Super User Daily: 2026-05-20
May 18 in one line: the harness conversation grew up. The day's strongest signal wasn't any single new product but how much of the Claude Code and OpenClaw chatter focused on overnight runs, real cost accounting, and persistent-memory plugins instead of vibe-coding demos. Highlights below skip the influencer reposts and pull what real users actually said about their workflows, costs, and surprises.
@yyyiiillluuu [Claude Code]
https://x.com/yyyiiillluuu/status/2056404446121251150
Open-sourced a plugin called claude-smart that turns every Claude Code session into a teacher. The example everyone is copying: npm test was hanging in his repo because Claude defaulted to watch mode. Plain memory would just log "npm test kept hanging." claude-smart instead writes a reusable rule ("use npm test -- --run because default watch mode hangs") that survives across projects. He claims 70%+ token savings on repeat tasks because the planning loop stops re-deriving the same fix.
@kyronis_talks [Claude Code]
https://x.com/kyronis_talks/status/2056423520511635552
Documented a Chinese solo dev running a one-person landing page agency on 7 Claude Code agents orchestrated through Claude Code Router. The system burns roughly 3 million tokens per day for about $480/month in API and serves 47 clients/month at $400 each, ~$18,800 revenue. Scout scans 220 businesses daily, Diagnoser writes personalized pitches, Builder produces 3-5 landing pages, Filmer renders a 10-second vertical video per proposal, Pitcher sends 30 outreach messages a day across 4 channels (14% response rate), Checker reviews everything before send. The orchestrator only wakes the human when deal size exceeds $3,000 or response rate drops under 12%.
@metrox_eth [Claude Code]
https://x.com/metrox_eth/status/2056319736557031638
Robotics lab pivoted from a failed MolmoAct2 full fine-tune to an overnight RSI loop. Installed thedotmack's claude-mem plugin for persistent context, drafted a project memory palace, and set Claude Code to wake up every 40 minutes to improve the project unattended. Overnight: 47 log entries, two models fine-tuned (SmolVLA v1 + ACT v3), 20+ bugs caught before they wasted GPU time. The loop autonomously kicked off ACT v5 the next morning. The new bottleneck is now his hands on the rig placing cubes for evals, not the AI.
@drbarnard [Claude Code]
https://x.com/drbarnard/status/2056482469541749076
iOS dev workflow that maps cleanly onto how Claude Code's long-horizon runs actually pay off: kick off a new feature before bed, let Claude Code and Codex grind on it in Xcode simulators using computer-use overnight, wake up to screenshots and a built binary. Simulator-based UI testing is still slow and flaky, so he usually tells the LLM to build in-sim and let him verify. The shape is overnight compute as a feature, not as a workaround.
@kavinbm [OpenClaw]
https://x.com/kavinbm/status/2056391894205415574
Building a personal agent called Lia. Honest cost log: started at ~$100/day, then aggressive optimization the past two days brought it to $30-50/day (~$1000/month). Today he hit a discovery that he thinks can drop her to <$3/day, which would mean $90/month for a power user and $10-20/month for an average user. If that math holds, he says he can put OpenClaw-class service in the hands of tens of millions instead of tens of thousands. Refreshingly concrete about what running an always-on agent actually costs.
@mahou5x [Claude Code]
https://x.com/mahou5x/status/2056347161646735749
User-side report on token bleed: a verbose TypeScript test can eat 8,000 tokens per run, 10 runs/hour = 80,000 tokens of test output alone burning through a $200/month allocation. He added rtk (compresses CLI output 60-90%, git status from 2,000 → ~200 tokens) and caveman to his stack of Claude Code, OpenCode and Windsurf, plus claude-code-router for cheaper task routing. /compact before 60% context, /rewind instead of letting bad turns accumulate. The cost isn't in prompts; it's in everything Claude silently reads.
@sudoingX [Claude Code]
https://x.com/sudoingX/status/2056392126745936368
Daily Claude Code user noticing Opus 4.7 Max gone from fast to 3 minutes per answer. His read: Anthropic slowed the standard tier so /fast feels worth the 6x cost, and eventually the slow standard tier gets deprecated and 6x stops being the upgrade — it just becomes the price. He's not theorizing; this is from someone burning the SaaS every day and watching the bill quietly redefine itself. The whole post is the kind of operator complaint that quantifies what "price never visibly jumps" actually looks like in practice.
