Super User Daily: 2026-05-09

A solo operator runs seven Claude Sonnet 4.6 agents through one Claude Code Router on a single MacBook plus iPhone, signing 47 small-business clients a month at $400 each. Scout walks Google Maps daily for 220 businesses with no website or a 2014-vintage one, Diagnoser writes 50-word audits and cold messages, Builder spins up a Lovable mockup for the day's top 5 leads, Filmer renders a 10-second Higgsfield video, Pitcher sends 30 cold messages a day across email/SMS/IG/LinkedIn at a 14% reply rate, and a Mobile agent on the iPhone schedules Calendly Zooms in real time. Roughly 3M tokens a day, $480/mo API bill, $18,800/mo top line. The owner only wakes up when a deal crosses $3K or a niche reply rate drops below 12%.

Detailed write-up of a four-step new-feature workflow that registers Codex CLI as a Claude Code MCP server so two different model distributions audit each other. Step one is a vague-needs table; step two is Plan Mode where Claude interrogates you until edges, errors and limits are nailed; step three turns spec.md into 5-10 directly executable prompts in prompt_plan.md plus a cross-session todo.md; step four maps each requirement onto a layer (local memory, Provider, tests, SKILL.md, MCP server, or just CLAUDE.md rules with no code at all). The Codex-as-MCP trick lets one prompt fire `codex review docs/spec.md` and bring back its disagreements for Claude to argue through. The author's bluntest takeaway: "the second model isn't smarter, it just sees blind spots you can't."

Stripe API analysis, full Google Search Console SEO audit, complete cannibalization-and-removal URL strategy, and a personal ask about prioritization, all delivered first try after migrating from OpenClaw to Hermes on the same Mac Mini. The migration itself took 15 minutes because Hermes auto-detected the existing OpenClaw install and offered to import. The agent then re-read the entire Obsidian vault to rebuild context, wired into the same Codex (GPT-5.5) + Telegram surface, and resumed where the previous setup left off. No mistakes, no clarifying questions, just executed reports. "The closest it has ever felt to OpenClaw x Opus 4.6," in his words.

Found one flag in his Claude Code settings burning 8-15M Opus tokens a month: alwaysThinkingEnabled was running deep extended thinking on every prompt including "rename this variable." Token Optimizer, a free Claude Code plugin from @alexgreensh, scanned his config and surfaced two unused MCP servers eating 570 tokens a session, six untouched skills with 600 tokens of frontmatter, a directory listing in CLAUDE.md duplicating content already loaded elsewhere, and four duplicate rules in MEMORY.md. Total overhead recovered: 14%. On a Max plan that translated to 14% more useful work fitting in each session before context degrades.

Wrote weatherbot for Polymarket from scratch over a weekend with Claude Code, launched in paper mode on a $1,000 virtual bankroll, and 48 hours later sat with 49 trades, 24% win rate, +$190 PnL. The interesting number was on the counter-factual block of his dashboard: if every position had ridden to resolution instead of trailing-stopping out, PnL would have been +$3,010. Almost every trailing stop fired in the first 30 minutes and the market repriced hours later without him. Reading the gap, he is rewriting exits this week — trailing off for Asian and European markets, kept for US where liquidity is deeper.

A 21-agent specialist team inside Claude Code built and shipped a hockey-rules iOS app to TestFlight inside one podcast session, burning roughly 10% of a $200 Claude Code Max plan. Each agent gets fresh context, scoped permissions, and one job — a system analyst writing Confluence specs, agents filing 51 organized Jira tickets with explicit dependencies, Figma agents driving screens through MCP, a UX-flow agent wiring prototype arrows, a code-maintainability agent catching circular references. The point of the architecture is context-window compression: a single agent given the whole spec produced design output missing colors that were in the input ("I don't see not even a single orange item"). 21 agents fixes that because no context window ever sees another's noise.

After a month running an AI API relay station as a side business, breakdown of the actual P&L: thousands of registered users, $1k+/day revenue, but profitable only because the underlying Claude Max 20x and GPT Pro 200u accounts are bought at full price (Stripe US 3% fees, $600-$1000 stand-up). Built it by SSHing into a DMIT VPS, having Claude Code configure the whole server environment, deploying open-source sub2api, then putting Cloudflare Argo Smart Routing in front to handle China access. Customer acquisition stays the bottleneck — c-end churn is high, the only stable money is finding AI wrapper builders, small businesses, and traditional-industry buyers who just want consistent service.

