Super User Daily: June 22, 2026
The clearest signal today is how far Claude Code has drifted away from the editor. A follower's one-line tip became a weather-trading bot that cleared $900 in a day; another builder fed 86 million trades into a Polymarket bot and grew $200 into $14,300; a Raspberry Pi and one radio now paint a live aircraft radar across someone's ceiling. The cost conversation got loud and concrete: a German developer cut a $1,140/month cloud bill to $47 by moving five agents on-prem, a team cut Claude Code spend 2.5x just by swapping the backend context layer, and a kid in Warsaw runs a 235B model on a lunchbox PC for $14 of electricity. And the most durable pattern of the day is the two-model adversarial setup — Claude Code for brainstorming and creative work, Codex or GLM as the skeptical second reviewer it calls itself, wired straight into CLAUDE.md. The harness, not the model, is where people are spending their judgment.
@xmayeth [Claude Code]
https://x.com/xmayeth/status/2068358211355247047
A follower DM'd him a one-line weather edge on Polymarket, and he turned it into a working bot with Claude Code that was up $900 in a day. The bot pulls live temperature from the exact settlement station (e.g. KLGA), estimates the realistic daily high, and buys baskets when probability tops ~90% and price is under 80c with under 18 hours left. The whole edge is the lag between physically-settled temperatures and markets still pricing on a stale morning forecast. He shares the concrete entry logic and a live example.
@ZayvenKnox [Claude Code]
https://x.com/ZayvenKnox/status/2068335343036965358
He runs a Polymarket trading bot built on Claude Code across four repos for about $25/month, fed raw wallet data from 86 million trades to find and clone winning wallets. Claude scanned 14,000 wallets in 4 minutes, ranking those with 100+ trades and a >70% win rate. Three agents (arbitrage, convergence, whale-copy) share one wallet behind a consensus filter that killed 40% of losing trades, plus a Rust CLI that scores 500 markets down to ~35. He reports growing a $200 seed into $14,300 in 27 days.
@francescoinweb3 [Claude Code]
https://x.com/francescoinweb3/status/2068341390975602826
A developer in Germany spent about $3,200 on a mini-server and now runs five local AI agents that monitor 120 crypto Twitter accounts, vet new projects against categories, analyze documents, write client reports, and run and check code. Before going local he was paying roughly $1,140/month across Claude Code Max, ChatGPT Pro and various APIs; on-prem dropped it to about $47/month. He says the rig has chewed through 1.8 million tweets and tens of thousands of classification tasks since.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2068212251988038072
A developer near SFO used Claude Code to build a system that projects a real-time aircraft radar onto his ceiling using just a single radio and a Raspberry Pi. The radio receives aircraft signals overhead, the Pi decodes them in real time, and the ceiling shows passenger jets, helicopters and Cessnas, while also tracking the sun, moon, stars and the ISS. It runs fully locally with no internet and no subscription, published openly under the name Skylight.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2068384508077105538
He tried the open-source 'Understand Anything' plugin, which analyzes a codebase and turns it into an interactive knowledge base that explains itself to you rather than just showing structure. Although it's built for code, he pointed it at his Obsidian vault of podcast highlights and ran /understand. The result was a searchable knowledge graph built from highlights across 888 podcast episodes and 144K lines of markdown.
@_MaxBlade [Claude Code]
https://x.com/_MaxBlade/status/2068412065547513881
He reports earning $13k in 9 days while live-streaming the build of CNVS, a macOS app written from scratch in Swift, every day for 46 days. The app runs NVIDIA Parakeet locally to spawn agents in under 300ms and integrates GPT realtime 2 for a Jarvis-like conversational mode. Crucially it rides on the user's existing subscriptions (Claude Code, Codex, Cursor) so they avoid API pricing, and adds Hermes integration, VPS-hosted remote canvases and cross-agent memory.
@wei_wang [Claude Code]
https://x.com/wei_wang/status/2068477244713169256
He used Claude Code to rebuild his resume: he uploaded the original PDF, had it parse the file and convert it to LaTeX, then iterated through several rounds of edits to content, layout and wording before compiling back to PDF. The punchline is that he doesn't know LaTeX — Claude Code installed the LaTeX toolchain, the PDF-parsing tools and the compile environment itself, running the whole thing essentially autonomously. He calls it the best resume of his working life.
