Super User Daily: 2026-06-04
The timeline was loud this week, but most of the noise was IPO math and Microsoft-on-stage theater. Underneath it, the real signal is that people have stopped using these agents as fancy autocomplete and started wiring them into things that run on their own. The pattern that keeps showing up: someone hands the agent a goal, a few guardrails, and a loop, then walks away. Trading desks, research pipelines, voice models, ad factories, even an LLM running on a second-hand FPGA. The interesting part isn't that the code gets written. It's how much of the surrounding system people are now willing to let the agent own.
@InvestwithMEH [Claude Code]
https://x.com/InvestwithMEH/status/2061885202247356655
Day one of building an agentic hedge fund, and the whole spine went up in a single ~45 minute Claude Code session. He wired the Claude Agent SDK to Robinhood's new agentic MCP endpoint, pulled the OAuth bearer straight out of Claude Code's macOS Keychain, and stood up a 6-layer architecture (data, research, execution, portfolio, monitoring, agent) where every layer is a real Python package. A launchd job fires the premarket routine every weekday at 8:30 ET, writes a state-of-book note, and appends to a decision log. The agent justifies every tool call and logs its reasoning, so auditability is built in rather than bolted on. This is the difference between "ask AI for a trade idea" and handing it the actual plumbing.
@shev_webmarke [Claude Code]
https://x.com/shev_webmarke/status/2061632481603031512
A non-designer's full playbook for getting consistent ad creatives out of GPT-Image-2, with Claude Code as the analysis brain. His core insight: vague prompts drift every single time, so you generate five variations of one simple prompt, feed the good and bad ones back to Claude, and ask it to identify exactly which instruction is causing the drift. Claude diffs the variations, tells you what to specify (model's gaze angle, text position, CTA color), and writes the tightened prompt for you. Then you repeat on the complex prompt until the wobble disappears. It's prompt engineering reframed as a measurable feedback loop, and the whole thing is run by someone who can't design.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2061839510053482777
He got tired of "self-updating memory" being one polite sentence in a config file that the model drops the moment context fills, so he built a real three-tier layer for Claude Code and spent five days hardening it. Advisory is a CLAUDE.md rule (rots, as always). Deterministic is the part that matters: one hook fires the instant a file is written and blocks any claim with no source attached; another fires at session end and won't let the session close until what you learned is filed. Self-improvement, where the agent rewrites its own operating rules, stays human-approved on purpose. The line that took five days to earn: knowledge self-update gets made deterministic, behavior self-update stays human-in-loop.
@kevinma_dev_zh [Claude Code]
https://x.com/kevinma_dev_zh/status/2061630933158232122
A genuinely clever non-coding loop for his product SentiaRead. He runs an agent called Eva on a cloud phone 24/7 scraping Xiaohongshu for real complaints from English learners (lookups breaking reading flow, no progress, abandoned saved material). Then he connects that to local Claude Code via Coze 3.0's multi-agent setup, where Claude Code has access to the codebase, project docs, and the Linear MCP to read issues. The split is deliberate: Eva collects and categorizes the pain points and pulls the users' exact wording, then he decides which feedback becomes a Linear product task and which becomes content (X threads, Xiaohongshu notes). External user feedback wired straight into both the product iteration and the marketing pipeline.
@dikibagast [Claude Code]
https://x.com/dikibagast/status/2061672624904876287
A private framework for testing systematic trading strategies through AI agents while actively fighting overfitting. He set up dedicated markdown files so agents (Claude Code, OpenCode) already understand how to write workflow-compatible scripts, so he just hands them an idea. It pulls OHLCV data across all timeframes, runs single backtests, then does full walk-forward optimization plus statistical validation across thousands of parameter combinations to find a robust sweet spot rather than one overfit curve. Everything is fully local, and once a strategy is confident enough, the backtest script converts to an automated trading strategy. This is the discipline most "AI trading bot" posts skip entirely.
@MushtaqBilalPhD [Claude Code]
https://x.com/MushtaqBilalPhD/status/2061767476493570362
Claude Code automating the title/abstract screening stage of a systematic literature review with two prompts. For anyone who's done academic research, that screening pass is the soul-crushing part: hundreds or thousands of papers, manually marked include/exclude. He walks through creating a "systematic review" folder and letting Claude Code do the first-pass filtering. Not a coding use case at all, which is exactly why it matters: the highest-leverage agent work is increasingly happening outside software, in the repetitive judgment tasks that used to eat a researcher's week.