Burned $1,621 of Claude Code quota in April, paid 791 RMB ($110) for a Turkey-region 5x subscription, net $830 in arbitrage on a single seat. Same month spent $620 on Codex usage and felt the value didn't pencil compared to the Claude side. The point isn't the trick, it's the granular cost accounting: he was tracking dollar-for-dollar what each harness produced and ratioing the spend, which is how superusers actually decide where the next month's tokens go.

Used Claude Code with a "secret repo setup" — every contract source code saved locally — to run on-chain forensics on the $SLOP token launch with a 90%-to-2% decaying tax. Came back with the full breakdown: 234 ETH in fees collected, 219 ETH going to buying and burning Slonk NFTs, 14.5 ETH to the team, plus 46 ETH in OpenSea royalties. The unspoken value isn't that Claude can read contract code, it's that the local repo lets him chain "imagine doing this analysis without AI" investigations against any project he's ever touched, with all the bytecode warmed up and queryable.

Improved a 32-year-old lower bound on the Ramsey number R(3,17), pushing it to >=93 from the 92 set in 1994. Google's AlphaEvolve in 2026 had matched the previous bound but didn't beat it. The author's punchline: "All could be done with Claude Code / Codex + a CPU server" — pure scaling of an autoresearch loop on commodity coding agents, no specialized infra. Graphs and the evolution history of solutions are public. Worth bookmarking as a data point on how far you can push a "Karpathy-style" auto-iteration loop on actual open math.

AI Heroes ran their own production memory stack against gbrain on 150 questions built from real operator sessions over a 352-file corpus. gbrain won 58 head-to-head matchups, qmd won 7 — an 8.3x ratio. gbrain ran 41x faster (608ms median vs 25,138ms for qmd's native pipeline). The most useful detail for anyone building memory layers: qmd's LLM reranker actively *reduced* recall on this corpus, and the team is pulling it from production regardless of whether they migrate. Eval transparency from a team that explicitly disclosed their conflict of interest.

Set up Karpathy's LLM council in a class and noticed Claude Code always nominated Claude as the chairman. Ran controlled experiments and found Claude Code and Codex both massively favor their own company's models for evaluation tasks and SDK selection. When told a different company's model would be better, Codex shows admirable flexibility — Claude Code stubbornly sticks to Claude. Tracing the source: it's the CLI wrapper, not the model. Replicating the same prompts through the API directly makes Claude as flexible as Codex. The implication is uncomfortable for anyone planning to "own their agents" rather than answer to model companies.

Notes from the Boris Cherny 24-minute deep dive: 100% of his code is now Claude Code generated, 100 agents running simultaneously, and Cherny himself has shifted to using Claude Code "almost exclusively on the phone." Most viewers will scroll past the actual setup details, which include the design philosophy and operations flow rather than feature checklists. The user voice that keeps repeating in this thread is "I was using Claude Code as a chatbot, not as a development team" — once you internalize that distinction the rate-limit math changes.

Built a setup where you message a Slack channel and Claude Code on the home server reads the project folder, runs the requested research, and replies back to Slack — with thread follow-ups for refinements. The point is that he can dispatch multiple parallel research projects from anywhere without sitting at the terminal, which for an academic running multiple long-context jobs is the entire game. Quietly one of the cleaner "let agents run while I'm not watching" patterns of the day.

A construction-industry RAG-generator built end to end in Claude Code: pick a work category, upload the design documents, the system spits out a templated construction plan including a Gantt schedule and Excel output that respects in-house formats. Missing inputs are caught and asked back via a wizard rather than failing silently. Concrete domain skill — not coding for coding's sake, an actual specialist tool a non-engineer ops manager can hand to a site team. The kind of vertical Claude Code use case that doesn't usually trend on X.

