Super User Daily: 2026-04-27
Yesterday belonged to the trading bots. An anonymous engineer turned $200 into $14,985 in 31 days running a Polymarket signal-extractor on a $5 VPS, with 80% of the code written by Claude Code. A 15-year-old in China shipped a Fortnite tycoon map for $23K. Another anon hammered through 86 million trades to rank wallets that nobody else copies yet. The pattern repeats: the people winning aren't reading Karpathy quotes — they're letting Claude Code grind through real-world arbitrage at 2am while they sleep. Meanwhile Anthropic finally posted the Claude Code postmortem (three harness regressions, all fixed), Boris Cherny dropped a 30-minute workshop that ate every coding feed, and DeepSeek V4 Pro started showing up in Claude Code config files at 25% of Anthropic's price.
@0xwhrrari [Claude Code]
https://x.com/0xwhrrari/status/2048098225471533449
31-day Polymarket experiment: $200 seed, 10 wallets filtered out of 14,000 by Opus 4.7, 86M historical trades parsed in 40 minutes on a $5 VPS. Four Python scripts wrote 80% by Claude Code on a weekend — profile_builder, event_listener, signal_processor, position_manager. The trick wasn't the code. It was finding wallets nobody copies yet, with single-category competence (83% win rate on crypto, 51% on politics) and 21-minute median entry delay after the source. End result: +$14,985 net, 253 trades, 72.7% win rate, Sharpe 2.67.
@MatznerJon [Claude Code]
https://x.com/MatznerJon/status/2048159736177283357
Built a website QA agent for a marketing agency that runs 50+ tests on every client site before launch. Clicks every button, submits every form, captures screenshots, flags broken links and broken layouts across desktop and mobile. Cost to build: $500 plus ongoing AI usage. No retainer, no seats. The pitch: "the boring part of every website launch, automated." Replicable across web shops, SEO firms, franchise groups.
@MoonDevOnYT [Claude Code]
https://x.com/MoonDevOnYT/status/2048185361357234639
288 Polymarket trades a day at a 60% win rate, all from a Claude Code-built bot. The trick is not standard candles — it's tick-level CVD (cumulative volume delta) computed by iterating every single trade and adding volume on buys, subtracting on sells. Most retail looks at 1-minute OHLC and misses half the story. CVD reveals when buyers are aggressive even if price hasn't moved yet. MACD 3-15-3 setting catches every five-minute cycle.
@cherry_mx_reds [OpenClaw]
https://x.com/cherry_mx_reds/status/2048063265314340932
Real engineering work, not a demo: starting late March, the team intentionally cut OpenRouter aggregate token usage by ~35% (down to ~400B tokens) ahead of the Apr 7 OpenClaw release. Different paths agents hit all day — oversized tool results, compaction/overflow recovery, background and subagent context, prompt caching. His piece was tool-result/overflow path; teammates handled cache boundaries, deterministic tool ordering, subagent light context. The kind of unglamorous infra work that compounds.
@cesarrpol [OpenClaw]
https://x.com/cesarrpol/status/2047954319261974594
Production-grade rollback playbook when an OpenClaw release breaks: three commands — npm install -g [email protected], systemctl restart openclaw, openclaw doctor. Then the test that actually matters: real messages through Slack and WhatsApp. ~15 minutes, no data loss. The whole post is a reminder that agent platforms now need ops discipline, not just prompts.
@aakashgupta [Claude Code/OpenClaw]
https://x.com/aakashgupta/status/2048039415491203338
The Mac mini agent pattern: install OpenClaw on a dedicated Mac mini in your closet, message it on WhatsApp, it runs the work and pings you back. Three things that make it different from Claude Code — delegation through channels you already use (so you stop being the bottleneck), full machine sandboxing (the Mac mini IS the sandbox, $599 to nuke instead of your real env), model agnosticism (no Anthropic rate limits). 32GB+ Mac minis now have a 10-18 week wait time as PMs buy them as personal AI compute boxes.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2048013195563172162
Workflow template v2 after 1.5 months of heavy use. 14 P0 rules, patterns.md at 2,366 lines, and the hard-earned insight: rules just hope you stay disciplined. Hooks are what actually enforce. Added Layer 3 — rules get force-injected before Claude moves. No longer trusting the model to remember.
