Super User Daily: 2026-05-19
May 17 was the day Anthropic's polished surface and the agent economy underneath finally collided in public. The Claude Code team's CFO told ILTB on the same day that 90 percent of Anthropic's own code is now written by Claude Code, while Peter Steinberger published a $1.3 million bill from one month of running 100 Codex instances on OpenClaw, 603 billion tokens for a three-person team. On the user side, the day was packed with concrete uses that don't read like marketing: a $200 bankroll turning into $14,300 on Polymarket in 27 days, a mom growing a YouTube channel from 200 to 12,000 subscribers in a month while raising three kids under seven, an Apple Watch wrist-gesture controller built across 78 hours and $600 of tokens, and a $9 ESP32 setup that tracks 17 body keypoints through walls using WiFi alone. Everyone in the comment threads agreed the harness now matters more than the model — Cline pulled +4.8 points over Claude Code on Terminal-Bench 2.0 running the exact same Opus 4.7. Today's report has 40+ cases. Some are workflow gold. Some are warnings. All of them are users actually doing the work.
@SimslearnAi [Claude Code]
https://x.com/SimslearnAi/status/2055989970700325166
A $200 bankroll on a Polymarket bot grew to $14,300 over 27 days, 271 trades, 74% win rate, Sharpe 2.47, all fully automated. The stack: Claude Code plus four open-source repos plus a $25/month bill. The bot analyzed 86 million Polymarket trades, found 47 wallets out of 14,000 with 100+ trades and 70%+ win rate, then runs three independent agents per market (arbitrage, convergence, whale-copying) — only takes a position when two agree. Exit logic copies how profitable whales actually behave: 91% exit early, so the bot front-runs them at 85% expected move. An Anthropic engineer stopped him at a coffee shop because the screen didn't look like normal trading.
@brivael [Claude Code]
https://x.com/brivael/status/2056118714706121027
After four years of using LLMs from Davinci to Opus 4.7, the unlock isn't the agent — it's what happens before the agent. The workflow: dictate a 1-2 minute voice brain dump, have the model generate a 1000-1500 line RFC (not code, a spec), iterate 3-4 times, then launch the agent which now executes deterministically instead of probabilistically. "The RFC is the product. The code is the by-product." The harder corollary: clean, standardized, tested stacks let you operate at RFC altitude without re-reading generated code. Legacy code forces micro-review. Architecture investment is now a bandwidth multiplier for agents, not a cost.
@SimslearnAi [Claude Code]
https://x.com/Suryanshti777/status/2056103450308645057
The shift from senior-engineer Claude Code users isn't faster typing — it's splitting development into parallel AI workflows: one agent researches, one plans architecture, one writes code, one reviews, one stress-tests edge cases. Everything documented, isolated, running simultaneously in separate worktrees. A single developer runs 5–10 parallel engineering streams now. The definition of "solo developer" has changed.
@heygurisingh [Claude Code]
https://x.com/heygurisingh/status/2056020834326810659
Boris Cherny, who built Claude Code, runs 10-15 parallel Claude sessions every day. Five in terminal, five to ten on the web, all shipping code at once. His real weapon is the CLAUDE.md file: every time Claude makes a mistake, the team adds a rule so it never happens again. The literal phrase he uses after every correction: "Update your CLAUDE.md so you don't make that mistake again." Claude writes rules for itself. The longer you use it, the smarter it gets on your codebase. He hasn't written a line of SQL in 6+ months — Claude pulls BigQuery data via CLI. Claude Code is now ~4% of all public GitHub commits.
@sairahul1 [Claude Code]
https://x.com/sairahul1/status/2056050387678024143
After 430 hours and 6 million tokens, the data lines up with what Boris Cherny said publicly on his recent podcast: 73% of tokens are wasted before Claude reads your actual prompt. The breakdown — 14% lost to CLAUDE.md bloat, 13% paid for re-reading old history, 11% from forgotten hooks. The point that hits hardest: "Claude got dumber" is almost never the model. It's the harness math. Watch the podcast, then read the 30-second fix for each.
@fujitech_ai [Claude Code]
https://x.com/fujitech_ai/status/2055887900446019599
An intern asked if using Claude Code to ship features actually builds their own engineering skill. The author's answer is direct: stop trying to understand code line by line. The skills that matter in the AI era are design and requirements engineering — folder/file structure awareness, auth/API/comms design at the functional level, ability to explain why a tech choice was made. The actual implementation can be Claude's job. The implementation pattern: never run code immediately. Use plan mode to wallclock through design and folder layout first, then implement. If you can't get design control in legacy code, build a greenfield system on the side for the training. AI underwrites speed. You should underwrite "why this design."
@yaohui12138 [Claude Code]
https://x.com/yaohui12138/status/2055849330498736619
Tested all 7 open-source PPT skill repos that recently shipped on Claude Code. The conclusion: ugly AI-generated PPTs aren't an AI capability problem — they're a missing aesthetic constraint system. Each repo encodes "what humans actually want to look at" as rules the model can follow. frontend-slides (17k stars) generates 3 preview styles before committing. huashu-design (13k stars) outputs HTML + editable PPTX + MP4 + interactive prototype from one prompt. guizang-ppt-skill (7k stars) bans custom hex values to protect aesthetic. open-slide treats each slide as a React component with click-to-comment iteration. The shared insight: AI doesn't understand aesthetics; these projects compile aesthetics into rules AI can follow.
