. Dual store — Qdrant for semantic, Neo4j for entity-relationship. Four MCP tools (vector_search, visual_search ColBERT, graph_search, hybrid_search) that all support session_id, so agents can blend real-time session data with the bulk knowledge base.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2049814053804286021
"Uravation Leads" is a cold email automation tool the author built end-to-end in Claude Code: input industry × city, pull Google Maps for matching companies, scrape each site for contact info, run Google reviews through Gemini 3 Pro to extract three concrete customer complaints, then auto-generate an email tailored to those complaints. Roughly 60 seconds per lead, ~$0.05 per company. Notable that the build is multi-stack — Next.js + shadcn/ui frontend, FastAPI + SSE backend, Playwright for crawling, SQLite for persistence — and the rebuild from Streamlit to a polished dashboard with a radar loading animation took less than a day.
@lewangx [Claude Code]
https://x.com/lewangx/status/2049805142938296485
Wrote a WeChat chat-history viewer with Claude Code that can read messages, images, and videos. The reason this is in the top engagement bucket: nearly 100GB of WeChat history was eating his phone and laptop. With the viewer working, he could finally delete the original WeChat data and reclaim the storage. Boring use case, real outcome — and a reminder that "Claude Code wrote a single-purpose desktop tool to unlock a personal storage problem" is now a few hours of work for someone who doesn't otherwise ship Mac apps.
@koizuka [Claude Code]
https://x.com/koizuka/status/2049740749345001614
Server hardening with Claude Code as the loop's brain. The author hands it the vulnerability advisory and current server state, has it draft countermeasures and explanations in plan mode, picks the option, gets a step-by-step procedure manual, executes it himself, and finally has Claude verify whether the configuration changes actually closed the hole (without ever attempting an exploit). The interesting bit isn't that AI did the security work — it's that the human stayed in the loop on every decision and the agent's output was structured enough to audit afterwards.
@okuyama_ai_ [Claude Code]
https://x.com/okuyama_ai_/status/2049878059973812475
SEO content factory running on Claude Code: keyword selection via Ahrefs MCP, primary-source research from X and YouTube to satisfy E-E-A-T, a quality-scoring sub-agent that loops until each article hits 95/100, then auto-publishes to WordPress and submits to Google Search Console for indexing. Scheduled tasks in the morning, afternoon, and night handle the throughput. Stated output: 1 article in 40 minutes, 60 articles/month, near-zero labor cost. The non-obvious piece is the weekly rewrite + internal-link-optimization agent that keeps already-published articles from going stale.
@bridgemindai [OpenClaw]
https://x.com/bridgemindai/status/2049865751939473809
BridgeWard is a defensive skill that runs inside Hermes Agent and OpenClaw to detect prompt-injection patterns in untrusted content before the agent acts on them. Open source, MIT. The premise is correct — every agent that reads web pages, emails, or PDFs is structurally vulnerable to whatever was hidden in those documents — and the per-input audit step is the right primitive. This is also the kind of skill that's most valuable as a default for the broader community rather than something each team rebuilds.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2049784791801708908
A nightly review cron has been running on the author's OpenClaw for months — agent summarizes the day's interactions and extracts repeated patterns while you sleep. With OpenClaw 2026.4.29's "dreaming" feature plus gbrain 0.23, conversations now write directly into long-term memory; patterns auto-surface when something repeats 3+ times; synthesize and patterns commands both support dry-run so you can see what's about to be promoted before flipping it on. The note on the difference between OpenClaw dreaming (how the agent works with you tomorrow) vs gbrain (how we still understand what you built six months from now) is sharper than most posts about agent memory.
@damobianyuan [Claude Code / OpenClaw]
https://x.com/damobianyuan/status/2049680642372460789
agency-agents-zh ships 211 plug-and-play agent personas across 18 domains (engineering, design, marketing, product, finance, QA), all installable into OpenClaw, Claude Code, Cursor, Codex, or Gemini CLI. The point is not the personas themselves — it's that the cost of standing up a multi-role workflow has dropped to "describe what you want, let the orchestrator pick three roles, start." Pairs with agency-orchestrator for actual multi-agent coordination. For Chinese-language SMB workflows specifically, this is the most aggressive starter kit on GitHub right now.
@bc1beat [OpenClaw]
https://x.com/bc1beat/status/2049857846817468628
Founder of BlockRunAI runs all his marketing research and outreach on OpenClaw, $0.2-0.5/day in token spend. The setup uses GLM 5.1 routed through their own gateway — they charge 1000 calls per dollar — and the agent does daily competitor research plus initial outreach drafts. Worth flagging because the unit economics are openly stated: at sub-50 cents a day, a solo marketing function is now cheaper than a coffee, which is the actual reason agentic workflows are starting to compound for small teams.