@KingBootoshi [Claude Code]
https://x.com/KingBootoshi/status/2047906118991798298
Overnight orchestration: Claude Code as manager spawns specialized subagents per task — research, scoping, implementation, review. One workflow can run 3-4 hours. Manager keeps small context window because subagents handle the heavy lifting. Sometimes runs multiple workflows simultaneously, with one agent handling 16+ subagents. Combines with eslint custom rules, strict design patterns, /skill chains. Calls it "engineering workflows automated with detailed agents," not vibe coding.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2047849886406599165
Universal starting prompt for in-flight projects. Reread AGENTS.md, then spin up a swarm of 4 Codex instances and 2 Claude Code instances, periodically clearing build artifacts to avoid running out of space, using the looper tool every 4 minutes to feed instructions to idle agents. Mark stalled beads as open again. Code review workflow only kicks in once all beads are done. The prompt itself is essentially a meta-controller for a small fleet of coding agents.
@MasterBorder [Claude Code]
https://x.com/dpetersen32/status/2048168735970619492
Built MasterBorder for the "Built with Opus 4.7" hackathon (500 selected from 20,000+ applicants, 76 countries). Solves trade compliance for exporters: input product details and target markets, six concurrent Claude Opus 4.7 agents research every required data point and produce a source-cited harmonized report in under 60 seconds. Replaces a 5-7 day, $2,000+ workflow.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2048024545710289260
$750 → $100/mo on Claude Code, same PM workflow. Anthropic fixed three of their bugs and reset all subscriber limits, but four root causes still live on your side: cache misses (lock tools and model at session start, target ~90% hit rate on the 5-min cache), context bloat (compact early, spawn subagents, /clear often), wrong model/effort (default reasoning burns ~2× medium, never switch model mid-session), wrong input format (use agent-browser instead of Claude in Chrome, pdftotext instead of Read, consider a code graph).
@yusaku_0426 [Claude Code]
https://x.com/yusaku_0426/status/2047910777261350970
Obsidian + Claude Code knowledge base architecture. Three-layer file structure: raw (raw material), wiki (AI-edited dictionary), schema (≒ CLAUDE.md). Four operating cycles: Ingest (accumulate knowledge), Compile (auto-generate summaries and TOCs), Query (questions update the wiki), Lint (housekeeping). The whole thing turns a Markdown vault into a personal Wikipedia that Claude can navigate.
@sitinme [Claude Code]
https://x.com/sitinme/status/2047894872473092115
Obsidian Mind, an open-source project that gives Claude Code an external "non-amnesiac brain." Stores goals, project progress, key decisions, fixes, work logs in Obsidian as Markdown. On startup, Claude Code reads the context; on session end, new decisions update the vault. Uses Obsidian backlinks, indexes, project directories, decision logs. Solves the problem where every new window forgets why you made a decision a month ago.
@LufzzLiz [Claude Code]
https://x.com/LufzzLiz/status/2048058284192715168
Official DeepSeek-V4 + Claude Code env vars: ANTHROPIC_BASE_URL points to deepseek API, ANTHROPIC_MODEL=deepseek-v4-pro[1m] (note the [1m] suffix unlocks 1M context — without it you're capped at 200K), CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash, CLAUDE_CODE_EFFORT_LEVEL=max. With the 25%-of-list-price promo through May 5, you get roughly 10-20× the tokens vs Claude API for coding workflows.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2048016721097945386
Built /page-cro skill that audits any marketing page for conversion issues across seven dimensions in order of impact: value proposition, headline, CTAs, visual hierarchy, social proof, objection handling, friction points. Each dimension has specific checks plus quick-wins, high-impact changes, and A/B test ideas. Part of an open-source Marketing Skills collection — 40 skills installable via npx skills add coreyhaines31/marketingskills.
@snakajima [Claude Code]
https://x.com/snakajima/status/2047875453135048907
Three months ago to today: personal Claude Code token consumption is 20× higher. Not a new workflow, not a new project — same engineer, same tasks. The "token explosion" has started. Calling that "100 nuclear power plants of electricity will be burned just for AI" is the sober prediction.
@KingBootoshi-related @fahdananta [Claude Code]
https://x.com/fahdananta/status/2048036866557935738
Saturday routine that runs in Claude Code: scans the past week across feedback, growth areas, focus, time spent, upcoming items, misses. Personal feedback loop on autopilot. The frame is exactly right — once you have a coding agent that can read your data, the boring weekly retro becomes free.
@hiro44_pino [Claude Code]
https://x.com/hiro44_pino/status/2047941177908040120
Two devs, same Mac, same Claude Code. One built a kids' math game and got praised at school. The other built an AI for prediction markets and made roughly ¥1.1 billion in a few months. Same tool, same actions — different target. The framing is uncomfortable: "Claude Code is equal access. Outcomes are not."