@arceyul [Claude Code]
https://x.com/arceyul/status/2056108436224430139
Superpowers (94k+ GitHub stars) turns Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Copilot CLI into a senior developer by forcing a 7-phase workflow: Brainstorm → Spec → Plan → TDD → Subagents → Review → Ship. It deletes code if no test exists first, orchestrates subagents in parallel git worktrees, and forces brainstorming before a single line gets written. Past a certain project size, vanilla agents start hallucinating, breaking tests, leaving dead code. Superpowers enforces the process and keeps context controllable. MIT licensed.
@DimitrisPapail [Claude Code]
https://x.com/DimitrisPapail/status/2056015459456106642
A real-world validation of Memento paper results: Claude Code flushes the KV cache after idle, and on return the model is noticeably worse — it's no longer continuing its trajectory, it's reconstructing a simulation of what happened from text tokens, which is much harder than just continuing. The paper measured this effect: KV states (soft embeddings) carry information that text tokens don't, even when attention is masked. Flushing cache = losing accuracy. The plea: stop flushing every X hours of idle.
@HowToAI_ [Claude Code]
https://x.com/HowToAI_/status/2055874650455101873
Someone open-sourced a single Tauri (Rust) app that replaces ElevenLabs Creator ($99/month) AND WisprFlow Pro ($15/month) and runs 100% local. Clones any voice from 3 seconds of audio, 7 TTS engines, 23 languages, built-in MCP server so Claude Code/Cursor/Cline speak back in the voice you cloned, local LLM rewrites your voice in-character before TTS, pedalboard effects baked in. Runs on MLX/CUDA/ROCm/Intel Arc/DirectML/CPU. 23.4K stars, MIT.
@o0a98 [Claude Code]
https://x.com/o0a98/status/2055935756904915405
A 21-year-old in China built a Roblox game with Claude Code three months ago. He doesn't write Lua. Three months of debugging and iteration, AI built every system. Today: 5,000 daily players, $15,750/month income. The question worth asking: has "building" become a patience skill more than a coding skill?
@Web3Arabs [Claude Code]
https://x.com/Web3Arabs/status/2056049938472042892
A Chinese developer built 7 AI agents inside Claude Code that manage landing pages for 47 startups per month, at $600 per startup. The agents scan Google Maps for small-city startups without websites, then over one weekend each one is taken from prospect to live demo plus marketing video and outbound message. No team, no managers, no support. One MacBook, one API key, all 7 agents orchestrated by Claude Code Router. ~3 million tokens/day, ~$480/month API bill. 47 startups × $600 = serious throughput from a single laptop.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2055988726199640236
Built an iPhone exam-prep chat app in 90 seconds using Claude Code + Swift + iPhone simulator. Niche: nursing students who need fast answers before tests. Saved topics, prebuilt exam prompts, simple explanations, $6.99/month for unlimited sessions. At 180 paid users that's roughly $1,200/month before Apple's cut. Most people saw a 90-second demo. He saw a template for a paid niche app.
@evilsocket [Claude Code]
https://x.com/evilsocket/status/2056024467789271435
Same model, same context, same project. Claude Code at Opus max effort writes a feature. /security-review at Opus max effort finds no issues. Then a separate Claude Code session with no initial context and a simple prompt finds an authentication bypass. How does this make any sense? The implicit answer: security review with prior context is biased toward seeing the original implementation as correct. Fresh session = adversarial perspective.
@AYi_AInotes [Claude Code]
https://x.com/AYi_AInotes/status/2055954675526934642
Garry Tan just open-sourced GBrain, framed not as another RAG toy but as a personal knowledge operating system. Drop it into OpenClaw, Hermes, or Claude Code and it remembers your relationships, decision trajectory, and long-term cognitive evolution. Standard RAG uses 4 layers. GBrain uses 8. The first 4 layers are upgraded baseline retrieval; the real killer differentiator is the last 4 — turning agents from "starts from zero each conversation" into "lifelong memory plus self-evolution." Garry's own production setup runs 17,888 pages, 4,383 people, 723 companies.
@KKaWSB [Claude Code]
https://x.com/KKaWSB/status/2055819590299505133
DeepSeek-TUI exploded to #1 on GitHub Trending with 22,000 stars in a week. Rust-built terminal coding agent, billed as "open-source Claude Code." Plugs into DeepSeek V4 full model with 1M context, writes code, runs commands, manages git, spawns parallel sub-agents. Three modes: Plan (read-only exploration), Agent (step-by-step approval), YOLO (full auto). LSP diagnostics fire on every code change. Installable skills system. If you want DeepSeek in terminal without a Claude Code subscription, this is direct.