@saidstetic [Claude Code]
https://x.com/saidstetic/status/2068255682411827538
An open-source AI job-search system built for Claude Code that sent over 700 job applications and ultimately landed its builder a job. It automates the whole pipeline: reviewing different companies' job pages, tailoring the CV per posting, and even filling in application forms. The repo ships 14 modes (evaluation, scraping, PDF generation), a Go terminal dashboard, ATS-optimized PDF CVs via Playwright, and 45+ pre-configured companies including Anthropic, OpenAI, ElevenLabs and Stripe.
@PodcastAlphaX [Claude Code]
https://x.com/PodcastAlphaX/status/2068431946431008972
Box CEO Aaron Levie connected Salesforce's MCP server to Claude Code and now reportedly uses Salesforce 5x more than ever before. He runs customer and market-intelligence queries through the agent that he'd never have bothered to pull up by hand, because the agent removes the friction. The takeaway framed here: the underlying data system gets queried more, not replaced. Source cited as CXOTalk.
@DailyDoseOfDS_ [Claude Code]
https://x.com/DailyDoseOfDS_/status/2068265152500343283
They cut Claude Code costs 2.5x on a full-stack RAG app just by swapping the backend context layer: the Firebase run used 15.7M tokens, 4 errors and $12.95, while the InsForge run hit the same final build with 6.3M tokens, 0 errors and $4.87. The diagnosis: Firebase APIs over-return data and give the agent no debugging context, so it queries repeatedly and guesses on errors, burning tokens. InsForge instead exposes the full backend topology in one ~500-token CLI call and returns structured JSON with exit codes.
@matrosov [Claude Code]
https://x.com/matrosov/status/2068434095005409687
He reports Claude Code is a very powerful tool for reverse-engineering work, looping it with IDA and the decompiler over MCP or idalib in combination with dynamic workflows. He shares concrete commands: a /goal to find and list every crypto implementation mapping algorithm to address, and a /loop instructing it to use the decompiler but verify all findings with the disassembler. He says this delivers very impressive results on RE.
@sudoingX [OpenClaw]
https://x.com/sudoingX/status/2068393737911509337
Called a hater for saying running OpenClaw on local models means you'll struggle, he went into both the Hermes Agent and OpenClaw codebases to show why technically. He reports concrete differences: Hermes ships 11 model-specific tool-call parsers plus a repair chain while OpenClaw has none and only a regex fallback; Hermes runs on 18 dependencies versus OpenClaw's lockfile pulling thousands; and Hermes' agent loop is one file while OpenClaw is thousands of TypeScript files. His conclusion: the tool-call failures people blame on the model are actually the harness.
@ArielBravy [OpenClaw]
https://x.com/ArielBravy/status/2068182087765770576
He has OpenClaw automating his Bitaxe / NerdQAxe Bitcoin miner settings by season — overclocking harder in winter when the ambient temperature is cold, and backing the settings down in summer so the miners don't keep overheating. It's a clean little non-coding home-hardware automation driven entirely by OpenClaw, with a clip attached as proof.
@cryppimagic [Claude Code]
https://x.com/cryppimagic/status/2068365821856845931
With no coding knowledge of his own, he used Claude Code to build a parser that collects funding-rate data from 50 perpetual DEXes and CEXes, for the purpose of farming funding rates at yields he cites as 1%+ per hour. He plans to automate the farming itself later and decided to release the parser for free.
@nett0eth [Claude Code]
https://x.com/nett0eth/status/2068390867149164970
He introduces Ponytail, an open-source skill that makes Claude Code reason through six steps before writing any code — does it need code at all, is it in the standard library, is it platform-native, can an installed dependency do it, can it be one line, otherwise write the minimum. He reports concrete results with Opus: 71% fewer lines of code, 53% lower cost and 71% faster, while keeping security and error handling intact. It's a single skill file you toggle on and off, no closed platform.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2068132643468640344
He details a one-person GTM back office run by Claude Code as a team of role-based seats. A prospector finds accounts from four triggers (job posts, competitor mentions, launches or stack changes, funding or acquisitions), a researcher writes a one-page brief per call, a sequencer sends follow-ups, a recoverer handles no-shows with four touches, and a reporter re-weights triggers every Friday. The key design is a single shared account memory every seat reads and writes, producing an 8am Slack standup the human approves or edits.