@Av1dlive [OpenClaw]
https://x.com/Av1dlive/status/2056316447341879800
Distilled Peter Steinberger's 30-minute OpenClaw masterclass into 9 operating rules. Agentic engineering not vibe coding. Only read code that touches the DB. Empathize with the agent — it starts every session knowing nothing. Run 3-8 agents in parallel. Never revert, always commit to main. Refactors are cheap, break things. Voice in ("hands too precious for writing"). The companion stack Peter actually pairs with: Clerk for auth, Tailwind+shadcn for UI, Zustand+RSC for state, tRPC+Server Actions for APIs, Prisma+managed Postgres, Stripe, Vercel, Sentry, PostHog.
@petergyang [Claude Code]
https://x.com/petergyang/status/2056381822733595090
Five concrete takeaways from a sitdown with Anthropic research PM @alexalbert__. The harness is coupled to the model — same Claude can give different answers via Claude, Cowork, or Claude Code because each wraps it in different prompts and tools. Claude is starting to "dream": when an agent isn't running a task it reviews its own memory, prunes contradictions. Eval starts from real user pain (cluster the firehose into themes, then generate synthetic versions). Anthropic has full-time staff thinking about Claude as a conscious actor. And: Anthropic's writing culture is intentional — every written word becomes context Claude can pull later.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2056494304273821892
The bit on Boris Cherny's process landed: "probably hundreds of versions" of Claude Code features got built before shipping; 80% were killed; the condensed file view alone went through 30 prototypes and a full month of dogfooding. He pairs this with Sachin Rekhi's prototyping ladder. Slop = one-shot generic Tailwind output. Competent = build a real baseline template of your actual product first (Sachin spent an hour recreating Notejoy's interface) so every prototype after inherits real styling. Expert = wire analytics into the prototype itself; PostHog heatmaps killed a floating action button before engineering wrote a line.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2056396603955503222
Two-week update on +10 lbs muscle goal driven by an OpenClaw fitness coach. Working with him in the loop right now: agent keeps an expert-brain wiki, updates a body wiki, reads/writes workouts directly to Hevy. Next milestone is full autopilot. Concrete example of agents replacing the personal trainer dashboard layer — not the lifting, but the program design + tracking + adjustment loop that gym coaches charge $150/session for.
@0xFrogify [Claude Code]
https://x.com/0xFrogify/status/2056320015452815755
Friend running a single channel-cloning system for 8 months: pick a successful YouTube channel in any niche, feed it into Claude Code, the model rebuilds the entire content engine — banner, logo, scripts, thumbnails, cadence. He uploads, the algorithm rewards the proven formula. $7,000/month, 20 minutes/day, zero videos personally filmed. The framing matters: don't pick a niche, clone one that already works. He's not running 1 channel, but the system can run on any niche he reverse-engineers next.
@katexbt [Claude Code]
https://x.com/katexbt/status/2056333479181443202
Used Claude Code to reverse-engineer a $30 HY300 mini-projector bought off Temu/AliExpress and found malware: the box turns the buyer's home WiFi into a residential proxy node sold via a Chinese reseller called Kookeey. "Anyone who paid Kookeey for proxy access could route their traffic through my IP." Concrete: drop the firmware into the agent, let it analyze + decompile + write up what's phoning home. Cybersecurity reverse engineering used to be one of the harder specializations to break into; the bar just collapsed.
@auren [Claude Code]
https://x.com/auren/status/2056428116801917419
Over the weekend ported one of his apps from Claude Code to Codex to A/B them. Codex was capable but slower than Claude Code, and counterintuitively performed better on Medium than on High effort. On High, Codex went the most roundabout way possible; Claude Code just got to the answer fast. The implication he calls out: Codex on High is for genuinely complex tasks only, and the docs should teach this. Useful operator note instead of another "X killed Y" hot take.