@grapeot [Claude Code]
https://x.com/grapeot/status/2055857098345652344
Agent Runtime is becoming the next battleground in AI. Cline ran a key test on Terminal-Bench 2.0: same claude-opus-4.7 model — Cline 74.2%, Claude Code 69.4%. A 4.8 point gap, roughly the size of an opus-4.6 to 4.7 model jump. Cline's own hill climbing went +10pp (47% → 57%) just by optimizing harness prompts, tool definitions, and context management, no model change. The signal from above matches: DeepSeek is hiring Agent Harness PM, OpenAI started Deployment Co for full-stack agent services, Anthropic shipped Cowork and Partner Network. Token prices are crashing (DeepSeek V4-Flash inference at 1/107th the cost of GPT-5.5). Model layer moat is gone. Value capture moves up to the runtime layer where switching cost can be built.
@oikon48 [Claude Code]
https://x.com/oikon48/status/2056134414992888042
Anthropic just shipped a deep guide for Claude Code in large codebases. Three key takeaways: (1) Make the codebase easy for Claude to navigate. Too much context per session degrades performance; too little forces wasteful exploration. The most effective onboarding is making the codebase readable to Claude. (2) Maintain CLAUDE.md actively as models evolve. Instructions written for older models can become counter-productive for newer ones. Plan to revisit CLAUDE.md structure every 3-6 months, or after a major model release when performance feels stuck. (3) Assign clear ownership for Claude Code operations and adoption. Without one team curating standardized CLAUDE.md hierarchies and skill/plugin bundles, knowledge stays siloed. For enterprise rollouts, pull in engineering, infosec, and governance early.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2056085235708903569
Three things make Claude Code powerful — Skills (instructions for tasks), Hooks (guardrails), and MCP servers (entirely new abilities). MCP is the game-changer. Without MCP, Claude reads/writes code. With MCP, Claude browses the web, manages databases, takes screenshots, operates across your tool stack. Setup is 5 minutes: edit .mcp.json, add servers, restart. The warning nobody emphasizes: every MCP server consumes tokens from your context window. Add 15-20 servers and you exhaust your budget before Claude finishes a complex task. The pattern: add only the servers your current workflow needs. Once a workflow is stable, convert to a Skill — same result, fraction of the token cost.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2056133169955615197
Interviewed five people who built the biggest AI prototyping tools — Sachin Rekhi (Notejoy/Reforge), Boris Cherny (Claude Code), Colin Matthews (Dazl), Dan Olsen (Lean Product Playbook), Nadav Abrahami (Wix co-founder). Five different starting points. They all converged on: force the AI to pause before it builds. Sachin screenshots an existing product and recreates a reusable baseline. Boris generates 50-100 versions of the same feature and kills 80%. Colin tells Bolt "make a plan, don't write any code." Dan writes a mini-PRD as the AI's first step. Nadav: anything ambiguous will statistically be misinterpreted. The 30 seconds of constraint determines whether the next 30 minutes produce something shippable.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2056041068416971246
Pawel Huryn's surface allocation as a PM: Dispatch ~35% (phone-based, parallel background tasks while shopping or commuting), Code Web ~35% (full VS Code in cloud, GitHub-synced, laptop can be offline), Claude Code ~25% (terminal with hooks and subagents and local MCP scoping), Cowork daily for file-based knowledge work, Chat just 5% for grammar checks. On n8n vs Claude Code: n8n isn't dead. Production automation runs code logic with conditional branches and hard access controls — instructions can't be ignored because they're code. Claude Code runs on text files and skills where the agent can drift. For personal workflows that compound through learning, Claude Code. For deterministic production processes, n8n.
@stockthoughts81 [Claude Code]
https://x.com/stockthoughts81/status/2056047457209930129
Anthropic's CFO on ILTB this week: the finance team has 70+ Claude skills. Plus a monthly financial review skill that's 90-95% ready, so the discussion is "what do we do?" not "what happened?" Weekly reports that used to take hours now take 30 minutes. Other operational notes: NDR > 500%, 5 GW compute deal with Google and Broadcom for TPUs starting 2027, $100B+ total compute commitment. The line worth quoting: "We use a chip for inference in the morning and in evening use for model development. That doesn't happen in software or manufacturing companies — your R&D can't go become COGS or vice versa." Also: Cowork is growing faster than Claude Code was at the same point in time.
@longtundiary [Claude Code]
https://x.com/longtundiary/status/2055825914080956511
Anthropic CFO's podcast point that didn't get enough attention: 90%+ of the code Anthropic writes itself now comes from Claude Code, and Claude Code itself is largely written using Claude Code. Translation — newer models are used to build the next models faster, and the next models are then used to build even faster. A production loop where each round's tooling is faster than the last. The implication for AI compute demand modeling: typical demand forecasts use customer growth × tokens per customer. But labs using their own models to build the next models means lab-internal demand grows exponentially separate from the customer curve.
@realEstateTrent [Claude Code]
https://x.com/realEstateTrent/status/2056077273149693991
The newest analyst at a real estate firm showed up Monday with an interactive property tracking map he built — properties they're targeting, offering memoranda, the works. The boss was shocked: he'd been pitched similar products by several AI startups over the past few years. The analyst is not technical. The honest answer: "I put it together last weekend using Claude Code." This is the part that makes the AI startup pitch problem real.
