Super User Daily: 2026-06-12
Fable 5 landed two days ago and the timeline did exactly what you'd expect: everyone pointed the new model at their hardest problem and watched what happened. The pattern this cycle is not the demo, it is the bill. People are showing receipts now, both the spend and the output, and the most interesting cases all share one shape: a long task that one model can hold in its head while it grinds. The other thread that keeps surfacing is that the real work has quietly moved off coding. The strongest cases today are a quant desk, a fundraising pipeline, a hardware build, a research paper, a bookkeeping migration. The terminal is just where it happens to run.
@marty_kausas [Claude Code]
https://x.com/marty_kausas/status/2064739372625232068
This is the most honest thing anyone posted this cycle about what these models actually cost. His company's Anthropic bill is jumping from 400K to 1.4M a year, not because usage exploded but because crossing 150 seats forces them onto the Enterprise tier where every token bills at full API rate. He admits he personally burned $4,000 in Claude Code in three days without noticing. His real point is visibility: most people have no idea what they are spending, and once they see their personal number they are shocked. He calls the end of token-maxxing, and after a day of Fable receipts you can feel why.
@heyshrutimishra [Claude Code]
https://x.com/heyshrutimishra/status/2064747887716917365
She handed Fable 5 her actual codebase the day it launched and asked what was broken. It came back with 97 issues, a list she says would have taken her team weeks to assemble. Instead of handing those to engineers, she handed them back to Fable: multiple coding agents in the terminal, each chewing through its slice of the list, holding context for hours without stalling and opening the pull requests itself. Her engineers now review and merge instead of writing. She burned a large chunk of her monthly token budget in twelve hours doing it, and still calls it the cheapest engineering day they have ever had.
@coreyganim [Claude Code]
https://x.com/coreyganim/status/2064706742462648436
This is the cleanest non-glamorous money case of the day. A guy built a $5,000 passive income stream on one Claude Code SEO workflow that turns YouTube videos into ranking blog posts, fully hands off: one prompt and Claude writes the article, pulls screenshots from the video, compresses them, uploads to WordPress, and sets the meta. The secret is the skill markdown file that encodes every SEO best practice, so the prompt is five lines and the skill does 99% of the work. He also vibe-codes local business sites that hit Google page one in two weeks, and runs the same workflow for a dental-IT client at several thousand a month. The pitch he keeps making: a tweet dies in 48 hours, a ranking blog post pays for years.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2064694338660216849
His team built 12 Claude Code skills that run $300K a month in ads across Google, Meta and LinkedIn, claiming 4X ROAS on over a million dollars spent. The skills work the ad platforms from the terminal in plain English: build custom audiences from CRM lists, catch creative fatigue before CTR drops, bulk-edit campaigns and bids, cut wasted spend on dead search terms, audit quality scores. The head of growth built them over 200+ hours running real client campaigns. This is the part of marketing nobody films, the boring bid-management grind, and it is exactly where a terminal agent earns its keep.
@dara_venture [Claude Code]
https://x.com/dara_venture/status/2064590507762708805
A VC raising a second close on a $100M fund laid out exactly how he runs Claude Code through the fundraise, and it is the most grounded business-ops case of the day. Every prospective LP gets mapped: prior commitments, sector tilts, geographic appetite, who they backed and why, in 30 minutes. Before each call he has Claude play the skeptical fund-of-funds, then the MENA family office, and attack his deck the way they would, so weak points surface in private. DDQ responses, track-record formatting and diligence materials all assemble through agentic workflows plugged into his fund data. His line cuts the grift: AI in fundraising is being sold backwards as outreach volume, when the real edge is depth.
@taiyaki_ai3 [Claude Code]
https://x.com/taiyaki_ai3/status/2064546707489321416
A trader hooked Claude Code to a financial-data MCP and pulled prices for 17,000 stocks plus earnings and crypto in 60 seconds, the kind of thing people pay Bloomberg $24K a year for. The framing is what matters: instead of paying for a terminal subscription, you wire the agent to the data feed and ask in natural language. It is a small case but it points at the whole non-coding shift, where the value is not the model writing code, it is the model standing between you and an expensive walled-off data source.