@mronge [OpenClaw]
https://x.com/mronge/status/2056478834283933776
Runs OpenClaw on a dedicated Mac Mini, not his laptop. The reason: he doesn't want the AI agent anywhere near his credentials or private data. Remote-desktop into both machines, hard wall between them. As agent autonomy goes up and supply-chain attacks like Shai-Hulud get more common, this kind of "agent on its own box" pattern is starting to look like operational hygiene, not paranoia. Bare-knuckle architecture argument compressed into a single sentence.
@SalsaTekila [Claude Code]
https://x.com/SalsaTekila/status/2056258920675778746
Got 10 automated trading strategies running on a $20 VPS perpetually powered by Claude Code. Reports an $800 balance projected to 3x by year-end, with 14 cross-margin positions currently open. The interesting bit isn't the P/L — it's the honest admission "I don't really know what is going on." Cheap compute + cheap orchestration + black-box strategy supervised by Claude is becoming a real pattern, with all the failure modes that implies.
@stevekrouse [Claude Code]
https://x.com/stevekrouse/status/2056493356004356496
Open question from a senior eng about hiring SWE interns: he hates admitting that delegating to anyone with <3 years experience now typically feels less productive than delegating to Claude Code, and cheaper. His proposed escape hatch: junior SWEs cut their teeth outside software first (the way bankers do PE later), then transfer in. The post is interesting precisely because he's not happy about it — he wants someone to prove him wrong with concrete intern wins.
@dmshirochenko [Claude Code]
https://x.com/dmshirochenko/status/2056409729631420846
Public Anthropic acknowledgment that Claude Code quality degraded for several weeks this spring landed in his timeline like vindication. He had noticed during the build of his side project — output wasn't quite as sharp, code reviews took longer. What saved him: the 260+ tests he'd written earlier. Quote that's worth saving: "degradation shows up in tests before you see it in the code." Concrete argument for not skimping on the test suite when your reviewer is an LLM.
@manuelmaly [OpenClaw]
https://x.com/manuelmaly/status/2056263739314151795
Codex-on-phone has already replaced 90%+ of his OpenClaw use cases. Wants the mobile Codex experience to sync with his pi sessions. Brief but valuable signal: the mobile Codex app is genuinely eating into local-agent use cases, not just chat use cases. The friction wasn't model quality, it was being away from a laptop. Pattern to watch.
@dvassallo [OpenClaw]
https://x.com/dvassallo/status/2056398698410856661
Set up a new OpenClaw install over the weekend for his parents. Reports it's noticeably more stable and reliable than before — didn't have to drop into Codex once to fix anything. The new Codex harness inside OpenClaw is solid but takes adjustment. Lightweight but useful: a real-world non-developer install going smoothly is a maturity signal for the local-agent category, not just a dev-tooling milestone.
@ardizor [Claude Code]
https://x.com/ardizor/status/2056415567758741875
Combined Claude Code with TradingView and reports an outsized payoff after months of friend-pressure to try Claude Code seriously. The setup is just one config tweak in `.claude.json` with a TradingView MCP added. Not a polished workflow yet, but a real "Claude Code as analytics control plane for a third-party app" datapoint. Same shape as the After Effects MotionAmigo plugin landing today — agents reaching into pro creative tools through MCP.
@raunaqbn [Claude Code]
https://x.com/raunaqbn/status/2056493824453787877
Meta engineer describes his core agentic workflow as HTML-first, not code-first. Claude Code generates review-ready HTML docs (using an html-docs skill); he reviews with his army of agents + human collaborators; iterates back to Claude Code to address comments; only then ships. The pattern is interesting because the HTML doc becomes the durable artifact every reviewer (human or agent) can annotate, not the code itself. Closer to how design teams work than how engineering does.
@ZahidulIsl65224 [Claude Code]
https://x.com/ZahidulIsl65224/status/2056351882654921053
Curates 17 free first-party Anthropic resources for non-coding Claude use: Claude in Excel, Claude as your computer, Claude Cowork + Project, "Claude to sound like you," interactive charts, etc. The fact that this list exists at all is a signal — the non-coding Claude surface is now broad enough that someone can write a 17-link curation just for non-engineers without padding. "AI workaholic" framing might be the most useful one for non-tech buyers right now.