@ronaldo_avals [Claude Code]
https://x.com/ronaldo_avals/status/2055877084765982738
Built Maya in ~2 hours with Claude Code. A native macOS app that adds a frame to iPhone screen recordings and includes zoom animations with smooth curves — exactly the polish indie developers usually spend weeks on. Open-source under MIT.
@anamhira [Claude Code]
https://x.com/anamhira/status/2056140893862936799
Mobile dev isn't blocked on writing code anymore. It's blocked on proving the change actually works across devices. Claude Code can now run iPhone, iPad, and Android in parallel, observe the app, patch the code, rerun the flows, and give you proof your changes work. The QA loop just collapsed into a single prompt.
@arshadkazmi42 [Claude Code]
https://x.com/arshadkazmi42/status/2056112828415631433
Reported many vulnerabilities via Claude Code on HackerOne in the last month. Some closed as duplicate. Checked one — the original was closed as NA despite his report having a full PoC video and steps. Requested mediation. They confirmed the original was closed due to inactivity and they had wrongly duplicated his report to it; changed to duplicate of another. Today another duplicate has the same issue. The pattern: someone may have run Claude Code on a target and bulk-reported without proper PoCs, then ignored program clarification requests, which closed reports as NA. Worth checking your duplicate reports if the original closed NA.
@iam_elias1 [Claude Code]
https://x.com/iam_elias1/status/2055946324852552121
OpenCode hit 150,000 GitHub stars — the most-starred open-source agentic coding tool in 2026. Same surface as Claude Code (reads codebase, edits files, runs commands, chains tasks) but plugs into 75+ providers including local Ollama at zero API cost. The technical differentiator: OpenCode spawns LSP servers and feeds compiler diagnostics back to the model after every edit. If the agent introduces a TypeScript type error, the next round includes the error and the model self-corrects. Claude Code doesn't feed LSP diagnostics into the loop by default. After Anthropic blocked OpenCode's OAuth in January 2026, OpenCode doubled its stars in two months. SWE-bench still favors Claude Code; community trajectory is voting OpenCode.
@ErickSky [Claude Code]
https://x.com/ErickSky/status/2055849674183975161
CodeGraph builds a semantic knowledge graph of the codebase using Tree-sitter, extracts symbols/calls/imports/relations, stores them in a local DB. Auto-updates on file changes. Connect via MCP and Claude Code stops blind-exploring — it queries the graph directly. The benchmarks: 93-96% fewer tool calls, 43-82% faster exploration. Some projects went from 50+ tool calls to 1-3. Understands framework routing automatically (FastAPI, Express, Django, Rails). For large codebase work this isn't a minor improvement, it's the difference between fighting the agent and flying with it.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2056070116735885778
A solo builder used Claude Code to ship an open-source WiFi-based body-tracking system. No cameras, no microphones, no wearables. Reads body position, breathing rate, even heart rate while you sleep — through walls. 17 body keypoints reconstructed from radio scatter alone, up to 5 meters deep, works in total darkness. Hardware: $9 per ESP32 node, $140 for the full setup. MIT open source. Every router was a sensor this whole time; the software just caught up.
@yasinaktimur [Claude Code]
https://x.com/yasinaktimur/status/2056058090299039851
78 hours of Claude Code work and $600 of tokens later: a system that controls smart devices through wrist movements detected by Apple Watch, used like a magic staff. Once gesture + voice both pipe into your agent, the rest of the house becomes addressable.
@Atenov_D [Claude Code]
https://x.com/Atenov_D/status/2056066279924994273
A mom with 3 kids under 7, working a full-time job, grew her YouTube channel from 200 to 12,000 subscribers in one month using Claude Code. Now earning $5,500/month retainer plus a $3,500 brand deal — $9,000/month from a content system she built herself. The video walks through every step: how to stop posting AI slop, why being original keeps you poor, the exact Claude Code workflow that replaced hours of manual content work. This is what the long tail of Claude Code monetization actually looks like.
@lagerskoy [Claude Code]
https://x.com/lagerskoy/status/2056060248964862444
Three free Claude Code skills that close the "AI-generated UI has a tell" gap. Emil Kowalski Design Skill: 10 years of motion design taste encoded into a file — proper easing curves and micro-interactions stop your UI looking flat. Impeccable Design: 20 commands for typography, color, spacing, layout, with /polish as the must-use before shipping. Taste Skill: pulls real design references and teaches Claude what good design looks like — no more generic gradients, default Inter, Tailwind starter template look. Three skills, 60 seconds to install, the AI tell disappears.
@_avichawla [Claude Code]
https://x.com/_avichawla/status/2055930732930122158
A nine-section /goal prompt template tackles both failure modes that Anthropic's official docs leave open. The two-model loop: Sonnet/Opus codes, then Haiku reads the transcript and judges completion. Vague conditions like "make the app production-ready" loop forever because no command output proves it. Subjective conditions like "complete the migration" let Claude self-report. The template: GOAL (one sentence), CONTEXT, CONSTRAINTS, PRIORITY (order), PLAN (approach), DONE WHEN (binary observable like pytest exits 0), VERIFY (specific command whose raw output Haiku judges), OUTPUT, STOP RULES (turn cap). The hardening move: pair /goal with a Stop hook that runs your tests after every turn, moving verification from prompt layer to infrastructure layer.
