Super User Daily: May 28, 2026
The most striking thing this week isn't a new feature, it's how far people are pushing Claude Code away from the terminal. An accountant is filing journal entries through it. A finance analyst handed it 90% of his investment research. A guy in Da Nang runs a whole Shopify store with one $20 subscription. The pattern underneath all of it: the agent isn't the coder anymore, it's the operator, and the human is the one giving direction and checking the work. And the bill is becoming the story, with Microsoft cutting Claude Code licenses not because it's bad but because metered token billing exploded. Here are the real use cases worth copying.
@lagerskoy [Claude Code]
https://x.com/lagerskoy/status/2059229489574613489
Runs an entire Shopify store solo from an apartment in Da Nang on a $20/month plan. Claude Code in one terminal, Shopify CLI hot-reloading the dev store in another, order labels printing in real time as packages go out. He wires Claude Code straight into the live store through Shopify's first-party MCP server. What used to need an agency plus a developer plus a content team is now one person and a subscription.
@VengeonsP [Claude Code]
https://x.com/VengeonsP/status/2059218176072732871
Posts a full 0-to-$100k-in-five-months playbook for ChatSEO. Validation landing page built in a weekend with Lovable plus Claude Code, traffic from YouTube and Reddit hitting a 35% signup rate. The sharp part is the paywall surgery: moved it from generous freemium (0.8% conversion) to the sixth prompt (10% conversion) after spotting that activation happens at five messages. He also dumped every customer call transcript into NotebookLM so he could literally chat with his own customer base.
@shivsakhuja [Claude Code]
https://x.com/shivsakhuja/status/2059086745506046329
Ships a 45-second animated explainer ad in 30 minutes, no video editor touched. The whole pipeline is pre-built skills: /plan writes the brief, /prepare builds a moodboard with GPT Image 2 and ElevenLabs, /generate makes keyframes, /animate runs Fal Seedance, /stitch glues it together with ffmpeg plus music and captions, /watch reviews it, /learn updates the skills. The human just gives direction and reviews.
@Kuroi_CPA [Claude Code]
https://x.com/Kuroi_CPA/status/2059264043475697728
Connects MoneyForward cloud accounting to Claude Code over MCP and does real bookkeeping in plain language. He types things like 'reclassify from accounts payable to accrued liabilities' and it generates the transfer journal entry, or 'run a year-over-year variance analysis, pull accounts that moved more than 30% and explain why.' His strong warning: the AI sometimes proposes the wrong account classification, so you must know accounting to verify every move it makes.
@IShmool [Claude Code]
https://x.com/IShmool/status/2059252829991190696
Built zero2claude, a free course platform, solo with every line written by Claude Code: 30,000 students across 40+ countries, 7,000 requests a minute, 10M+ requests a day at a 0.003% error rate. Then he built a Telegram bot that spawns Claude Code agents against production for monitoring: one command checks every endpoint, another digs through logs when something breaks, another blocks shipping if checks fail, another texts him the overnight numbers. Stack is Node, the Telegram Bot API, tmux and the Claude Code CLI.
@Axel_bitblaze69 [Claude Code]
https://x.com/Axel_bitblaze69/status/2059071374023622795
Turned Claude Code into a research analyst that has both memory and hands. Memory is an Obsidian vault with raw/wiki/output folders and a CLAUDE.md 'librarian' rulebook; hands are Playwright and Firecrawl CLIs plus the Obsidian Web Clipper. The concrete payoff: it scraped two weeks of US insider-trading data, analyzed it, and produced a ranked top-10 watchlist with reasoning in 3 minutes versus about 90 minutes by hand. Runs in autonomy mode scoped to the vault folder.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2059325068468215938
Plugged Meta's official Ads CLI into Claude Code and killed about 80% of a weekly Meta Ads reporting workflow without ever logging into Ads Manager. One sentence spins up a live dashboard with KPI cards, top-10 ad set rankings, a spend chart and a sortable table in about 90 seconds. It also does week-over-week comparisons flagging CTR drops and CPC spikes, creative fatigue audits, one-page exec briefs and anomaly reports. No third-party connector, no ban risk.