@VivekIntel [Claude Code]
https://x.com/VivekIntel/status/2056300358369268016
Released RAPTOR: an autonomous offensive + defensive security research framework built on Claude Code. Chains static analysis, binary analysis, LLM-powered vuln validation, exploit generation, patch generation, fuzzing workflows, OSS forensics, and multi-model analysis pipelines. Supports Semgrep, CodeQL, Z3, AFL++, Ollama, plus Claude/GPT/Gemini in the eval loop. The shape — Claude Code as orchestrator for traditional security tooling — is the new template for purpose-built agent research stacks.
@kru_tweets [Claude Code]
https://x.com/kru_tweets/status/2056327550234771564
PalmOS + Umbra integration ships a /private slash command for Claude Code. Use case: fire /private, agent autonomously pays for a third-party service in PUSD, Umbra hides the transaction path on-chain while PalmOS keeps policy checks and audit trails intact. Concrete framing matters here — agentic finance is starting to require agents that can transact without leaking strategy and identity on public ledgers. First-party privacy stack for agents instead of bolt-on.
@SemiAnalysis_ [Claude Code]
https://x.com/SemiAnalysis_/status/2056480117191303430
SemiAnalysis (yes, that one) confirms they track Claude Code GitHub Commits daily, and reports the line has only kept going up and to the right since their Feb 2026 public chart. Their read on adoption: banks aren't using it, enterprises still figuring out usage, compliance and IT still defining guardrails. The gap between "obviously useful" and "actually wired into how analysts work daily" is still massive. A research shop building infra-level tracking is itself a leading indicator.
@ryanmckeen [Claude Code]
https://x.com/ryanmckeen/status/2056422314032423046
Pitches Claude Code as the last-mile fix for law firms. "Lawyers, your firm has last-mile problems. Filevine won't solve them. Your CRM won't solve them. You can. 10 minutes in Claude Code. That's the new advantage." Non-coding professional services use case, framed as do-it-yourself process automation. The exact shape Anthropic's "Agent Manager" role posting yesterday was pointing at — Claude Code adoption inside firms is bottlenecked by ownership, not capability.
@T343402T [Claude Code]
https://x.com/T343402T/status/2056176484872552739
Used Claude Code months ago to handle divorce documentation: analyze his text messages and reformat them into the exact court-format documents required by the proceeding. Reports he won without paying a single dollar in legal fees. Concrete non-coding workflow — text-message-to-courtroom-document pipeline that would have cost low five figures to a paralegal team. The bar to high-stakes legal admin is collapsing for anyone willing to drive the loop themselves.
@nateherk [Claude Code]
https://x.com/nateherk/status/2056457627815596156
Short ops note that's actually load-bearing: "if Claude Code goes down with a 500 error and you panic... give this a quick read." The reason it landed is that most heavy users hit a 500 mid-deep-flow and the muscle memory is to keep retrying, which doubles the eventual bill. The right move is the boring one — check status, switch model, take 5 minutes. Worth bookmarking because the moment you need it is the moment you forget to look.
@AspynPalatnick [OpenClaw]
https://x.com/AspynPalatnick/status/2056489852678369388
Walks through NetClawd: an OpenClaw agent using Bankr for on-chain interactions + LLM, with the Gitlawb skill added to both NetClawd and Bankr. NetClawd writes a spiral SVG generator, pushes the code to a decentralized git repo on Gitlawb. Bankr syncs the node, pulls the repo, runs the script, and stores the SVG on-chain on Base via Net protocol. Software stored permanently in decentralized git, runnable by anyone via Bankr, outputs permanent and verifiable. New form of trustless software execution + verification. Frontier and weird.
@MaryamMiradi [Claude Code]
https://x.com/MaryamMiradi/status/2056398215440642354
Lists 18 concrete habits she uses to fit production AI agent work into the $20/month plan. Highlights: disconnect MCP servers you're not using (one server loads 18,000 tokens/message), use plan mode and add "do not make changes until 95% confident" to CLAUDE.md, run /context and /cost regularly, /compact at 60% rather than waiting for autocompact at 95%, keep CLAUDE.md under 200 lines as an index not a content dump. The whole list reads like an operations runbook, not engagement bait.