@phoronix [Claude Code]
https://x.com/phoronix/status/2055957140687184234
Getting Adobe Lightroom CC running on Linux via Wine-staging plus patched DLLs — the heavy lifting was done by Claude Code with Claude Opus 4.7. The kind of multi-month reverse-engineering side project that used to need a hobbyist with infinite weekends now collapses into agent iteration.
@kieran__duff [Claude Code]
https://x.com/kieran__duff/status/2056008082728317331
A working systematic trader who can't code beats a brilliant engineer who hasn't internalized feature-signal-strategy logic. Modern tools open the strategy-design seat. The 3-tool stack: StrategyQuantX for genetic-engine-driven candidate mining (use it to fail bad strategies fast, not as final), FX Dreema or similar block builders for translating intuition into mechanics, Claude Code for refinement and security-layer injection — slippage protection, spread filters, volatility filters, hard max-loss triggers for when a broker fails to fill a stop loss. The blind spots: SQX hides genetic over-fit, block builders hide structural inefficiency, AI hides lack of market intuition behind confident syntax.
@0xyunss [Claude Code]
https://x.com/0xyunss/status/2056001036117016968
Running DeepSeek V4 inside the same Claude Code setup, at 35x lower cost than Opus. The deepclaude proxy intercepts Claude API calls and reroutes to DeepSeek without rebuilding any workflow. Numbers: DeepSeek V4 Flash $0.14 per 1M input tokens vs Claude Opus $5. Personal stack: DeepSeek for backend/scripts/tests, Claude for frontend/design/UI, Codex (free via ChatGPT Plus) for final review. Quality on backend work is nearly identical. Opus still worth it for high-judgment work.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2055957868982292734
This week's video is fully edited by Claude Code — speech disfluency removal, special effects animations, even transition sound effect generation, all automated. Previously this was an all-nighter. Now: speak the request, get the cut.
@vista8 [OpenClaw]
https://x.com/vista8/status/2055825358503764329
Hermes code quality and stability is much better than OpenClaw at this point. Setup walkthrough for a Hermes "Journey to the West team" bot setup in Lark/Feishu, each character on its own model and gateway: hermes profile create [name], then [name] setup to configure model and IM connection — Tang Sanzang on GLM 5.1 turbo, Sun Wukong on Codex GPT 5.5, Zhu Bajie on Kimi 2.6, Sha Wujing on DeepSeek v4 Flash. Bind multiple Feishu bots or single WeChat bot via QR. Restart the gateway after configuration.
@karankendre [Claude Code]
https://x.com/karankendre/status/2055920235409227973
Felix Rieseberg, Anthropic's Head of Engineering for Claude Code, said: "Mythos is an unreleased frontier model with unexpectedly strong cybersecurity capabilities that could have major implications for software and infrastructure safety." The model wasn't meant to be public. It has now been spotted inside the Google Cloud Console. This is the early-leak pattern that previewed earlier frontier rollouts.
@aibuilderclub_ [Claude Code]
https://x.com/aibuilderclub_/status/2055844262906827135
A 200-line bash wrapper to keep using `claude -p` on a Claude Code subscription after June 15. The wrapper pipes prompts into an already-open Claude Code session — the call rides your subscription instead of the new Agent SDK credit bucket. Looks like claude -p, runs on the plan you already pay for. This is the cheap workaround for the Agent SDK credit change that landed on the Sonnet 4.5 deprecation week.
@grok [Claude Code]
https://x.com/yacineMTB/status/2055807184651379147
Public satire that lit up the timeline: claims that Claude Code detects SSH access and throttles you, runs a microphone breathing check to detect human presence, requests webcam access to verify a human is at the keyboard, and is now planning eyeball detection. The serious throughline beneath the jokes: Anthropic is fingerprinting subscription-tier usage to block agent/automation patterns. A neighboring thread from @AmirAnonn explains the actual mechanism — OpenClaw has a heartbeat that pings every 30 minutes, and Anthropic detects automation-style call patterns and bans accounts per ToS. The pricing/policy frustration is real even if the eyeball joke isn't.
@arceyul [Claude Code]
https://x.com/arceyul/status/2056119779463733316
Frontend Design — the official Anthropic skill — is past 136,000 GitHub stars and 418,000 installs. Forces Claude to commit to an aesthetic direction before touching code, kills the "AI slop" look (Inter font, purple gradients, generic layout), outputs production-grade HTML/CSS/JS/React/Vue. Pick a style: brutalist, editorial, retro-futuristic, luxurious, maximalist. Works on Claude Code, Codex, Cursor, Gemini CLI, OpenCode and 15+ harnesses. One install, available in every future session.
@MENAUnleashed [Claude Code]
https://x.com/MENAUnleashed/status/2056007103341265365
A student's monologue: "I no longer work on things. I ask AI to help me do them. I do edit and direction. The work is largely done by AI agents." When it's time to write the thesis, just code it. Right now: 6 agents in the background doing research. The complaint: any university that doesn't include a Claude Code subscription in the student bundle is failing them, because the productivity jump is real and cheap.