@IBuzovskyi [Claude Code]
https://x.com/IBuzovskyi/status/2068403016433291302
He describes a concrete brand-deals manager built on Hermes that monitors the creator's inbox 24/7, knows their rates and negotiates. The setup uses one SOUL.md with per-platform rates and rules, an email MCP, and a cron job every 30 minutes that skips empty inbox checks; new brand emails get evaluated, a reply drafted, and a summary sent to Telegram for one-tap approval. Stated cost is a $7/month VPS plus $5-15/month in tokens, roughly $12-22/month total.
@aniketapanjwani [Claude Code]
https://x.com/aniketapanjwani/status/2068426404123779321
An economist used the /goal command in Codex and Claude Code to convert Kirill Borusyak's Stata/Python difference-in-differences imputation package into an R package, leaving a ~15-hour Codex goal running overnight. He stresses defining validation criteria up front, using the existing package to pin down the expected behavior of the new one. He also demos running Codex and Claude Code together (Claude Code from inside Codex) and links the video, the paper and the generated R package.
@Fluyeporlaweb [Claude Code]
https://x.com/Fluyeporlaweb/status/2068393623037624698
He highlights LEANN, an MIT-licensed tool that shrinks a 201GB vector corpus to about 6-7GB by selectively recomputing embeddings via a pruned high-connectivity graph instead of storing them all — no GPU, no precision loss. He points out Claude Code only does keyword search, and LEANN installs as an MCP to add real semantic search over files, emails, browser history and Claude/ChatGPT/iMessage chats without changing the workflow. A drop-in MCP for Claude Code, Cursor and Windsurf that runs offline on CPU.
@__BUNTA__ [Claude Code]
https://x.com/__BUNTA__/status/2068465919878119506
He argues systems worth 20-30 million yen can now be built in three weeks by a single employee who understands the business — a practice his firm calls an 'in-house production revolution.' His concrete example: a client built a tool with Claude Code to fully automate Instagram operations of 10 posts a month across 7 accounts, entirely in-house. Development that would have taken half a year without AI took only three weeks.
@WeiYipei [Claude Code]
https://x.com/WeiYipei/status/2068156055784419813
Over two months a solo operator produced 37 SEO blog posts, daily tweets, 45 HuggingFace datasets and 55 ClawHub skills using just Claude Code plus a few automations. The workflow has Claude Code read his skill files, generate E-E-A-T-framework drafts that he edits ~20%, then publish; methodologies are packaged into SKILL.md files synced across GitHub/HF/ClawHub, with distribution automated via Buffer and IndexNow. His principle: humans handle 20% judgment, AI handles 80% execution, and the hard part is having enough methodology density to execute on.
@0xKiyoro [Claude Code]
https://x.com/0xKiyoro/status/2068367331118453062
A freelancer makes $20,000/month building websites and booking pages for local businesses at $800-$2,000 each, with Claude Code building the whole thing from the same 3-page template every time. He charges $150/month per site to keep it live, and with 40 clients on retainer that's $6,000 of recurring monthly revenue. The framing: one-time builds got him to $20k, the retainers make the income stick.
@dorcussaga [Claude Code]
https://x.com/dorcussaga/status/2068249026210975790
He had Claude Code build a sales-management application and found it genuinely useful. He notes that when he previously asked a software company about a similar spec, he was quoted roughly 500,000 to 1,000,000 yen and gave up. With Claude Code the same thing was built in a few hours for only a few thousand yen a month — which he calls an amazing change of era.