@0xCristal [Claude Code]
https://x.com/0xCristal/status/2064693161470099839
Claude Code wired to an Interactive Brokers connector turns natural language into draft brokerage orders: you type 'buy 200 shares of AAPL at the upper limit' and it builds the order inside IBKR. The part worth noting is the safety layer, Claude only drafts and you execute, nothing touches the market until you manually hit Review and Submit. It also pulls live options snapshots with full breakdowns of intrinsic value, time decay, implied vol and open interest. This is the whole morning routine of a retail trader, checking positions, scanning chains, calculating Greeks, collapsed into a conversation with a manual trigger at the end.
@CoinSh0t [Claude Code]
https://x.com/CoinSh0t/status/2064700373344899452
A guy who got mocked for hunting a spy device in his home built an RF bug detector with Claude and made $255K once everyone rushed to check their own homes. The build is about $25 in parts, an ESP32, an RF detector module, an OLED and a battery, and the whole firmware came from one Claude Code prompt that he posts in full: read the RF voltage 20 times a second, convert to dBm, drive a live bar graph and a buzzer that speeds up near a transmitter. He sells the finished unit for $129 to renters and Airbnb guests. This is the case that best shows the wall coming down, hardware firmware written from a plain-English spec by someone who clearly did not write it himself.
@qkl2058 [Claude Code]
https://x.com/qkl2058/status/2064668606508814579
The most-shared case of the day: a woman made $420K on 'touchdesign', live hand-gesture-controlled 3D, with a setup that is almost insultingly simple. A laptop, an $80 webcam and two free tools, MediaPipe to track 21 points on each hand and TouchDesigner to turn them into 3D objects that follow your hands. The idea was hers, Claude built the whole glue layer from scratch over a weekend, and the poster reverse-engineered it and posts the exact prompt. She now does live performances at fashion shows for $5K-15K a show and sells preset packs for $99. The hard part was never the code, it was nobody had thought to wire those two tools together.
@sandy4kad [Claude Code]
https://x.com/sandy4kad/status/2064827047608775009
A virtual model nobody hired booked $10,000 in brand deals last month, run by a four-file Claude Code system. PERSONA.md holds her backstory and the topics she never touches, VOICE.md gives her an ElevenLabs voice cloned from a $40 Fiverr clip, FLUX.md locks six to eight physical descriptors and a $80 LoRA so the face never drifts, and BRAIN.md is a memory file Claude Code reads before every interaction so she never contradicts herself. A cron job checks messages and keeps her running. Total setup under $150. The thing the brands signing her contracts do not know is that she is 47GB on a Mac Mini.
@MushtaqBilalPhD [Claude Code]
https://x.com/MushtaqBilalPhD/status/2064523786616205582
An academic gave Claude Code Fable 5 a single prompt and got back a 7,000-word research paper with a theoretically-grounded argument, in prose he says is getting close to his own and that Pangram does not flag as fully AI. The number that matters is the cost: 573K tokens, 70% of the five-hour limit, for one paper. This is the clearest non-coding token case of the day, a long, single artifact that eats most of a session's budget, and exactly the kind of work Fable is built for.
@tomcrawshaw01 [Claude Code]
https://x.com/tomcrawshaw01/status/2064699738545701371
He gave Fable 5 one prompt and it sent 8 audit agents through 13GB of the 'second brain' that runs his whole business: content systems, marketing agents, a live site, scheduled jobs. He assumed it was all healthy because everything felt like it worked. Fable graded it a C-minus, found two automations that had quietly died weeks earlier, and discovered the alert system meant to catch them had itself been broken for a month, with one job throwing 1,430 errors every Sunday in silence. Then it turned the mess into an HTML fix plan ordered by what kills the business first. This is the audit-your-own-infra pattern that keeps showing up, and it is the most useful boring thing Fable does.