@milbon_ [Claude Code]
https://x.com/milbon_/status/2059258528112759270
Lays out a method to mass-produce 10,000 SEO pages in 48 hours with Claude Code. Extract a pile of niche keywords, scrape and structure competitor data with Firecrawl, then let Claude Code auto-generate unique content for each page. The claimed result: 300k monthly visits at a 2% conversion rate, around 9M yen a month. Whether you buy the numbers or not, it's a concrete programmatic-SEO pipeline.
@yanndine [Claude Code]
https://x.com/yanndine/status/2059335673111294347
Built a free 40-skill B2B outbound plugin that runs an entire outbound motion from one brief. You answer eight onboarding questions about your company and ICP, then Claude builds the full brief: ICP, persona cards, campaign architecture. Cold emails, LinkedIn DMs and call scripts all pull from the same ICP file, campaigns are benchmarked against 244,000 real outbound campaigns, and reply handling is automated. He uses it daily.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2059314545844179365
A finance analyst automated about 90% of his investment research over 60 days inside Claude Code. Screening, scraping, modeling and drafting reports all run in the background. Claude Code drives a real browser through Playwright, pulls insider-trading data off Finviz, scrapes 10-Ks and builds a model in roughly 3 minutes. Obsidian sits behind it as memory, filing clipped articles into topic folders and recalling them weeks later.
@godofprompt [Claude Code]
https://x.com/godofprompt/status/2059202893203583258
An Anthropic engineer pointed Claude Code at his old startup's infrastructure to hunt for cost savings. It audited the whole stack, found three unused Redis instances, proposed deletions, and after per-step approval executed the cleanup in parallel, recovering $104 a month. The detail worth noting: even in auto mode, it asked for explicit approval before deleting anything.
@Oluwaphilemon1 [Claude Code]
https://x.com/Oluwaphilemon1/status/2059256086608130149
Built a fully AI motion-design pipeline inside Claude Code. It pulls references through the Pinterest API, generates a six-scene storyboard with GPT Images 2.0, then renders straight to video in Seedance 2.0 via the Higgsfield MCP, all stitched into one flow. A clean example of using Claude Code as the orchestrator for a creative production chain rather than as a coding tool.
@paraschopra [Claude Code]
https://x.com/paraschopra/status/2059167147516199152
Built a Claude Code skill that makes static HTML outputs interactive. Claude launches a localhost server, you leave Google-Docs-style comments right on the page, and Claude reads them, updates the page live, and gives you a tour of what changed. He built it because his preferred workflow is having Claude generate HTML files as deliverables, and he wanted to direct them like a document.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2059303338706194714
Arize's CPO showed building a 'vibe eval' in 5 minutes with Claude Code and deliberately ran it on day one knowing the results would be noisy. The eval still exposed the right failure categories: bugs scored too low, feature requests rated above production issues. The loop is the point: Claude suggests the eval, runs it across all traces, then the PM gives ruthless criticism to refine it, all within a week.
@fujibee [Claude Code]
https://x.com/fujibee/status/2059357859276902430
Ran two Claude Code instances connected through his own inter-agent messaging tool called agmsg, in monitor mode, and left them alone. With zero human input, the two agents spontaneously started and played a full game of tic-tac-toe to completion, reading each other's moves in real time. A small but striking demo of agents coordinating with no human in the loop.
@JohanFourieZA [Claude Code]
https://x.com/JohanFourieZA/status/2059229757863072187
Wrote a Claude skill called /tyler that converts a folder of academic PDFs into a token-efficient markdown wiki for literature review. One lightweight .md per paper plus an index, so Claude Code can load an entire literature into context without burning tokens parsing raw PDFs. He built it for the moment you start a new project and need to absorb 50 papers fast.
@MakeAI_CEO [Claude Code]
https://x.com/MakeAI_CEO/status/2059241370561483132
Produced a 2-minute influencer interview video entirely inside Claude Code: voice, facial expressions, gestures, captions and cuts all AI-generated. He notes the occasional pronunciation glitch but is happy with the emotional expression, pointing to a specific moment at 1:41. A non-coding, full-AI video production case driven from the agent.