@koffuxu [Claude Code]
https://x.com/koffuxu/status/2048057548734165041
Android AOSP source code navigation with DeepSeek V4 + Claude Code: feeding it a precise knowledge base, asking it to find every config touch-point for a specific display panel. Initial benchmark says it's roughly Opus 4.6 quality. Cost per task ~¥1 (after the 25%-off promo) vs ~¥5 list, so ~¥5/task post-promo. Next step: hardware driver development.
@daken_in_market [Claude Code]
https://x.com/daken_in_market/status/2047918013492691236
Earnings season is coming and instead of doing the usual prep work, this trader has been letting Claude Code build the toolchain. The bet: Claude-generated outputs are now good enough to gamble on at earnings. Honest skeptic note in the post itself — "if I can build this, anyone can. Where's the alpha?" Either way it's a real signal that public-info trading workflows are going through a step change.
@xdadevelopers [Claude Code]
https://x.com/xdadevelopers/status/2047828046913114249
Gave Claude Code persistent memory and "now it's unstoppable." The point everyone keeps re-discovering: a coding agent without long-term memory restarts every conversation. Bolt on memory (Obsidian, claude-mem, GitMemo, take your pick) and it stops being a parlor trick. The friction is no longer "can the model code?" — it's "can it remember why we made this decision in February?"
@kerotaro27 [OpenClaw]
https://x.com/kerotaro27/status/2047884763726582214
Got OpenClaw running on a Raspberry Pi 5 as an AI secretary. Honest assessment: still mostly doing what a chat AI could do. The hardware is ready before the use cases are. A reminder that the bottleneck on personal-AI-on-the-edge isn't compute anymore — it's "what do I actually want this thing to do all day."
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2048167256899665986
Built a voice app that uses OpenAI Realtime voice API to talk to the full OpenClaw agent. Posted Friday. Saturday OpenClaw maintainers reposted it. Sunday OpenClaw shipped it natively. The tightest open-source feedback loop you'll see this week.
@VivekIntel [OpenClaw]
https://x.com/VivekIntel/status/2047903395516678502
ClawSec — full security skill suite for OpenClaw, NanoClaw, Hermes. Prompt-injection defense, CVE threat intel, file integrity checks, agent self-audits, real-time advisory feed, checksum verification, drift detection. The agent-security stack that nobody had a month ago is now a packaged skill set.
@PhalaNetwork [Claude Code/OpenClaw]
https://x.com/PhalaNetwork/status/2048047251734368285
Clawdi v2.0 — "iCloud for AI agents." Install once on any device and your OpenClaw, Hermes, Claude Code, Codex agents share the same memory, keys, skills, and files. Runs on Phala's TEE so everything stays encrypted by default. The natural extension of "agents need persistent state" once you stop pretending one machine is enough.
@traversymedia [Claude Code/OpenClaw]
https://x.com/traversymedia/status/2048027509900423260
Has Claude Code, Cowork, and Travis (OpenClaw) all working together, talking to each other, giving each other tasks, using Obsidian as a shared brain. Says he couldn't have imagined this two years ago. The pattern is becoming common: instead of one mega-agent, multiple specialized agents with a shared knowledge layer.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2047940234114220067
Chris Camillo's network: 15 people called him last week ready to quit their jobs after setting up a $650 Mac Mini. The math has shifted. v1 of a product used to take 6-12 months, which meant 12-18 months of savings, which meant only VC-runway people could quit. Claude Code compresses v1 to a weekend. Runway requirement drops from 18 months to 3. The constraint that held for 40 years is gone.
@chongdashu [Claude Code]
https://x.com/chongdashu/status/2047970936003916179
Warp's latest update with first-class Claude Code + Codex integration is the thing that finally unified his game-dev workflow. Vertical tabs for multiple processes, built-in file explorer/editor, native code review panel. Ran a demo with Phaser game clients + Colyseus server alongside both coding agents. The pattern of replacing scattered terminal+editor combos with a single agent-aware shell is showing up everywhere.
@grapeot [Claude Code]
https://x.com/grapeot/status/2047917779932819851
Built a Claude Code skill that lets a PM produce investment-bank-grade Excel models. Then realized: skills can't be sold cleanly. Plain-text files mean the first buyer leaks them. Hosted = you're reselling AWS. API + free skill = the skill is just an SDK for the API. Three failed monetization paths. Argues skills have a "self-destructing" property, and the four real moats in an AI-native business are relationships, the moment, physical side-effects, and judgment/taste.