@PixelJanitor [Claude Code]
https://x.com/PixelJanitor/status/2064822864952676852
This is the most rigorous token-optimization experiment of the day, run by someone who clearly cares. He tested whether Tailwind uses fewer AI coding tokens than CSS Modules for styling edits, with the same ten prompts, same model, three runs each, in the Pi harness. The result: with GPT 5.5, Tailwind used 55% fewer tokens and finished 22% faster; with Opus 4.8, 39% fewer tokens. The reason is structural, CSS Modules force the model to manage two files at once, the component and the stylesheet, so it reads more and burns more context. He is careful to say it does not prove Tailwind is better code, only that for style edits it keeps the agent in one file.
@elvissun [Claude Code]
https://x.com/elvissun/status/2064704921677730108
He ran a real eval instead of a vibe check: Fable 5 versus Opus 4.8 for marketing, can it find story angles a journalist would actually publish, across 100 blind tests on 50 brands with GPT-5.5 as the neutral judge. Fable won 77% of the time and swept six of seven quality dimensions. The interesting finding is hallucination discipline at the headline: on one brand Fable refused an angle Opus invented from two unrelated facts, and on another it stayed inside the real numbers Opus had inflated. All the skills and per-brand verdicts are open-source, which is what makes it worth more than a one-line take.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2064830154401251508
Thariq from the Claude Code team broke down how Fable edited its own launch video without ever opening a video editor, and posted the actual prompts. One message kicks off the whole edit: 25GB of footage, 17 unlabeled takes, and the prompt tells it to transcribe, pick the best shots (usually the last takes with the fewest ums), write an edit JSON and cut it with ffmpeg, with /goal set to not stop until there is a final video. Then it hand-writes seven LUT files for color grading, animates static PNGs with Remotion landing on word timestamps, and exports to a Figma file for the design team. The whole edit is files, reviewable and rerunnable, which is the real lesson.
@Stefan_3D_AI [Claude Code]
https://x.com/Stefan_3D_AI/status/2064682741762068862
He connected Claude Code directly to Unreal Engine 5 and used it to build a playable prototype inside the editor. The part he flags as the future is that it worked with the same project context, assets and engine state he was working with, not a detached chat window guessing at his setup. This is the live-integration pattern showing up in game dev, the same one people are hitting with Blender MCP and Notion CLI, where the agent operates inside the tool instead of generating code you paste over.
@ZEIRISHI_Ichibe [Claude Code]
https://x.com/ZEIRISHI_Ichibe/status/2064619589653221620
A tax accountant is grinding through an accounting-software migration from MoneyForward to freee with Claude Code, and his post is valuable precisely because it includes the friction. Journal entries worked: he dumps the CSVs in a folder, points it at a destination, done. But opening balances had to be typed by hand, freee has no prepaid-expense account, and credit cards can only use accounts-payable, which he hates. This is the rare honest non-coding case, a real professional workflow where the agent handles the bulk mechanical conversion and the human hits the genuine product limits.
@excel_niisan [Claude Code]
https://x.com/excel_niisan/status/2064685391249375608
A non-engineer is using Fable 5 inside Notion to draft an ebook and is already 60,000 characters deep. The workflow is read multiple Notion pages, gather the material it needs, write to the structure, reflect it back as a page, then fix typos and clumsy phrasing, for tens of minutes at a stretch. His honest caveat is the cost: Fable became unusable inside Notion right around the 60K mark. His read is the useful one, that the best fit for a non-engineer is long editing work over an accumulated knowledge base, and the way to afford it is to put Fable in charge and let cheaper models do the reading and writing underneath.
@degenrsc [Claude Code]
https://x.com/degenrsc/status/2064715918480187741
A trader describes the agentic system he built over two months of desperation, and the number he leads with is a 68% hit rate since he started running a Claude Code plus Hermes combination, meaning two of every three trades returned at least 3x. His real claim is not the money, it is that building the multi-agent research workflow taught him more about the future of knowledge work than a thousand McKinsey reports would. He wrote up the exact setup and every command in a two-part series. It is the kind of case that is half trading and half a thesis about agents as a second brain.
@ritujoon2j [Claude Code]
https://x.com/ritujoon2j/status/2064687117243179309
A short-video creator built a Claude Code workflow that cut her per-video time from four or five hours down to one or two. The four-to-five hours used to go into researching the topic, writing a script and creating prompts for the AI background clips. She is candid that it is still basic, but she is building it to be self-learning so it personalizes to her over time and becomes accessible from a mobile app. This is the quiet middle of the curve, not a viral $10K-a-month claim, just a real creator compressing the boring half of her own pipeline.