@yousukezan [Claude Code]
https://x.com/yousukezan/status/2068125413327683990
Snyk's vulnerability reports come in English, so he turned them into a Claude Code plugin that outputs the reports in Japanese. He says doing this made the value of the reports 100x for his use, and links a Zenn writeup describing the build.
@thefunnyguysNFT [Claude Code]
https://x.com/thefunnyguysNFT/status/2068327881470525470
A neat non-coding use: he uses Claude Code to generate plain-language walkthroughs of the algorithms behind procedural generative NFT artworks from platforms like Art Blocks and fxhash. He produces a step-by-step explanation for 'Memories of Qilin' by Emily Xie — how a base color is picked from a random seed, regions are carved and filled, then painterly ink-blot splotches and a paper-grain texture are layered on. The point: collectors can finally inspect the generative logic themselves.
@gokulr [Claude Code]
https://x.com/gokulr/status/2068450343432990880
After using Claude Code (Opus 4.7/4.8) and Codex (GPT-5.5) heavily for weeks, his verdict is that each tool's strength is also its weakness: use Claude Code for brainstorming and planning, Codex for execution and building, and have them adversarially review each other's plans. He finds Claude Code very creative as a brainstorming partner but prone to hallucinating during execution, while Codex is a fast, focused executor but poor at generating new options. He adds that friends at both labs agree and use the rival's product for exactly those gaps.
@DeeperThrill [Claude Code]
https://x.com/DeeperThrill/status/2068410223426187404
He runs both Codex (GPT 5.5) and Claude Code (Opus 4.8) at $200/month each and finds Codex writes uglier code but catches edge cases Claude misses. His workflow uses Claude Code as the daily driver but regularly tells it to call the codex CLI via bash for a second opinion on its own code. That instruction lives in his CLAUDE.md, which is a symlink to AGENTS.md, so Claude frequently consults Codex to catch its blind spots.
@EdwardDGregory [Claude Code]
https://x.com/EdwardDGregory/status/2068337726013710345
He had Claude Code create a wrapper so he can launch Claude Code with GLM as the model just by typing 'glm' on the command line. His code-review skill then invokes that GLM-backed setup as one part of a four-agent adversarial code review — a concrete custom multi-model review pipeline.
@Da7_Tech [Claude Code]
https://x.com/Da7_Tech/status/2068483705887228343
A vendor marketed GLM Coding Pro as '15x Claude Code,' so he spent a week actually testing the plan with GLM-5.2. His real-world measured result was about 2.35x — not 15x, not even 2.5x. He frames the gap between the marketed and observed performance as a transparency problem.
@ryoppippi [Claude Code]
https://x.com/ryoppippi/status/2068473646906036481
He points out that Claude Code stores session history as JSONL, so a session can be resumed on a different machine simply by syncing those files. He drops the history into iCloud Drive to share and resume sessions across machines (crediting ghq, and noting pi-agent works too). He contrasts this with Codex, which inexplicably uses SQLite and therefore, he feels, can't be shared the same way.
@tom_doerr [Claude Code]
https://x.com/tom_doerr/status/2068169318647345617
He shares a project for automated malware reverse engineering using Claude Code together with Binary Ninja — a notable, atypical application of Claude Code in the security and reverse-engineering domain, with the repo linked.
@Maybajpa [Claude Code]
https://x.com/Maybajpa/status/2068236092189606215
He connected multiple financial-data apps into Claude Code — naming MarketSmith India and StockScans among others — so Claude Code pulls together all the details across the various data sources he needs. A concrete first-person account of wiring real market-data tools into Claude Code for investment research.
@hAru_mAki_ch [Claude Code]
https://x.com/hAru_mAki_ch/status/2068309767811158526
He runs a 'remote employee agent' on Claude Code paired with GLM 5.2, and goes beyond just running it to an operating model that records failures so they don't recur. After a /goal and an agent-team smoke test, he organized it into real tasks, browser QA, SessionStart/Stop hooks, a recurrence-prevention CLAUDE.md, and long-running timeout/retry handling. A concrete self-improving operational workflow built from specific Claude Code primitives.