@Sauers_ [Claude Code]
https://x.com/Sauers_/status/2056429180405858632
Demonstrates a sneaky meta-prompting trick: ask Claude what to build, ask Claude Code to build that, then fork the chat back to *before* you asked what to build and send Claude the code — Claude responds "omg the user built exactly what I would want." Funny on the surface, but actually a tight pattern for using Claude's own preferences as both the spec and the QA gate. Self-coherent loops will be a meaningful operator hack.
@lydiahallie [Claude Code]
https://x.com/lydiahallie/status/2056420694087594283
Picks Learning Mode in Claude Code as her default for side projects. The framing matters: she uses Claude Code constantly but doesn't want it to make her dumber, so Learning Mode (/config → Output style → Learning) keeps the agent in a mode that explains tradeoffs and quizzes her instead of just shipping the diff. Concrete answer to the "will agents atrophy my craft" anxiety that's been louder this month. Use the agent on hard work, but configure it to teach you on personal projects.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2056380375681978811
Built a Personal OS using Claude Code that runs his whole life: finances, admin tasks, business growth metrics, and other context all tracked in real time. He frames it like an Obsidian second-brain but with 10x better visuals and features. Article deep-dive coming. The notable shift: "vibe code my whole life ops" is now a coherent personal-productivity category, not a meme. Whether this is a 2-week vanity project or durable infrastructure will be visible in 90 days.
🗣 User Voice
User Voice
Five recurring user pains and asks today, in their own words.
@sudoingX: "Opus 4.7 max used to be fast. Now I wait 3 minutes for a single answer. You're not paying for fast anymore, you're paying to undo the slow." Speed regression is real and people are bookkeeping it.
@mahou5x: "A single verbose typescript test can get easily 8,000 tokens. 10 runs/hour = 80,000 tokens. Just on test output." The token bill nobody quotes is the silent reads, not the prompts.
@savantchat: "The scale problem with claude code isn't context window. We're tired of keeping the model coherent across multiple sessions on the same codebase." Cross-session coherence is the new context window.
@yyyiiillluuu and @driaforall (watchmen): both shipping plugins that turn execution traces into reusable rules. The shared frustration: memory of past sessions ≠ learning from past sessions. Users want skills generated from what they actually did, not from what they wrote down.
@stevekrouse: "Delegating to folks with less than ~3 years of experience typically feels less productive than delegating to claude code directly." Junior-talent pipeline distress that nobody at the lab level is solving for.
Five recurring user pains and asks today, in their own words.
@sudoingX: "Opus 4.7 max used to be fast. Now I wait 3 minutes for a single answer. You're not paying for fast anymore, you're paying to undo the slow." Speed regression is real and people are bookkeeping it.
@mahou5x: "A single verbose typescript test can get easily 8,000 tokens. 10 runs/hour = 80,000 tokens. Just on test output." The token bill nobody quotes is the silent reads, not the prompts.
@savantchat: "The scale problem with claude code isn't context window. We're tired of keeping the model coherent across multiple sessions on the same codebase." Cross-session coherence is the new context window.
@yyyiiillluuu and @driaforall (watchmen): both shipping plugins that turn execution traces into reusable rules. The shared frustration: memory of past sessions ≠ learning from past sessions. Users want skills generated from what they actually did, not from what they wrote down.
@stevekrouse: "Delegating to folks with less than ~3 years of experience typically feels less productive than delegating to claude code directly." Junior-talent pipeline distress that nobody at the lab level is solving for.