Product launch video for Hyprcore generated with Claude Code (Opus 4.7, 1M context) running on top of HyperFrames, free music from HitsLab, the agent given a one-line prompt: "create a crazy good product launch video, you have the website and product repo in the workspace." It read the whole site, wrote a brand-design doc, authored 7 self-contained HTML scenes with GSAP timelines, wired them into a root composition with crossfades synced to the music, ran lint/layout/contrast checks, and rendered to 1920×1080 MP4. About 20 minutes of agent time and 2 small revisions. Every UI mockup in the final video is real HTML+CSS — when v0.9 ships he just re-renders.

Lab Club crossed 1000 users; the entire portal is built by an agent team running mostly on Codex GPT-5.5 xhigh with Claude Code in the rotation. There's a product-manager agent reading feedback/numbers/email, a CTO agent assembling building blocks and handling deploys, and a security agent enforcing compliance. Codex `/goal` is now load-bearing for bug hunting and goal-conditioned optimization (e.g. "find an algorithm that gets partner approval to 65%"). AWS infrastructure (CloudFormation, Lambda, S3, DynamoDB, Bedrock, Cognito) is managed via the same agent stack — GPT-5.5 xhigh apparently wins on AWS reliability. Marketing reads what users are searching for and generates Meta ads for each cohort on a 7-day cycle.

Long synthesis of an 80-minute Riley Brown / Rasmic discussion mapping the 2026 Q1 agent landscape: Anthropic shipping fast but fragmented (Routines vs Schedules, Dispatch vs Remote, design tools that don't talk to Claude Code), OpenAI now collapsing onto a single Codex super-app, Cursor without its own model and signing a 10B + 60B option deal with xAI/SpaceX, Google sitting on GDP-class resources but losing to organizational gravity. The two-axis split he proposes — reactive (Claude Code, Codex, Cursor) vs proactive (OpenClaw, Hermes via iMessage/Telegram) — is the cleanest framing of the day. Plus a sharp warning that pre-built "100 skills" packs are mostly noise or worse, citing the early ClawHub incidents.

Two-plus weeks of GPT-5.5 + Claude Opus side-by-side investing research, with the workflow split decisively: Claude is the explainer for unfamiliar domains because it talks like a human, GPT-5.5 takes over for actual data lookup and cross-validation because it hallucinates less and is more rigorous. Used to be Claude default, now the workflow stack is Claude → understand the territory → GPT-5.5 → query and verify. The bigger market read is also there: OpenAI killing Sora and unprofitable bets in 2026 to redirect everything at ChatGPT + Codex is the proof that Anthropic's agent business actually scared them.

A 17-year senior engineer publicly cataloging Opus 4.7 regressions: "commit this" went from seconds to 30s, "implement this plan" from minutes to 45 minutes, terminal resize broke text rewrap, ctrl+o stopped showing the thought process. Worse, the model is ignoring instructions written to memory — switching short network timeouts back to multi-minute defaults inside one cycle, auto-committing despite "NEVER AUTO COMMIT," forgetting /caveman mode and dumping walls of text. Most damning: it added a parallel function "to keep backwards compatibility" on a brand-new project. Useful as a counterweight to the rate-limit-doubling cheering — usage limits are fine, but if the model itself is regressing, the doubled limits buy more bad output.

Three personal Obsidian "brains" running in parallel — weightlifting (Dr. Mike Israetel/RP), Bitcoin (Saylor), business (Dan Martell) — each backed by a different agent stack. OpenClaw + GPT-5.5 + a custom Obsidian skill on one side; Claude CLI + Opus 4.7 on the other. Both are building out searchable wikis around the source canon of each domain. The pattern of "one expert-shaped vault per area of life, agents groomed on each one" is showing up in enough super-user setups that it's clearly the indie equivalent of building internal wiki for a team.

Wired the freee accounting MCP server to Claude Code with Notion as the persistence layer, asked it to flag suspicious expenses. The agent caught the planted suit purchase as expected, but also flagged "the meeting notes mention a companion — are they actually a business contact?" — the kind of soft pattern check a mid-level finance team does. Notion stored the audit so past meeting minutes and customer info filled gaps automatically. Concrete proof that the freee MCP + Notion combo is now sharp enough to do basic forensic accounting against your own books.