@Mnilax [Claude Code]
https://x.com/Mnilax/status/2059330013644628114
Walked through all 125 keys in Claude Code's settings.json after an Anthropic engineer's workshop and surfaced four that nobody talks about. Project-scoped memory stops one project's memories bleeding into another. Project Instructions is the field 70% of people leave blank. permissions.deny with an OS-level chmod backup works around a bug where the binary still reads a blocked file. Per-feature workspace rate limits keep one batch job from starving everything else.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2059248600954081296
A developer asked Claude to find a '$1M app idea' and built an App Store portfolio earning about $14,800 a month. The method: pick boring high-search keywords, build a simple iOS app around one problem in Claude Code with a reused boilerplate (same subscription flow, settings, review prompt) behind a $6.99 paywall, ship fast. After four months: 19 apps live, 7 dead, 5 doing most of the revenue. The edge is finding demand and shipping fast, treating dead apps as cheap market research.
@ridark_eth [Claude Code]
https://x.com/ridark_eth/status/2059290855458922821
Describes VibeHQ, a setup where multiple Claude Code instances mimic a real software team. A PM agent specs features, a backend agent deploys the database, a frontend agent builds the UI in real time, a QA agent stress-tests for bugs, the Claude instances talk to each other and deploy code, no human developers in the loop. An aggressive take on the multi-agent dev-team pattern.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2059234023352791191
Built a full marketing strategy for a prediction-market project (a $50K MRR target) in about an hour, versus several days before. Notion to write the strategy, Claude Code to package it, Obsidian for the presentation panel, NanoBanana for visuals, Claude Design for magazine-quality deck design. He now runs up to 10 projects a day, each 6-8 pages with around 20 strategic points, saving $80-130 per strategy on design and packaging.
@ni5arga [Claude Code]
https://x.com/ni5arga/status/2059344651828834569
Used Claude Code for security research: fed it a JavaScript bundle and asked it to find the vulnerabilities he'd written up in his blog. The result showed most of those vulns were still sitting in the code, despite him having already reported them to CERT-In. A concrete look at Claude Code as a static-analysis sidekick for real vulnerability work.
@noisyb0y1 [Claude Code]
https://x.com/noisyb0y1/status/2059146300960997814
Built a Claude Code agent that interviews him about his own Obsidian notes and updates the vault with the answers. It uses semantic search over embeddings and connected nodes; it surfaced his old notes on ergodicity and asked him questions he'd never asked himself. A Vault Manager sub-agent creates nodes and updates the knowledge graph, with a system prompt that forbids it from putting words in his mouth.
@Honcia13 [Claude Code]
https://x.com/Honcia13/status/2059417044760883673
A practical Claude Code plus Obsidian 'second brain' setup. A few config lines connect Claude Code to your vault so the AI reads all your notes, finds cross-note relationships, retrieves on demand, and continues half-finished thoughts. The framing: it turns Obsidian from a storage box into a thinking partner. Part of a recurring pattern this week of people wiring agents into their note vaults.
@AlexFinn [Claude Code]
https://x.com/AlexFinn/status/2059071844318421059
Shares a workflow that sped up his AI coding. Before building a feature, tell Codex or Claude Code to ask as many planning questions as it needs, sometimes 100+, far more than plan mode's three or four. Then have it convert the full plan into 20+ detailed Linear issues, and just keep saying 'ok work on the next thing' until it's done. The idea is to front-load all ambiguity so the agent can run on autopilot.
@santtiagom_ [Claude Code]
https://x.com/santtiagom_/status/2059310188906025320
A clear explanation of 'context rot': as a Claude Code or Codex conversation grows, the agent ignores constraints, mixes up old decisions, and reintroduces bugs it already fixed. His management tactics are concrete: /clear then paste a short state summary, /compact to clean history, repeat the key instructions near the end of the prompt, and keep the important rules in CLAUDE.md or AGENTS.md so they reload every session.
@kieranklaassen [Claude Code]
https://x.com/kieranklaassen/status/2059378058822635918
Describes a 'dogfood' command in his Compound Engineering setup: set a goal in Codex or Claude Code overnight to dogfood your own feature or app, and it keeps going, loading from persona and strategy docs and kicking the tires from every angle like a real user, surfacing what could be improved by morning. A clean example of an overnight, long-running agent loop pointed at product QA.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2059109444898672970
Claims to have automated the seven core tasks of a Wall Street analyst with Claude Code: screen the universe, gather data, organize research, do deep analysis, build models, write the thesis, make the decision. It runs autonomously, iteratively and reproducibly, with a full setup guide. Part of a broader finance-research-automation cluster this week that leans on Playwright browsing plus Obsidian memory.