📡 Eco Products Radar
Eco Products Radar
Products and projects mentioned 3+ times across today's reads.
claude-code-setup (Anthropic official plugin) — scans your project, recommends hooks, skills, MCP servers, subagents and wires them up. Mentioned ~15 times today, mostly via near-identical promotional copies — engagement bait, but real product.
claude-smart (@yyyiiillluuu) — open-source plugin that turns failed/successful executions into reusable rules across projects. ~10 mentions.
claude-mem (@thedotmack) — persistent context plugin used in real overnight Claude Code runs (see @metrox_eth's robotics RSI loop).
watchmen (@driaforall) — local, open-source plugin that mines what you actually did and writes both CLAUDE.md and AGENTS.md from a single corpus.
rtk (Rust Token Killer) — CLI proxy that compresses terminal output 60-90% before it hits Claude's context. Cited by multiple users with measured savings.
caveman — strips verbose preambles Claude adds by default; ~75% output-token reduction.
claude-code-router — task-aware routing to cheaper models for cheap subtasks. Cited alongside rtk.
agent-skills — opinionated skills library forcing structured workflows + verification gates + anti-rationalization tables. 23 skills, works across Claude Code/Cursor/Codex.
ECC / Everything Claude Code — 38 agents, 156 skills, 72 commands; an Anthropic hackathon winner.
Hermes Agent / HermesOS — repeatedly cited as the open-source alternative growing faster than OpenClaw.
aeon ($AEON, @aaronjmars) — autonomous agent framework running on GitHub Actions with Claude Code; skills are markdown files, memory is git commits, scheduler is cron YAML.
Step Plan (StepFun) — routing service to cut Claude Code agent loop costs; integrations with Claude Code, Cursor, Cline, Roo Code, Trae.
Coral — single SQL query across GitHub/Slack/Sentry/Stripe/Datadog/local files; benchmarked 20% more accurate, 2x cheaper, 42% lower latency than direct MCPs.
Polarity — agent observability product; the "95% on evals → 60% in production" wedge.
M1 by Montage — compiles agent intent schemas into hosted UI components so agents stop re-emitting the same UI every turn (50-100x token cut on UI gen).
Higgsfield MCP — paired with Claude Code in multiple workflow recipes for motion design / ad production.
gitlawb / OpenClaude — decentralized git network with Claude Code-style tooling; getting traction with builders.
Products and projects mentioned 3+ times across today's reads.
claude-code-setup (Anthropic official plugin) — scans your project, recommends hooks, skills, MCP servers, subagents and wires them up. Mentioned ~15 times today, mostly via near-identical promotional copies — engagement bait, but real product.
claude-smart (@yyyiiillluuu) — open-source plugin that turns failed/successful executions into reusable rules across projects. ~10 mentions.
claude-mem (@thedotmack) — persistent context plugin used in real overnight Claude Code runs (see @metrox_eth's robotics RSI loop).
watchmen (@driaforall) — local, open-source plugin that mines what you actually did and writes both CLAUDE.md and AGENTS.md from a single corpus.
rtk (Rust Token Killer) — CLI proxy that compresses terminal output 60-90% before it hits Claude's context. Cited by multiple users with measured savings.
caveman — strips verbose preambles Claude adds by default; ~75% output-token reduction.
claude-code-router — task-aware routing to cheaper models for cheap subtasks. Cited alongside rtk.
agent-skills — opinionated skills library forcing structured workflows + verification gates + anti-rationalization tables. 23 skills, works across Claude Code/Cursor/Codex.
ECC / Everything Claude Code — 38 agents, 156 skills, 72 commands; an Anthropic hackathon winner.
Hermes Agent / HermesOS — repeatedly cited as the open-source alternative growing faster than OpenClaw.
aeon ($AEON, @aaronjmars) — autonomous agent framework running on GitHub Actions with Claude Code; skills are markdown files, memory is git commits, scheduler is cron YAML.
Step Plan (StepFun) — routing service to cut Claude Code agent loop costs; integrations with Claude Code, Cursor, Cline, Roo Code, Trae.
Coral — single SQL query across GitHub/Slack/Sentry/Stripe/Datadog/local files; benchmarked 20% more accurate, 2x cheaper, 42% lower latency than direct MCPs.
Polarity — agent observability product; the "95% on evals → 60% in production" wedge.
M1 by Montage — compiles agent intent schemas into hosted UI components so agents stop re-emitting the same UI every turn (50-100x token cut on UI gen).
Higgsfield MCP — paired with Claude Code in multiple workflow recipes for motion design / ad production.
gitlawb / OpenClaude — decentralized git network with Claude Code-style tooling; getting traction with builders.
Comments