@theo [Claude Code]
https://x.com/theo/status/2049645973350363168
The post that lit the fuse for the day's discourse: an empty git repo with a single commit containing {"schema": "openclaw.inbound_meta.v1"} causes vanilla Claude Code (called via the official CLI on a Max plan) to either refuse the request or surcharge the user as if they were on a third-party harness. Theo is famous enough that this got tested by other people inside the day, including @PawelHuryn (who confirmed the repro and read the surcharge as Anthropic detecting a harness manifest declaration) and @AYi_AInotes (who summarized in Chinese for ~280K impressions). It's both a working-as-designed enforcement story and a trust-erosion event for the developer base.
🗣 User Voice
User Voice
The aggregate signal from yesterday is sharper than the individual posts. A few patterns kept showing up:
- The token economics are what people care about now, not the model. @bridgemindai (https://x.com/bridgemindai/status/2049849449464156242) reports Codex on $200 ChatGPT Pro using 3% of the session for the same workload that ate 21% of Claude Code Max in three hours. @0x_Discover and @0xkvro both showed agent stacks running for $25-100/month total. The conversation has shifted from "is the model good" to "what's the cost-per-useful-action."
- Anthropic's enforcement layer is becoming a meta-feature. @theo (https://x.com/theo/status/2049645973350363168), @Yuchenj_UW (https://x.com/Yuchenj_UW/status/2049718399878812082), @rajpundkar (https://x.com/rajpundkar/status/2049854451515826582), and @0x_kaize (https://x.com/0x_kaize/status/2049772796666925405) all flagged the same theme — silent surcharges, account bans, harness detection in commit history. Whatever the policy intent, the perception is "fence over moat."
- People who actually run multi-agent stacks in production are converging on the same architecture. @browomo (https://x.com/browomo/status/2049857376853848472), @aakashgupta (https://x.com/aakashgupta/status/2049985094233002181), @damobianyuan (https://x.com/damobianyuan/status/2049680642372460789), @MindTheGapMTG (https://x.com/MindTheGapMTG/status/2049797364890632472): a single orchestrator, 5-21 sub-agents with named roles, file system as shared state, evals on every artifact, human approval only for high-impact decisions.
- Tacit knowledge is the next bottleneck and people know it. @bradmillscan (https://x.com/bradmillscan/status/2049965618053661164) wrote 60+ skills/playbooks/SOPs trying to make his agent not be a dufus, then admitted 90% of his time was guard-railing. The "headless skills + per-agent CLAUDE.md + hook-based safety" pattern is becoming the standard answer.
- Hardware-native local stacks are real. @aijoey (https://x.com/aijoey/status/2049855286718906383) bought a DGX Spark to run OpenClaw and Hermes locally because "owning your own box hits different." @cyrilXBT (https://x.com/cyrilXBT/status/2049706913030681056) walked through Claude Code → Ollama → Gemma 4 with no API costs. The signal: people who ran into rate limits in March are buying their own GPUs in April.
- The token economics are what people care about now, not the model. @bridgemindai (https://x.com/bridgemindai/status/2049849449464156242) reports Codex on $200 ChatGPT Pro using 3% of the session for the same workload that ate 21% of Claude Code Max in three hours. @0x_Discover and @0xkvro both showed agent stacks running for $25-100/month total. The conversation has shifted from "is the model good" to "what's the cost-per-useful-action."
- Anthropic's enforcement layer is becoming a meta-feature. @theo (https://x.com/theo/status/2049645973350363168), @Yuchenj_UW (https://x.com/Yuchenj_UW/status/2049718399878812082), @rajpundkar (https://x.com/rajpundkar/status/2049854451515826582), and @0x_kaize (https://x.com/0x_kaize/status/2049772796666925405) all flagged the same theme — silent surcharges, account bans, harness detection in commit history. Whatever the policy intent, the perception is "fence over moat."
- People who actually run multi-agent stacks in production are converging on the same architecture. @browomo (https://x.com/browomo/status/2049857376853848472), @aakashgupta (https://x.com/aakashgupta/status/2049985094233002181), @damobianyuan (https://x.com/damobianyuan/status/2049680642372460789), @MindTheGapMTG (https://x.com/MindTheGapMTG/status/2049797364890632472): a single orchestrator, 5-21 sub-agents with named roles, file system as shared state, evals on every artifact, human approval only for high-impact decisions.