@VincentBuilds [Claude Code]
https://x.com/VincentBuilds/status/2048101489818677708
"Definitely getting tired of Claude Code — the Anthropic harness keeps making Opus 4.7 do stupid stuff." Counter-vote against the Boris Cherny love-fest. Reading between the lines: Anthropic optimized for pro-tier monetization at the harness layer (rate limits, effort defaults, telemetry) and power users feel it.
@JeremyNguyenPhD [Claude Code]
https://x.com/JeremyNguyenPhD/status/2047856741434745315
Concrete tip: downgrading Claude Code from 2.1.120 to 2.1.71 saves 5× usage on Opus 4.6. Command: npm install -g @anthropic-ai/[email protected]. Whether or not the savings hold for everyone, the fact that experienced users are version-pinning their CLI says a lot about the trust in the latest harness changes.
@AntoineRSX [OpenClaw]
https://x.com/AntoineRSX/status/2047876305908384008
Tested OpenClaw 4.23 with GPT-5.5 inside: built a Pomodoro app in 1 minute, generated a photoreal image (image 2 natively supported), edited its own code mid-build to add CSV export plus streak tracking. The "self-editing during a build" behavior is the part that makes OpenClaw feel different from a coding assistant.
🗣 User Voice
User Voice
The pain stories in this batch cluster around four themes:
Token cost is suddenly real money. @snakajima reports 20× monthly consumption growth in 3 months. @PawelHuryn cut $750 to $100/mo by tuning his harness. @VaibhavSisinty burned 15K extra credits in 12 hours. The "$20 Pro plan" mental model is dead — the question now is whether your usage justifies $1k+/mo. — @snakajima, @PawelHuryn
Cross-session memory is the real moat. @xdadevelopers, @sitinme (Obsidian Mind), @yusaku_0426 (Obsidian schema), @ElaraGrace_AI (GitNexus), all converging on the same point: a coding agent that forgets every Monday is a parlor trick. The community is racing to build the memory layer Anthropic hasn't shipped. — @sitinme
The harness is now the product. @VincentBuilds and @JeremyNguyenPhD both express skepticism that the latest Claude Code harness is actually serving users. @0xSero analyzes Pi vs OpenClaw vs Hermes harness efficiency. The model isn't the bottleneck — it's the wiring around it. — @VincentBuilds
Migration tooling is becoming standard. @fkysly notes Codex now offers one-tap migration of Claude Code settings, skills, and CLAUDE.md → AGENTS.md. @Reboottttttt switched OpenClaw → Hermes via :hermes claw migrate, kept all skills and wallet keys. The competitive terrain has shifted to "how easy is it to leave?" — @fkysly
The pain stories in this batch cluster around four themes:
Token cost is suddenly real money. @snakajima reports 20× monthly consumption growth in 3 months. @PawelHuryn cut $750 to $100/mo by tuning his harness. @VaibhavSisinty burned 15K extra credits in 12 hours. The "$20 Pro plan" mental model is dead — the question now is whether your usage justifies $1k+/mo. — @snakajima, @PawelHuryn
Cross-session memory is the real moat. @xdadevelopers, @sitinme (Obsidian Mind), @yusaku_0426 (Obsidian schema), @ElaraGrace_AI (GitNexus), all converging on the same point: a coding agent that forgets every Monday is a parlor trick. The community is racing to build the memory layer Anthropic hasn't shipped. — @sitinme
The harness is now the product. @VincentBuilds and @JeremyNguyenPhD both express skepticism that the latest Claude Code harness is actually serving users. @0xSero analyzes Pi vs OpenClaw vs Hermes harness efficiency. The model isn't the bottleneck — it's the wiring around it. — @VincentBuilds
Migration tooling is becoming standard. @fkysly notes Codex now offers one-tap migration of Claude Code settings, skills, and CLAUDE.md → AGENTS.md. @Reboottttttt switched OpenClaw → Hermes via :hermes claw migrate, kept all skills and wallet keys. The competitive terrain has shifted to "how easy is it to leave?" — @fkysly
📡 Eco Products Radar
Eco Products Radar
DeepSeek V4 Pro/Flash — 75% off API through May 5. 1M context via the [1m] suffix on the model name. Native integration shipped for Claude Code, OpenCode, OpenClaw. The "Anthropic alternative" story everyone has been waiting for, with concrete config files in circulation.
claude-code-setup (Anthropic official plugin) — Scans your project, recommends hooks/skills/MCP servers/subagents, auto-configures the lot. /plugin install claude-code-setup@claude-plugins-official. The most-recommended single command of the day.
forrestchang/CLAUDE.md (78-82K stars) — One file, four behavior rules derived from Karpathy's critique of how LLMs write code. Drop it into any project and Claude stops over-engineering, stops touching unrequested code, stops assuming. Boris Cherny's workshop is the meta-event of the week.