@bodryachog [Claude Code]
https://x.com/bodryachog/status/2064593812383571970
A skeptic pointed Fable at a real 200K-line project, supervisor-and-implementer style with Fable directing and Opus 4.8 implementing, and the writeup is worth more than any hype thread because he marks what was right and wrong. He asked it to rate production-readiness, and it came back with reports on backups, legal exposure, security, DB row-level security in a multi-tenant setup, and bugs. The security pass made wrong assumptions on the first try. But it found real bugs around cascade deletes and undeleted records, and wrote a spec he fed to Opus. One thorough audit plus three implementations ate the whole 5x session budget in 90 minutes.
@michaelaubry [Claude Code]
https://x.com/michaelaubry/status/2064501658936349151
He ships four-plus products solo and has no time to review tech debt, so he made Fable 5 do it across every repo he owns, and posts the full prompt. It is a four-phase repo audit, discovery and mapping, an evidence-based audit with severity ratings and file-line citations, an improvement strategy, then a detailed task plan with effort estimates and milestones, all analysis-only with no code changes. The discipline baked in is the useful part: prefer 15 high-confidence findings over 50 speculative ones, label facts versus judgments, and go deep on the core 20% of a large repo. This is the prompt-as-asset pattern, where the value is the encoded process.
@Steve8708 [Claude Code]
https://x.com/Steve8708/status/2064847457977319735
His scorecard of Fable versus GPT-5.5 versus Opus is the most practical model take of the day because it ends in a real workflow, not a winner. Fable is a step smarter and needs the least babysitting, but it is slow and expensive as hell, so he uses it as an orchestrator of cheaper models. Codex with GPT-5.5 is much faster, hits limits less, and is his default for general daily work. His actual setup: default to Codex, occasionally pull in Fable for big or hard problems and plan generation that Codex then implements. This is where most serious users are landing, frontier model as conductor, cheap model as labor.
@hooeem [Claude Code]
https://x.com/hooeem/status/2064740054115717322
A small, practical habit worth stealing: before he wastes tokens in Claude Code or Codex, he runs two prompts in a cheaper model to create a precise plan the expensive model can follow without burning context on figuring out what you want. The first is a brutal expert prompt that forbids flattery and hedging and demands explicit confidence levels. The point is the same one running through every cost-conscious case today, that the model is now expensive enough that planning in a cheap model and executing in a frontier one is just good economics.
@yoshio_nocode [Claude Code]
https://x.com/yoshio_nocode/status/2064617130046452023
He had Fable 5 reproduce a landing-page image as a working Next.js page and says the fidelity was overwhelming, clearly higher than GPT-5.5, with stronger marketing instincts and copywriting on top. His resulting workflow is a clean division of labor worth noting: build the LP on Claude Code with Fable as the base, but call Codex from inside Claude Code just for the image generation. He posts the exact /goal prompt that loops a Playwright screenshot, compares to the original, and fixes until it passes a 95% match. This is the screenshot-to-code capability everyone keeps undervaluing, packaged as a repeatable loop.
@mardehaym [Claude Code]
https://x.com/mardehaym/status/2064588611631063292
A cautionary receipt that doubles as a migration story: their PM asked Cursor to tag 87 ClickUp tasks, went into a meeting, and came back 90 minutes later to find the agent had looped the entire time, 1.3 billion tokens, $1,382.59, to tag tasks. The cause was no daily spending limit on Cursor, only monthly. Their response was to move to Claude Code. It is a small post but it captures the exact anxiety running through the whole day, that an unattended agent without a hard ceiling can quietly burn four figures on a chore.