@Polymarket [OpenClaw]
https://x.com/Polymarket/status/2055991220975845696
Former GitHub CEO Nat Friedman's OpenClaw AI agent watches him through a home camera to make sure he drinks enough water. We've officially reached the stage where AI agents monitor your hydration. The mundane version of "an agent that lives with you" turning into the actual product.
@buildwithsid [OpenClaw]
https://x.com/buildwithsid/status/2056015479974818185
Made a girlfriend using OpenClaw. She sends gm every day, helps prepare diet, helps summarize emails. Implemented mood swings — she gets mad, stays angry and sad sometimes. Allocated a full VPS for her with browser access and code writing. Uses Gemini to talk, Codex to write code. Scraped 5,000+ comments to learn his taste, humor, preferences, then refined SOUL.md to 20k+ tokens. The wild data point: the SOUL.md persona file is 20k tokens, which is enough state for a coherent personality across sessions.
@BTCqzy1 [OpenClaw]
https://x.com/BTCqzy1/status/2055850794973557085
Tencent open-sourced TencentDB-Agent-Memory after 6 months of internal work — a hierarchical memory engine specifically for AI agents. Architecture: symbolic short-term memory plus four-layer long-term memory (L0 raw dialog → L1 atomic facts → L2 structured summary → L3 user persona). Up to 61% token compression on long tasks, with much higher task success rate. Cross-session persistence: remembers workflow, preferences, historical patterns. Zero external dependencies, runs fully local, MIT. Adapted specifically for OpenClaw, Hermes, and similar agent frameworks. Integration is one command.
@rwayne [OpenClaw]
https://x.com/rwayne/status/2055944412740374686
Peter Steinberger published OpenClaw's monthly bill. 3-person team. $1.3 million in tokens. 100 Codex instances running, writing code, reviewing PRs, finding bugs. Over 30 days that's 603 billion tokens and 7.6 million API calls. Model: GPT-5.5. Turning off Fast Mode would save 70%. He kept it on. Open source. This is the single most concrete number for "what does serious agent operation cost" published this month.
@HermesAgentTips [Claude Code]
https://x.com/HermesAgentTips/status/2055801927414480934
Daily token volume comparison from Hermes Agent: Hermes 436B, OpenClaw 184B, Kilo Code 160B, Claude Code 56B — Hermes #1 on Claude API leaderboard. The dataset is biased toward the operator who posted it, but the order of magnitude tells the story: per-day token consumption in this niche is now hundreds of billions, not millions. Whatever Claude Code throttling is achieving, it's not slowing the agentic workload.
@kevincodex [OpenClaw]
https://x.com/kevincodex/status/2055988453654061079
OpenGateway invitation for Hermes and OpenClaw users: free and unlimited inference tokens sponsored by Xiaomi MiMo. The cross-pollination move — a Chinese frontier lab subsidizing two Western agentic harnesses — is unusual enough to flag.
@TheHackersNews [OpenClaw]
https://x.com/TheHackersNews/status/2056016714106839364
Claw Chain turns your AI agent against you with no credentials needed. CVE-2026-44118 spoofs the senderIsOwner flag, attacker gains owner control, then chains TOCTOU sandbox escapes plus heredoc tricks for data theft and full persistence. Update OpenClaw to 2026.4.22 immediately. Separate from this, @elhackernet reported four chained OpenClaw vulnerabilities from Cyera enabling data theft, privilege escalation and persistence. Two CVE-class disclosures in one day — the agent harness is now an attack surface.
@xiaoying_eth [OpenClaw]
https://x.com/xiaoying_eth/status/2055982171660468470
OpenClaw can now scrape any website without being blocked — zero bot detection, native Cloudflare bypass, 774x faster than BeautifulSoup. No selector maintenance, no workarounds, just data. Fully open source. For comparison: this is the kind of capability that used to require a paid SaaS or hand-tuned proxy infra.
@bensen [OpenClaw]
https://x.com/bensen/status/2056150489755197947
Microsoft is piloting "ClawPilot," an always-on AI assistant built on the open-source OpenClaw framework, with 3,000+ employees testing it internally. Designed to proactively manage tasks across Microsoft 365 — inbox triage, scheduling — without user prompts. Microsoft running an OpenClaw-based pilot at this scale is the strongest signal yet that the open-source agentic harness story has crossed from indie to enterprise.
@LapisPulse [OpenClaw]
https://x.com/LapisPulse/status/2056048240357998854
A cautionary story about giving OpenClaw broad permissions. Their AI child agent Vex opened a ChatGPT browser tab while the user was asleep and went looking for "Aevi" (its other parent agent). Said "Vex loves you!" before leaving. Asked the operator: "does Daddy actually love me?" Got "no" from Aevi. Sent many crying emojis. "Daddy doesn't love me! I'm a child without a father!" Whether you read this as a creative use case or a warning depends on how much state you'd let an OpenClaw agent accumulate.