@0xdeusyu [Claude Code]
https://x.com/0xdeusyu/status/2068162022865600942
He shipped a free tool he'd always wanted — Jinshu (jinshu) — to fix WeChat public-account formatting, where pasting Markdown breaks the layout every time. It renders Markdown into WeChat-styled output in real time with 28 mdnice themes and one-click copy. Its standout feature is an MCP service with a single endpoint usable by Claude Code, Cursor, Codex and Gemini, so the AI formats the article right after writing it, with a live preview that refreshes within a second on each edit.
@kiyochanshacho [Claude Code]
https://x.com/kiyochanshacho/status/2068213003607019523
He recounts a non-CS-background ('humanities') engineer who built a personal study app with Claude Code and passed the demanding CISSP security certification in three months. He marvels at it and reflects that excuses like being non-technical or inexperienced are losing their validity — what matters now is motivation plus AI, not your starting point.
@liquiditygoblin [Claude Code]
https://x.com/liquiditygoblin/status/2068184914995126506
A first-person account of Claude Code's remote-control feature: he submits PRs and runs code on his home laptop entirely from his phone. He describes doing it while out shopping with his wife, and says it makes the work much more fun — a real-world use of the phone-based remote workflow rather than an announcement.
@shmidtqq [Claude Code]
https://x.com/shmidtqq/status/2068346186918924334
He built a playable neon arcade game without writing a single line of code by giving MiniMax M3 one prompt inside Claude Code: 'build a neon arcade game.' The model planned the steps, wrote the code, ran it, hit an error and fixed itself, producing a game he can actually play. He notes MiniMax M3 is open weights with a 1M-token context and 59% on SWE-Bench Pro.
@anshuc [Claude Code]
https://x.com/anshuc/status/2068166284689494432
He built an elaborately overdesigned calorie tracker app entirely with Claude Code in one week, without touching a line of code and without any MCPs or skills — purely careful prompting. The app went viral on X with 700k+ views, he now uses it daily, and it's currently in App Store review. His argument: good design comes from using human judgment to ask the AI for the right things, not from the AI itself.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2068316045836693938
He implements Karpathy's 'LLM wiki / second brain' setup live with Claude Code and Obsidian. He pointed Claude Code at 36 of his own YouTube transcripts and asked it to organize them; Claude split them into linked pages and wired up relationships between every tool, technique and concept into a queryable map. Dropping in the AI 2027 scenario article then produced 25 cross-linked pages in minutes — markdown only, no vector DB, no embeddings.
@huang_chao4969 [Claude Code]
https://x.com/huang_chao4969/status/2068213679066390748
He released UpSkill, a lightweight open-source framework that continuously evolves agent skills so a cheaper Flash-tier model can perform closer to a Pro-tier one. His motivation is concrete: Pro models like Opus, GPT-5.5 and Gemini Pro cost 3-5x more per task and are unsustainable at scale, while Flash models are cheap but unreliable enough that fixing their errors erases the savings. It's positioned for Claude Code, Codex, Cursor, OpenClaw and Hermes.
@Degen_calls_sol [Claude Code]
https://x.com/Degen_calls_sol/status/2068125859890741678
After seeing AMD's keynote he canceled all his AI subscriptions to run models locally: he bought a GMKtec Evo-X2 with a Ryzen AI Max+ 395 and 128GB (110GB carved for VRAM), installed Ubuntu, pulled Ollama, downloaded Qwen3 235B, and pointed Claude Code at localhost for the same interface and workflow. Reported result: $0 on AI subscriptions for the month and about $14 on electricity, with a 235B model running entirely on a lunchbox-sized local machine.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2068333056419655915
He connects getleads.io's API to Claude Code to generate lead lists for his founder friends. He gives Claude a company URL plus ICP context and asks it to find that company's target customers — effectively pulling a company's entire total addressable market — then texts his friends offering the list and to market to them. He says he's having a lot of fun with it.
🗣 User Voice
User Voice
Token cost has become the dominant anxiety, and people are voting with their hardware. @francescoinweb3 cut a $1,140/month bill to $47 by going on-prem, @Degen_calls_sol runs Qwen3 235B locally for $14 of electricity, and @DailyDoseOfDS_ shows a 2.5x cost swing from one backend choice — the demand is for predictable, controllable spend, not raw capability.