- Tacit knowledge is the next bottleneck and people know it. @bradmillscan (https://x.com/bradmillscan/status/2049965618053661164) wrote 60+ skills/playbooks/SOPs trying to make his agent not be a dufus, then admitted 90% of his time was guard-railing. The "headless skills + per-agent CLAUDE.md + hook-based safety" pattern is becoming the standard answer.
- Hardware-native local stacks are real. @aijoey (https://x.com/aijoey/status/2049855286718906383) bought a DGX Spark to run OpenClaw and Hermes locally because "owning your own box hits different." @cyrilXBT (https://x.com/cyrilXBT/status/2049706913030681056) walked through Claude Code → Ollama → Gemma 4 with no API costs. The signal: people who ran into rate limits in March are buying their own GPUs in April.
📡 Eco Products Radar
Eco Products Radar
Products mentioned 3+ times across yesterday's data, with rough why-they-matter notes:
Codex / GPT-5.5 — dominant alternative narrative, with most posts framed as "Codex over Claude Code now" or hybrid use (plan in Claude Code, execute in Codex).
DeepSeek V4 Pro / Flash — the cost narrative. Multiple authors running DeepSeek V4 through Claude Code or OpenCode for pennies.
Cursor / Cursor SDK — the harness story. Cursor SDK shipping yesterday made it possible to embed Cursor's agent loop in CI/CD and other products.
Hermes Agent — consistent #2 alongside OpenClaw for self-hosted, multi-channel deployments. Multiple posts describe migrating from OpenClaw to Hermes.
Skills / ClawHub / npx skills — the package manager layer. Every workflow post now ships a skills install command.
Obsidian + claude-obsidian + Tolaria — knowledge-base agents are converging on Obsidian as the substrate.
NotebookLM — the deep-research offload destination. Multiple workflows route research through NotebookLM to save Claude Code tokens.
GPT Image 2.0 / Seedance 2.0 / Higgsfield — the creative-stack triple, used together in iOS app builds, anime production, and ad creative pipelines.
Stripe Link agent wallet / Machine Payments Protocol — yesterday's Stripe Sessions launch landed in nearly every "agentic future" post.
Manus Cloud Computer — non-engineer-friendly always-on agent infrastructure, positioned against OpenClaw + Mac Mini.
Ollama + Gemma 4 + GLM-4.7-Flash — the local-stack triple for people exiting API subscriptions.
n8n + n8n-MCP — the workflow automation that Claude Code can now create and edit directly via MCP.
Trendtrack + Meta Ads CLI — the marketing-agent triple, used together to spy competitors, generate creatives, and manage ad accounts from Claude Code.
Walrus + WalrusProtocol MemWal — long-term agent memory that persists across sessions and providers.
Claude Security — public beta yesterday, scans repos for vulnerabilities and drafts patches, packaged into Claude Code on the web.
Codex / GPT-5.5 — dominant alternative narrative, with most posts framed as "Codex over Claude Code now" or hybrid use (plan in Claude Code, execute in Codex).
DeepSeek V4 Pro / Flash — the cost narrative. Multiple authors running DeepSeek V4 through Claude Code or OpenCode for pennies.
Cursor / Cursor SDK — the harness story. Cursor SDK shipping yesterday made it possible to embed Cursor's agent loop in CI/CD and other products.
Hermes Agent — consistent #2 alongside OpenClaw for self-hosted, multi-channel deployments. Multiple posts describe migrating from OpenClaw to Hermes.
Skills / ClawHub / npx skills — the package manager layer. Every workflow post now ships a skills install command.
Obsidian + claude-obsidian + Tolaria — knowledge-base agents are converging on Obsidian as the substrate.
NotebookLM — the deep-research offload destination. Multiple workflows route research through NotebookLM to save Claude Code tokens.
GPT Image 2.0 / Seedance 2.0 / Higgsfield — the creative-stack triple, used together in iOS app builds, anime production, and ad creative pipelines.
Stripe Link agent wallet / Machine Payments Protocol — yesterday's Stripe Sessions launch landed in nearly every "agentic future" post.
Manus Cloud Computer — non-engineer-friendly always-on agent infrastructure, positioned against OpenClaw + Mac Mini.
Ollama + Gemma 4 + GLM-4.7-Flash — the local-stack triple for people exiting API subscriptions.
n8n + n8n-MCP — the workflow automation that Claude Code can now create and edit directly via MCP.
Trendtrack + Meta Ads CLI — the marketing-agent triple, used together to spy competitors, generate creatives, and manage ad accounts from Claude Code.
Walrus + WalrusProtocol MemWal — long-term agent memory that persists across sessions and providers.
Claude Security — public beta yesterday, scans repos for vulnerabilities and drafts patches, packaged into Claude Code on the web.
Comments