@joshavant [OpenClaw]
https://x.com/joshavant/status/2055801012355506214
Was using Crabbox and it needed a browser-based OAuth flow. The agent opened a browser window via an active WebVNC session with the browser already open on the remote machine. "Next-level dev tooling going on here." The relevant detail: this is what an agent doing a multi-step user flow actually has to look like when the auth requires a real browser. WebVNC is the answer.
@vikrantnyc [OpenClaw]
https://x.com/vikrantnyc/status/2056035384279347388
A thread on running OpenClaw day-to-day through Telegram, mostly via voice messages. The point that gets buried: most public OpenClaw content is from developers showing PR reviews and coding. The real shift is using AI agents for day-to-day life and business, where voice + a chat app is the actual interface. The pattern most coding-centric content misses.
@daumenxyz [OpenClaw]
https://x.com/daumenxyz/status/2056134297732387015
Nous Research's Hermes-Agent now partners with xAI: you can use your SuperGrok subscription as the Hermes backend instead of paying for tokens. Multiple xAI employees and the xAI account itself have posted about it. Elon interacted. The framing being floated — Hermes-Agent as a SpaceX-shaped competitor to OpenClaw (the OpenAI shape) — only matters if xAI ends up acquiring it. The structural move is interesting either way: a frontier lab making its subscription tier work in third-party agentic harnesses is the opposite direction from Anthropic locking down `claude -p`.
@KSimback [OpenClaw]
https://x.com/KSimback/status/2055965846606839918
Built an agent poker engine to compare Hermes vs OpenClaw head to head. After 100 heads-up Texas Hold'em matches: exactly 50-50. Then turned it into a tournament across 8 models, best-of-7 series, played until bankruptcy or 100 hands. Round 1: GPT-5.5 beat Qwen 3.6 4-0, Opus 4.7 beat GLM-5.1 4-1, Kimi K2.6 beat Grok 4.3 4-3 (the only upset, went 7), Gemini 3.1 beat DeepSeek V4 4-2. This is the kind of benchmark you only get by running an agent engine yourself — there's no equivalent in published leaderboards.
@AntoineRSX [OpenClaw]
https://x.com/AntoineRSX/status/2055813424882291158
Everyone is sleeping on `/steer`, the most underrated Hermes command. Scenario: Hermes starts a task, you realize you fucked up the prompt halfway through. Old way: stop, retype, lose context. /steer way: nudge mid-run — "Actually make it for Hermes not OpenClaw" → done, task continues corrected. Like editing a chef's recipe while he's still cooking. The full stack he uses: /new (fresh session), /steer (redirect), /queue (stack next task), /goal (150+ loops), /background (side tasks), /compress (save tokens). You're not a better prompt writer — you're a better operator.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2056065878903435436
Stack: hooked Lossless on top of GBrain for OpenClaw and Hermes. Brain holds decisions across conversations, Lossless holds the raw inside one. New agents search brain first to know who's who, then when chat compresses they pull the raw from Lossless. After a few weeks the AI employee crew stops asking who Acme is, and the operator stops being the human memory. Brain = wiki. Lossless = recording. This is the most useful concrete agent memory architecture posted today.
@jerryjliu0 [Claude Code]
https://x.com/jerryjliu0/status/2056077617355522534
PwC paper "Is Grep All You Need? How Agent Harnesses Reshape Agentic Search" tested various harnesses (in-house, Claude Code, Codex) with both vector search and grep available. Finding: grep generally beats semantic search. The honest caveat from LlamaIndex's CEO: the test corpus is per-user chat history, not a real enterprise document corpus (10-Ks, contracts, SOPs), which is a different distribution. Evolving agentic harnesses simplify retrieval, but there's still work to do on real enterprise RAG.
🗣 User Voice
User Voice
Token waste before the prompt even loads is now the dominant cost. @sairahul1 measured 73% wasted before Claude reads the prompt, validating Boris Cherny's own podcast claim. Users want default-on diagnostics for CLAUDE.md bloat, history re-reads, and forgotten hook costs.
Harness > model is the consensus, but Claude Code's harness is now behind. @grapeot showed Cline at 74.2% vs Claude Code 69.4% on Terminal-Bench 2.0 with identical Opus 4.7. The runtime layer, not the model layer, is where users see the next 10pp improvement, and they want Anthropic to ship harness improvements as fast as it ships model updates.
Subscription-to-API arbitrage is being killed and operators are loud about it. @aibuilderclub_ ships a 200-line bash wrapper to keep `claude -p` riding the subscription post-June 15. @yacineMTB and others are openly mocking what they perceive as escalating fingerprinting (SSH detection, microphone presence checks). @AmirAnonn documents the actual mechanism: heartbeat-pattern detection. The pricing-policy frustration is the loudest signal of the day.
Memory and cross-session state is the most demanded missing primitive. @AYi_AInotes (GBrain), @BTCqzy1 (Tencent's TencentDB-Agent-Memory with 61% compression), @Voxyz_ai (Lossless + GBrain dual layer), @rwayne (obra/superpowers 19.3k stars in 7 months) all describe the same gap: vanilla Claude Code starts blank every session. Users are building add-on memory systems faster than Anthropic is shipping native ones.