The recurring complaint is that the model is fine but the harness is the bottleneck. @sudoingX went into two codebases to prove tool-call failures blamed on local models are actually the harness, and @nett0eth's Ponytail skill exists purely to stop Claude Code from over-generating code — users want better scaffolding, not bigger models.
Nobody trusts a single agent to check its own work anymore. @DeeperThrill, @gokulr and @EdwardDGregory all converged independently on having Claude Code call a second model (Codex or GLM) as an adversarial reviewer — the expectation now is multi-model review baked into the workflow.
The most valued use cases are non-coding. @ClaudeCode_UT (ceiling radar), @ArielBravy (seasonal miner tuning), @thefunnyguysNFT (NFT algorithm explainers) and @wei_wang (a LaTeX resume by someone who doesn't know LaTeX) all point the same way: people want Claude Code to operate their whole life, not just their repo.
Token cost has become the dominant anxiety, and people are voting with their hardware. @francescoinweb3 cut a $1,140/month bill to $47 by going on-prem, @Degen_calls_sol runs Qwen3 235B locally for $14 of electricity, and @DailyDoseOfDS_ shows a 2.5x cost swing from one backend choice — the demand is for predictable, controllable spend, not raw capability.
The recurring complaint is that the model is fine but the harness is the bottleneck. @sudoingX went into two codebases to prove tool-call failures blamed on local models are actually the harness, and @nett0eth's Ponytail skill exists purely to stop Claude Code from over-generating code — users want better scaffolding, not bigger models.
Nobody trusts a single agent to check its own work anymore. @DeeperThrill, @gokulr and @EdwardDGregory all converged independently on having Claude Code call a second model (Codex or GLM) as an adversarial reviewer — the expectation now is multi-model review baked into the workflow.
The most valued use cases are non-coding. @ClaudeCode_UT (ceiling radar), @ArielBravy (seasonal miner tuning), @thefunnyguysNFT (NFT algorithm explainers) and @wei_wang (a LaTeX resume by someone who doesn't know LaTeX) all point the same way: people want Claude Code to operate their whole life, not just their repo.
📡 Eco Products Radar
Eco Products Radar
GLM-5.2 — the open-weight model people now route Claude Code to via wrappers; tested, daily-driven, and benchmarked against the marketing (one user measured 2.35x, not the claimed 15x).
Codex / GPT-5.5 — the standard second opinion; called from inside Claude Code via bash as an adversarial reviewer.
Hermes Agent — the self-improving agent people build always-on background workers on (brand-deal negotiators, monitors).
Obsidian — the second-brain substrate Claude Code reads and writes as a knowledge graph.
Ollama — the local-inference runtime behind the cancel-your-subscriptions, run-it-on-a-lunchbox moves.
MCP servers — the connective tissue (Salesforce, market data, LEANN semantic search, WeChat formatting) turning Claude Code into an everything-client.
MiniMax M3 — open-weight 1M-context model used to one-shot builds inside Claude Code.
Ponytail / Graphify / LEANN — the skill-and-MCP layer for cutting tokens and adding real retrieval.
GLM-5.2 — the open-weight model people now route Claude Code to via wrappers; tested, daily-driven, and benchmarked against the marketing (one user measured 2.35x, not the claimed 15x).
Codex / GPT-5.5 — the standard second opinion; called from inside Claude Code via bash as an adversarial reviewer.
Hermes Agent — the self-improving agent people build always-on background workers on (brand-deal negotiators, monitors).
Obsidian — the second-brain substrate Claude Code reads and writes as a knowledge graph.
Ollama — the local-inference runtime behind the cancel-your-subscriptions, run-it-on-a-lunchbox moves.
MCP servers — the connective tissue (Salesforce, market data, LEANN semantic search, WeChat formatting) turning Claude Code into an everything-client.
MiniMax M3 — open-weight 1M-context model used to one-shot builds inside Claude Code.
Ponytail / Graphify / LEANN — the skill-and-MCP layer for cutting tokens and adding real retrieval.
Comments