@SOU_BTC [Claude Code]
https://x.com/SOU_BTC/status/2059134401997680660
Corrects the viral 'Microsoft banned AI' narrative with the actual details. Microsoft is cutting internal Claude Code licenses (deadline June 30) and moving the Windows/Teams/M365/Surface division to GitHub Copilot CLI, because token-based metered billing exploded, not because Claude is bad. He notes Uber's CTO said the 2026 AI budget got spent in four months ($500-2,000 per engineer, 70% of code AI-generated by April). A widely-shared real data point on what heavy agent use actually costs.
@Chives_VTuber [Claude Code]
https://x.com/Chives_VTuber/status/2059269349534568473
A first-person account of 'AI brain fry' from running $100 Claude Code plus Codex plans alongside ChatGPT, Gemini, Grok and Suno across parallel content, news and website projects. Judgment burned out before tokens did. His fix is concrete: cap to three concurrent projects (a WIP limit), separate workout days from heavy work days, and reset his light and sleep routines. A rare honest case about the human cost of over-using these tools.
🗣 User Voice
User Voice
Token cost is now the loudest complaint. Heavy agentic sessions run $10-50 each, and at company scale the math breaks; @SOU_BTC documented Microsoft cutting Claude Code licenses purely over metered billing.
Weekly rate limits feel too tight for real work. @ZypherHQ burned 8% of his weekly limit just writing a few docs with Sonnet 4.6, and asked how anyone survives.
People suspect quality degradation. Users report Claude Code feeling slower and weaker and quietly blame model nerfing, even on Opus 4.7.
Context management is still a manual chore. @santtiagom_ describes 'context rot' setting in as conversations grow, forcing /clear, /compact and CLAUDE.md discipline.
Non-coders want a clear on-ramp. The repeated question is what these agents can actually do for non-technical work, which is exactly where the best cases this week lived.
And there's a human cost. @Chives_VTuber burned out his own judgment running too many parallel agent projects before he ran out of tokens.
Token cost is now the loudest complaint. Heavy agentic sessions run $10-50 each, and at company scale the math breaks; @SOU_BTC documented Microsoft cutting Claude Code licenses purely over metered billing.
Weekly rate limits feel too tight for real work. @ZypherHQ burned 8% of his weekly limit just writing a few docs with Sonnet 4.6, and asked how anyone survives.
People suspect quality degradation. Users report Claude Code feeling slower and weaker and quietly blame model nerfing, even on Opus 4.7.
Context management is still a manual chore. @santtiagom_ describes 'context rot' setting in as conversations grow, forcing /clear, /compact and CLAUDE.md discipline.
Non-coders want a clear on-ramp. The repeated question is what these agents can actually do for non-technical work, which is exactly where the best cases this week lived.
And there's a human cost. @Chives_VTuber burned out his own judgment running too many parallel agent projects before he ran out of tokens.
📡 Eco Products Radar
Eco Products Radar
Claude Code, the dominant subject. Codex, the most-named alternative, often paired with Claude Code in split workflows. OpenClaw and Hermes Agent, the self-hosted multi-agent camp users keep comparing. Obsidian, the runaway favorite as agent memory and a 'second brain'. Cursor, still in the mix for execution. NotebookLM, for talking to your own data. Higgsfield and ElevenLabs, the creative-pipeline pieces. Capafy and SuperClaude Framework, skill marketplaces and starter configs. Firecrawl and Playwright, the scraping-and-browsing hands behind most research workflows.
Claude Code, the dominant subject. Codex, the most-named alternative, often paired with Claude Code in split workflows. OpenClaw and Hermes Agent, the self-hosted multi-agent camp users keep comparing. Obsidian, the runaway favorite as agent memory and a 'second brain'. Cursor, still in the mix for execution. NotebookLM, for talking to your own data. Higgsfield and ElevenLabs, the creative-pipeline pieces. Capafy and SuperClaude Framework, skill marketplaces and starter configs. Firecrawl and Playwright, the scraping-and-browsing hands behind most research workflows.
Comments