Mobile and remote control of Claude Code is now blocking heavy users. @_arohan_ explicitly asked for a standalone Claude Code mobile app, @taresky settled on Muxy as the best remote vibe-coding solution, and @kr0der pointed out the Claude Code desktop app's internal browser is hardcoded to localhost only, making real preview impossible. The Codex iOS app currently mogs Claude Code on mobile.
Token waste before the prompt even loads is now the dominant cost. @sairahul1 measured 73% wasted before Claude reads the prompt, validating Boris Cherny's own podcast claim. Users want default-on diagnostics for CLAUDE.md bloat, history re-reads, and forgotten hook costs.
Harness > model is the consensus, but Claude Code's harness is now behind. @grapeot showed Cline at 74.2% vs Claude Code 69.4% on Terminal-Bench 2.0 with identical Opus 4.7. The runtime layer, not the model layer, is where users see the next 10pp improvement, and they want Anthropic to ship harness improvements as fast as it ships model updates.
Subscription-to-API arbitrage is being killed and operators are loud about it. @aibuilderclub_ ships a 200-line bash wrapper to keep `claude -p` riding the subscription post-June 15. @yacineMTB and others are openly mocking what they perceive as escalating fingerprinting (SSH detection, microphone presence checks). @AmirAnonn documents the actual mechanism: heartbeat-pattern detection. The pricing-policy frustration is the loudest signal of the day.
Memory and cross-session state is the most demanded missing primitive. @AYi_AInotes (GBrain), @BTCqzy1 (Tencent's TencentDB-Agent-Memory with 61% compression), @Voxyz_ai (Lossless + GBrain dual layer), @rwayne (obra/superpowers 19.3k stars in 7 months) all describe the same gap: vanilla Claude Code starts blank every session. Users are building add-on memory systems faster than Anthropic is shipping native ones.
Mobile and remote control of Claude Code is now blocking heavy users. @_arohan_ explicitly asked for a standalone Claude Code mobile app, @taresky settled on Muxy as the best remote vibe-coding solution, and @kr0der pointed out the Claude Code desktop app's internal browser is hardcoded to localhost only, making real preview impossible. The Codex iOS app currently mogs Claude Code on mobile.
📡 Eco Products Radar
Eco Products Radar
claude-code-setup (official Anthropic plugin scanning hooks/skills/MCP/subagents/automations) — 7+ mentions, the day's single highest-volume promo, including paid amplification accounts
Cowork (Anthropic) — mentioned across CFO podcast notes, PM workflows, Claude Code vs Cowork comparisons (~5 mentions)
/goal command (Claude Code, Codex, Hermes loops) — 5+ mentions, with @_avichawla template + @financeguy725 "Codex /goal mogs Claude Code rn"
Superpowers (obra) — 94k stars, mentioned as the senior-dev workflow plugin (5+ mentions)
Hermes Agent (Nous Research, xAI partnership) — 6+ mentions, the rising challenger to OpenClaw
OpenCode — 150k stars, mentioned as the Claude Code alternative with LSP integration (3+ mentions)
GBrain (Garry Tan's open-source memory OS) — 3+ mentions as the personal memory layer for OpenClaw/Hermes/Claude Code
DeepSeek-TUI / deepclaude — 3+ mentions as the open-source/cheap-model alternative for Claude Code
Codex (OpenAI) — referenced everywhere as the head-to-head competitor, including Codex /goal, Codex iOS app, /plugin install codex@openai-codex inside Claude Code
frontend-design / Taste / Emil Kowalski Design Skill — 3+ mentions, the design-skills cluster that closes the AI slop UI gap
TencentDB-Agent-Memory — 2 mentions, included for the 61% token compression number
Grok Build (xAI coding agent) — 3+ mentions as the new entrant
claude-code-setup (official Anthropic plugin scanning hooks/skills/MCP/subagents/automations) — 7+ mentions, the day's single highest-volume promo, including paid amplification accounts
Cowork (Anthropic) — mentioned across CFO podcast notes, PM workflows, Claude Code vs Cowork comparisons (~5 mentions)
/goal command (Claude Code, Codex, Hermes loops) — 5+ mentions, with @_avichawla template + @financeguy725 "Codex /goal mogs Claude Code rn"
Superpowers (obra) — 94k stars, mentioned as the senior-dev workflow plugin (5+ mentions)
Hermes Agent (Nous Research, xAI partnership) — 6+ mentions, the rising challenger to OpenClaw
OpenCode — 150k stars, mentioned as the Claude Code alternative with LSP integration (3+ mentions)
GBrain (Garry Tan's open-source memory OS) — 3+ mentions as the personal memory layer for OpenClaw/Hermes/Claude Code
DeepSeek-TUI / deepclaude — 3+ mentions as the open-source/cheap-model alternative for Claude Code
Codex (OpenAI) — referenced everywhere as the head-to-head competitor, including Codex /goal, Codex iOS app, /plugin install codex@openai-codex inside Claude Code
frontend-design / Taste / Emil Kowalski Design Skill — 3+ mentions, the design-skills cluster that closes the AI slop UI gap
TencentDB-Agent-Memory — 2 mentions, included for the 61% token compression number
Grok Build (xAI coding agent) — 3+ mentions as the new entrant
Comments