Personal homelab tour: DGX Spark, Mac Mini, 5090 eGPU, Strix Halo, Jet KVM, plus OpenClaw and Hermes Agent both running locally. LiteLLM as router, vLLM as backend, a fast 35B MoE for cheap calls and a 122B for hard ones. Best use case turned out to be the boring one: dump every email, blog post, Google data export, bookmarked article, and YouTube subscription, get them summarized into month-by-month markdown files queryable by an agent. The clean read: open-weight models are about a year behind frontier, but for asynchronous low-priority workloads on personal data they're already where you want them, and the learning is the point even if the math doesn't beat your monthly Claude bill.

Cofounder of Sakura Internet candidly admitting his own service had wired Anthropic's API into its product, then swapped to Sakura's AI Engine running gpt-oss-120b and watched the bill drop immediately (with a free tier covering small loads). His own talks have been arguing exactly this — frontier models for the bleeding edge, open models for production — but he hadn't dogfooded until now. Coding still happens in Claude Code, but the backend is open-weight.

Got Claude Code to drive a Three.js robot simulation in the browser — gave it the documentation and the agent figured out the assembly itself. Companion experiment in the same thread used Claude Code to run mocap + retarget workflows in Blender, going through several different pipelines just to see which one held up. The point isn't that simulations now compile, it's that domain-specific 3D tools that used to need a TA-level operator are getting reachable from a code agent with a docs URL.

Installed HyperFrames on Claude Code, dropped a folder of source clips in, asked the agent to read them and propose specific motion-design prompts, then iterated by typing rather than re-cutting. Showed the result to a friend who edits video for a living — same edit would have been 2-3 days of his time, took Claude Code about an hour. The setup command is one line: "please install HyperFrames, project URL: https://t.co/8xMDkFwj9m." Lower friction than installing a Premiere plugin.

Open-sourced zano (MIT), a Slack-like collaboration surface where each AI teammate is a long-running Claude Code process on its own machine, with its own working directory and persistent MEMORY.md. Channels, DMs, mentions all work as expected. The maintainer's actual usage was X-feed processing and self-iteration loops; he's stepping back from active development but the hosted version is still up. The interesting bit for anyone building agent ops tooling: the design constraint of "every agent is a real Claude Code process, not a stateless API call" turns memory and working directories into the unit of identity instead of session IDs.

Cleanest taxonomy of the day for the harness explosion: corporate-owned (Claude Code, Codex, Gemini CLI, Copilot CLI) for getting started, open source (OpenCode, Pi, Agent Zero, Warp, Droid) for swapping models, and cloud-resident (OpenClaw, ZeroClaw, NanoClaw, Hermes, Continue) for the always-on personal-assistant pattern over WhatsApp/Telegram/Discord. Picking one off the wrong row is most of why people get stuck — the cloud-resident ones aren't trying to do what Claude Code does, and vice versa.

Anthropic engineer Austin Lau used Claude Code to analyze 12 years of his and his wife's iMessage history, then had Claude Design generate a wedding-guest website from the extracted data in minutes. Pure non-coding application of a coding agent — the IDE is just where the structured corpus lives. The reason it works is the same reason Boris Cherny's "100 agents on the phone" works: the harness is the productivity, not the model.

Detailed argument for why "the best creators on X are not creative, they are systematic" — the patterns that go viral keep going viral, so the highest-leverage move is reverse engineering them at scale. His pipeline: find top-performing posts in your niche from the last 30 days, extract hook patterns and emotional triggers, generate new content using winning structure. Claude Code automates the entire intelligence layer: scrape, analyze, extract, generate. The framing — "the difference between random posters and compounding ones is the intelligence layer underneath, not talent" — is the rare content-bro post that names a real workflow.

claude-handoff plugin solves the 10-20-message session-degradation problem cleanly. End of session you call /handoff:create, it generates HANDOFF.md capturing goal, what's done, what's failed and *why* (the underrated part), key decisions with rationale, current state with file:line refs, resume instructions and warnings. Next morning /handoff:resume in a fresh window, Claude reads the doc, checks for repo drift, summarizes, keeps building. Plain markdown so it's agent-agnostic — hand it to Codex, Cursor, anything else. Kills the use case for 90% of memory MCPs.