@scalperof [Claude Code]
https://x.com/scalperof/status/2061761245942886566
Built an entire KPSS exam-prep app with Claude Code while studying for that same exam, and it's live on the Play Store now. He scraped and rewrote 2011-2021 exam questions into original ones, loaded 400+ knowledge cards (heading toward 1000+), added an AI question-generation feature and practice exams, and made the whole thing free. The detail that lands: the tool he needed didn't exist, so the person taking the exam shipped it for everyone else taking it. That's the new default, and the barrier to it just keeps dropping.
@irl_danB [Claude Code]
https://x.com/irl_danB/status/2061929789355962493
He keeps most of his company's operations in a directory (tenets, specs, code, analytics, burn model, sales CRM) and got tired of manually telling Claude Code to "fast-forward" each model when reality changed. So he built the Reactor Harness: think React, but instead of keeping the DOM up to date, it keeps an ideal world-model up to date by re-rendering declared facets when upstream events change. It's a DAG of agent sessions, each memoized, so downstream sessions only re-run when their inputs actually change. He'd tried putting his agent scripts on a cron and it got expensive fast; memoization is the fix. The goal he's chasing is wild: inference cost for maintaining a world-model that scales with surprise, not wall-clock time.
@zach_coredump [Claude Code]
https://x.com/zach_coredump/status/2061754287173378285
Spent the last month migrating an LLM-on-FPGA demo from a KR260 board to a second-hand Xilinx U50, and credits Claude Code (self-funded, he notes) with the modify/refactor/debug work that let him get an LLM running on FPGA inside a single month. He even benchmarked it: speed is close to his MacBook Pro M3 Max 128GB, though the energy efficiency is meaningfully better. This is about as far from "vibe-coded landing page" as it gets, an agent doing real systems work on unfamiliar hardware, and it's a good counter to anyone claiming these tools only work on web CRUD.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2061797074816123017
Original voice models, nearly hands-off, by instructing Claude Code. The flow: hand Claude a base voice, have it tune voice and caption inputs in Irodori-TTS VoiceDesign to mass-produce training data, then train a reusable model. The honest caveat he flags is the actual insight: Claude can't hear what attributes a voice has, so a human still reviews the output. But solve that one gap and you get "describe the voice you want in natural language, get the model," which opens the door to natural-language-driven AI VTuber voices and purpose-built AI personas. A non-coding application where the human is the ear, not the engineer.
@maverickecom [Claude Code]
https://x.com/maverickecom/status/2061842175877399036
A fully automated AI UGC factory built on Claude Code plus Fastmoss, Grok, and CapCut that turns one consistent AI authority figure into hundreds of shoppable creator-style videos across niche accounts daily. The $300/month stack (manus for research, fastmoss for viral competitor content, nano banana pro for images, kling for video) replaced what he says was a $50k+ creative budget, with CPMs as low as $0.10. The sharp part isn't the videos, it's the volume of learning: testing 500+ creatives a month instead of 25, so you find 20x more winners. He's running it like Facebook ads in 2015, before everyone caught on.
@KSimback [Claude Code]
https://x.com/KSimback/status/2061828182454698290
The setup he uses to advise startups, demoed at a cyberfund event. Obsidian is the second brain where everything gets indexed; research and advisory workflows pull from and feed it; agents do too. He uses Claude Code and Codex at his machine, then Hermes agents when remote and for crons and orchestration (and notes he doesn't really use OpenClaw anymore). At the core is a custom skill pack that walks companies through every major deliverable in the 0-to-1 phase. The second brain pattern-matches across his advisory work and research, so the more he runs it, the more it connects dots he wouldn't have.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2061607289417842820
A content engine built entirely in Claude Code that writes in a brand's exact voice, refuses to fabricate stats, and scores every draft 0-100 for AI slop. He pointed it at Grüns, fed it the homepage, and in seconds it built a brand-voice profile (the playful anti-hype tone, the exact cadence, the audience, the review proof). One post came back publish-ready, scored 89/100 on the first pass, answer-first with quotable question headings and zero AI-slop phrases. The architecture is an orchestrator plus writer/editor/researcher subagents. The thing that makes it credible is the slop scorer: a built-in grader that grades the agent's own output before you ever see it.