Boris Cherny "Mastering Claude Code in 30 Minutes" — Workshop by the creator of Claude Code. Got translated into Japanese, Korean, Spanish, Arabic within 24 hours. Showed how Anthropic engineers actually use Claude Code (Plan Mode, /verify, CLAUDE.md, subagents, hooks).
Hermes Agent — Python-based, persistent memory, self-evolving skills. The most-named OpenClaw alternative. Migration via :hermes claw migrate carries skills and wallet keys over.
Vercel npx skills — Cross-tool skill installer for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, 45+ agents. find-skills meta-skill ranks options by install count. The "package manager for agent skills" the ecosystem needed.
claude-mem — Persistent memory across Claude Code sessions. 46K stars in 48 hours.
OpenClaude — Open-source Claude Code clone, 21K stars, MIT license. Same tool loop (bash, file edits, grep, glob, agents, MCP), but provider-agnostic — works with GPT-4o, Gemini, DeepSeek, local Ollama. Free DuckDuckGo web search baked in.
HyperFrames — Motion graphics generation for Claude Design and Claude Code. Skill file + npx hyperframes render produces design-to-animation-to-MP4 in one flow. Got natively integrated into Claude Design.
GitNexus — Knowledge graph engine that maps every dependency, call chain, and execution flow in your repo, then plugs into Claude Code, Cursor, Windsurf via MCP. npx gitnexus analyze. Pre-computes blast radius before any change. 9.4K stars.
Mac mini agent box — 32GB+ models have a 10-18 week wait time as PMs buy them as personal AI compute. Run OpenClaw or Claude Code on it 24/7, message via WhatsApp/Telegram. The hardware shortage is itself a leading indicator.
DeepSeek V4 Pro/Flash — 75% off API through May 5. 1M context via the [1m] suffix on the model name. Native integration shipped for Claude Code, OpenCode, OpenClaw. The "Anthropic alternative" story everyone has been waiting for, with concrete config files in circulation.
claude-code-setup (Anthropic official plugin) — Scans your project, recommends hooks/skills/MCP servers/subagents, auto-configures the lot. /plugin install claude-code-setup@claude-plugins-official. The most-recommended single command of the day.
forrestchang/CLAUDE.md (78-82K stars) — One file, four behavior rules derived from Karpathy's critique of how LLMs write code. Drop it into any project and Claude stops over-engineering, stops touching unrequested code, stops assuming. Boris Cherny's workshop is the meta-event of the week.
Boris Cherny "Mastering Claude Code in 30 Minutes" — Workshop by the creator of Claude Code. Got translated into Japanese, Korean, Spanish, Arabic within 24 hours. Showed how Anthropic engineers actually use Claude Code (Plan Mode, /verify, CLAUDE.md, subagents, hooks).
Hermes Agent — Python-based, persistent memory, self-evolving skills. The most-named OpenClaw alternative. Migration via :hermes claw migrate carries skills and wallet keys over.
Vercel npx skills — Cross-tool skill installer for Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, 45+ agents. find-skills meta-skill ranks options by install count. The "package manager for agent skills" the ecosystem needed.
claude-mem — Persistent memory across Claude Code sessions. 46K stars in 48 hours.
OpenClaude — Open-source Claude Code clone, 21K stars, MIT license. Same tool loop (bash, file edits, grep, glob, agents, MCP), but provider-agnostic — works with GPT-4o, Gemini, DeepSeek, local Ollama. Free DuckDuckGo web search baked in.
HyperFrames — Motion graphics generation for Claude Design and Claude Code. Skill file + npx hyperframes render produces design-to-animation-to-MP4 in one flow. Got natively integrated into Claude Design.
GitNexus — Knowledge graph engine that maps every dependency, call chain, and execution flow in your repo, then plugs into Claude Code, Cursor, Windsurf via MCP. npx gitnexus analyze. Pre-computes blast radius before any change. 9.4K stars.
Mac mini agent box — 32GB+ models have a 10-18 week wait time as PMs buy them as personal AI compute. Run OpenClaw or Claude Code on it 24/7, message via WhatsApp/Telegram. The hardware shortage is itself a leading indicator.
Comments