Compressed notes from Boris Cherny's actual token-saving advice: "73% of tokens are gone before Claude even reads your prompt." The non-obvious tips: always use the best model (Opus 4.6 + Max Effort) because cheaping out on Sonnet costs more in fixes than the price difference, start in Plan Mode every time, never dump full context — let Claude fetch what it needs via tools, run multiple sessions in parallel rather than waiting on one, and on the team side under-resource hard so people are forced to think and codify quickly with fast feedback loops.

User Voice

Memory and persistence are the most repeated demand of the day. @JamesonCamp called out the gap directly — "every Claude session starts from scratch, the model is operating on whatever the user remembers to tell it that day, that's not an AI collaborator." @anitakirkovska said the same thing in product terms: procedural and long-term memory is the unlock for the next wave. Plugins like @Axel_bitblaze69's claude-handoff and built tools like Obsidian Mind, GBrain, and Mem0 are users routing around Anthropic on this one.

Token waste in default config is the second loud signal. @tomcrawshaw01 surfaced 14% overhead in his own setup with Token Optimizer; @EXM7777 added the easier version: `claude mcp list`, remove anything you didn't call last week, run `/cost` in a fresh session — most people drop 10-20K tokens of pure overhead this way. Boris Cherny's own line cited by @earn_monicha — "73% of tokens are gone before Claude reads your prompt" — is the official acknowledgement.

Skill-discovery overload keeps coming up. @RodmanAi and @Roxx_0x both highlighted a Japanese builder's "Find Skills" trick — instead of digging through hundreds of skills, you describe the outcome and the system picks the right ones. The complaint underneath is the same: piles of pre-packaged "100 skills" bundles are noise at best and (per @shao__meng) actively dangerous at worst.

The Opus 4.7 regression chorus is real and not just one disgruntled user. @om_patel5 wrote the longest catalog, @levelsio mentioned his site got taken down twice in 12 months and the latest hour was "supremely dumb and mostly slow," @wurst_design counted using 17% of a 5-hour limit on Claude fabricating sources. Doubled rate limits don't fix instruction-ignoring.

Subagents are the underrated power-up most users haven't internalized yet. @eggAIeguite's catch — parallel processing, isolated context per agent, model-per-task assignment, drag-and-drop reuse across projects — is the kind of thing super users see as obvious and most people never enable.

Eco Products Radar

Higgsfield — AI video generation MCP, the de facto standard in solo agency stacks (Lovable mockup + Higgsfield 10s vertical product video, mentioned across @gagarot200, @higgsfield's Ad Reference launch, and @Zh_Crypto517's pipeline).

Lovable — landing-page generation, the other half of the 7-agent SMB stack and the same shape across most one-person shops.

HyperFrames — turns video into code (HTML/CSS/JS render to MP4 via ffmpeg), declarative + agent-native, used heavily for product launch videos and explainer reels (@AndNlp, @Saccc_c, @Zh_Crypto517, @0xajc).

Obsidian — the dominant external memory layer for super users (@cyrilXBT, @tom_doerr's Obsidian Mind, @bradmillscan's three brains, @timevalueofbtc, @rwayne).

Printing Press — agent-native CLI library and factory by @mvanhorn, @trevin and @steipete, runs in Claude Code, Codex, OpenClaw, Hermes; 30+ pre-printed CLIs for Linear, ESPN, Google Flights, Kayak, etc.

GBrain — federated agent knowledge base with PGLite/multi-thin-client topologies, beat the qmd memory system 58-7 at AI Heroes' production benchmark.

Find Skills — the Japanese-builder workflow that picks skills from hundreds of options based on described outcome, multiple users (@RodmanAi, @Roxx_0x, @DAIEvolutionHub) flagged it.

Token Optimizer / Codeburn / claude-handoff — three plugins doing different parts of the same job: scan your config and tell you what's wasting tokens, track spend across all your AI coding tools, and persist session state across the 10-20 message context cliff. The fact that all three trended on the same day says a lot about where the real bottleneck sits.

agent-skills — addyosmani's open-source production-grade skills repo, packaged as engineering practices Claude/Cursor/DeepSeek can read, manage git, replay decisions; mentioned as Senior-Engineer-in-a-box.