@moacirmoda [Claude Code]
https://x.com/moacirmoda/status/2061829390980260104
A small but genuinely useful habit for compounding your skills. After Claude Code finishes a process that's roughly mapped out, ask it for the session ID, open a fresh chat, and paste that ID with a prompt: analyze the whole session and tell me the errors, the wins, what was asked that the skills didn't cover, and what adjustments would make next time faster and burn fewer tokens on mistakes. List only, don't act yet. It's a manual self-improvement loop, turning your own work history into the material that sharpens the skill, which is exactly the move most people skip.
@kota33000604 [Claude Code]
https://x.com/kota33000604/status/2061807399137370317
"Content as Code" is landing hard at his company, and it's a non-engineering use that's quietly clever. Marketing and product-marketing changes now run through Pull Requests, same as engineers ship code: PMM changes positioning, marketing edits the landing page, a lead reviews. The reviewer doesn't write code, they look at a preview URL, leave line comments, and approve. The unlock is that Claude Code lets PMM and marketing ship HTML without being able to write it. If content is also a deliverable, version-controlling it like code turns out to be faster.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2061741864664109377
Stopped checking social media for AI news and says he's more informed than ever, because Claude Code reads the internet for him every morning at 6am, filtered by his focus areas, and drops a brief into his Obsidian vault. No doomscrolling, no 200 tweets about the same launch. It's a tiny personal automation, but it's the kind of thing that used to require stitching together an RSS reader, a summarizer, and a notes app, and now it's one scheduled agent task pointed at a vault.
@wilderko [Claude Code]
https://x.com/wilderko/status/2061764221432385926
A genuinely future-feeling setup: he runs an active Claude session entirely by voice through Even Realities AR glasses. The connection is "Terminal Mode," activated by scanning a QR code on the page where Claude Code is running, with both sides bridged over a Tailscale VPN so the link isn't unencrypted over the open internet. He switches between options using an R1 ring or by tapping the side of the glasses, and can code or solve problems hands-free. It's buried in a long post about navigation and live translation, but the Claude-over-glasses bit is a real glimpse of computing that follows you instead of the other way around.
@itsbrianscherer [Claude Code]
https://x.com/itsbrianscherer/status/2061854000476303576
Hard data from a solo and small law firm tech-stack survey, and the AI numbers are striking. Claude is the single most-used tool in the entire report (84% adoption, 4.5/5 satisfaction). On legal research, Claude now leads on adoption (36%) over Westlaw (31%), a thing that would have sounded absurd a year ago. And the "vibe coding" category, which barely existed twelve months ago, is led by Claude Code at 73%: lawyers are now building their own software. When a famously conservative profession is shipping its own tools, the adoption curve isn't a tech-bubble story anymore.
@kj_kombat [Claude Code]
https://x.com/kj_kombat/status/2061777850781573531
A useful reality check amid the hype. He burned $50 having Claude Code batch-process 1.2 GB of PDF files for his father's audit work, and his honest take is that he could have paid a data engineer a similar hourly rate instead, concluding "I don't think AI is replacing us." It's a real, specific job with a real cost, and the skepticism is the value: not every batch task is a 100x win, and the people actually running the bills are the ones worth listening to on where the economics break even.
🗣 User Voice
User Voice
Rate limits are the number-one source of pain right now, and they're killing flow at the worst moment. @Saanvi_dhillon asks straight up if there's any way to use Claude Code without hitting the usage limit so fast, "every time I get into a productive workflow, I hit the limit." Even an Opus 4.8 subagent bug that silently burned people's quota (forcing Anthropic to reset everyone's limits) shows how fragile the budget feels to heavy users.
Cost anxiety is the close second, and it's specific, not abstract. @kj_kombat burned $50 on one PDF audit batch and openly questioned whether a data engineer would've been cheaper. The whole Mac-mini-local-Ollama wave exists because people are doing the math on token bills, and they don't like the answer.
Memory that doesn't survive tool-switching is a real structural complaint. @himanshutwtxs nails it: memory is harness-scoped, so what Claude Code remembers means nothing to Codex, and the fix is to treat memory as infrastructure, not a setting inside one tool or a markdown file. Multiple people are building portable memory layers because the built-in ones rot or don't travel.
Opus 4.8 quality regression keeps coming up from real users. @Fabiobuilds noticed Opus 4.7 in the web app keeps correcting Opus 4.8 in Claude Code "like it was a summer intern," and a widely-shared news roundup flagged a wave of developer reports that 4.8 struggles with pronoun tracking and following arguments. When your power users say the newest model feels dumber, that's worth listening to.
People love Workflows but want to see what's happening. @dani_avila7 argues the biggest potential of Dynamic Workflows unlocks once it's in Claude Code Desktop with a UI that visualizes the entire agent orchestration, because for systems like this "observability is everything." @daptonai compares an over-eager agent to a memory leak: everything looks fine until a real job spins up background work you never asked for, and you need to see what your agent is actually doing in real time, not just the output.
Rate limits are the number-one source of pain right now, and they're killing flow at the worst moment. @Saanvi_dhillon asks straight up if there's any way to use Claude Code without hitting the usage limit so fast, "every time I get into a productive workflow, I hit the limit." Even an Opus 4.8 subagent bug that silently burned people's quota (forcing Anthropic to reset everyone's limits) shows how fragile the budget feels to heavy users.
Cost anxiety is the close second, and it's specific, not abstract. @kj_kombat burned $50 on one PDF audit batch and openly questioned whether a data engineer would've been cheaper. The whole Mac-mini-local-Ollama wave exists because people are doing the math on token bills, and they don't like the answer.
Memory that doesn't survive tool-switching is a real structural complaint. @himanshutwtxs nails it: memory is harness-scoped, so what Claude Code remembers means nothing to Codex, and the fix is to treat memory as infrastructure, not a setting inside one tool or a markdown file. Multiple people are building portable memory layers because the built-in ones rot or don't travel.
Opus 4.8 quality regression keeps coming up from real users. @Fabiobuilds noticed Opus 4.7 in the web app keeps correcting Opus 4.8 in Claude Code "like it was a summer intern," and a widely-shared news roundup flagged a wave of developer reports that 4.8 struggles with pronoun tracking and following arguments. When your power users say the newest model feels dumber, that's worth listening to.
People love Workflows but want to see what's happening. @dani_avila7 argues the biggest potential of Dynamic Workflows unlocks once it's in Claude Code Desktop with a UI that visualizes the entire agent orchestration, because for systems like this "observability is everything." @daptonai compares an over-eager agent to a memory leak: everything looks fine until a real job spins up background work you never asked for, and you need to see what your agent is actually doing in real time, not just the output.
📡 Eco Products Radar
Eco Products Radar
NotebookLM and Obsidian are the breakout combo this cycle, showing up again and again as the "research pipeline" pairing with Claude Code: NotebookLM eats the sources, Obsidian remembers the decisions, Claude Code routes the work. Codex remains the constant comparison point and co-pilot (many run Claude Code and Codex side by side). Cursor still appears throughout, though several posts note developers dropping back to raw Claude Code. Hermes Agent shows up repeatedly as the remote/cron counterpart to local agents. Supabase and Vercel are the default backend-and-deploy stack for the "build a whole app/social network in 4 weeks" posts. Ollama anchors the local-model movement (route Claude Code at a local model, zero API cost). Claude Code Router (ccr) keeps appearing as the way to point Claude Code at free or cheaper models. And MCP is now table stakes, the connective tissue under nearly every serious workflow above.
NotebookLM and Obsidian are the breakout combo this cycle, showing up again and again as the "research pipeline" pairing with Claude Code: NotebookLM eats the sources, Obsidian remembers the decisions, Claude Code routes the work. Codex remains the constant comparison point and co-pilot (many run Claude Code and Codex side by side). Cursor still appears throughout, though several posts note developers dropping back to raw Claude Code. Hermes Agent shows up repeatedly as the remote/cron counterpart to local agents. Supabase and Vercel are the default backend-and-deploy stack for the "build a whole app/social network in 4 weeks" posts. Ollama anchors the local-model movement (route Claude Code at a local model, zero API cost). Claude Code Router (ccr) keeps appearing as the way to point Claude Code at free or cheaper models. And MCP is now table stakes, the connective tissue under nearly every serious workflow above.
Comments