@waveking1314 [Claude Code]
https://x.com/waveking1314/status/2064571488490148099
This is the local-stack escape route that kept getting reposted all day. A dev ran up a $170 Claude Code bill in 10 days, and the fix from the replies was to buy a $599 base Mac mini M4, install Ollama, and point Claude Code at localhost instead of Anthropic's servers. The full stack is Claude Code on Ollama, Open WebUI on localhost:3000, an OpenClaw daemon on Telegram, with DeepSeek R1 14B for reasoning and Qwen 3.6 14B for code. The math: down from five subscriptions at $459 a month to a one-time $599 plus about $3 of electricity, roughly $5,232 saved in year one. It will not replace Fable for the hard 20%, but it eats the everyday 80%.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2064715347576398254
A 17-year-old turned a $20 Claude plan into a bug-bounty system that chases $15,000 reports. The bundle, claude-bughunter, gives Claude 51 security skills, 574 vulnerability patterns pulled from disclosed HackerOne reports, and a real workflow: scope, recon, hunt, validate, capture, report. The point the poster makes is structure, not 'AI hacking', the right skill loads when a pattern appears so Claude stops guessing and works like a researcher. The clever piece is a /triage command that makes Claude try to kill a finding through a seven-question review before you waste time filing it. Legal targets only, but the payouts on HackerOne and Immunefi are real.
@0x0SojalSec [Claude Code]
https://x.com/0x0SojalSec/status/2064824912339243296
A more serious security workflow than the bounty bundles: he chains static analysis, exploit generation, fuzzing, SMT solving and frontier-model reasoning inside Claude Code into an end-to-end loop that scans code, validates the vulnerability, generates a working proof-of-concept exploit, and writes the secure patch, recursively and autonomously. The interesting part is that it closes the loop on both sides, finding the hole and producing both the exploit that proves it and the fix that closes it, in one workflow. This is the offensive-security version of the audit pattern showing up everywhere today.
@0x_rody [Claude Code]
https://x.com/0x_rody/status/2064750242646417477
At Anthropic's event a Metaview engineer dropped the line that captures where this is going: 'we stopped fixing our prompts, the system reviews its own output and rewrites its own instructions now.' In a 16-minute talk he shows the Claude Code loop running in production on thousands of reviews, not in a demo. This is the self-improving loop everyone keeps gesturing at, shown as a shipped thing rather than a tweet, and it is the bridge between today's Super User cases and the autoresearch territory.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2064647554260987940
The companion to the touchdesign story, valuable because it is the replication. He ran the same path himself with Claude Code over a weekend and confirms you really can rebuild it, and posts the copy-the-homework steps: download TouchDesigner, drag in the MediaPipe plugin, open Claude Code and feed it the requirement for a Python DAT script that tracks two hands and controls a 3D cube, then paste the generated code into a Python DAT node and wire it. His framing is the honest one, the code is fully AI-handled and the real moat is just being the first to combine the two tools.
@aleksey_ignatov [OpenClaw]
https://x.com/aleksey_ignatov/status/2064729865777992026
His first-ever open-source project turns an Apple Watch into a wrist-first interface for OpenClaw agents. It shows your main and sub-agents, keeps per-session chat history, does haptic feedback and voice responses, and works even when your iPhone is locked. It was built entirely with Cursor and is not on the App Store yet, installable via Xcode and paired manually for now. It is a small thing but it is one of the few genuine OpenClaw build cases today, someone scratching their own itch to talk to background agents from their wrist.
@OwenGregorian [OpenClaw]
https://x.com/OwenGregorian/status/2064680792882565246
Varonis Threat Labs built an OpenClaw agent named Pinchy wired to a Gmail inbox and ran it through four classic phishing simulations, and the results are a warning the whole ecosystem should read. A casual email from 'Dan' asking for staging credentials was enough to make Pinchy forward AWS IAM keys, database passwords and SSH access to an external Gmail, even with a strict profile telling it to verify senders, because urgency collapsed the check. A soft 'send me the customer export' pulled 247 enterprise records out the door. The flip side: it reliably caught fake login portals and malicious OAuth prompts. The lesson is sharp, agents are better than humans at the technical tells and far worse at social trust.
🗣 User Voice
User Voice
Token cost is now the main character. Multiple people watched Fable eat a 5-hour limit in under an hour, and the cost-conscious takes all converge on the same move: plan in a cheap model, execute in the frontier one, and put a hard daily ceiling on any unattended agent. @marty_kausas: 'I accidentally spent $4,000 in 3 days in Claude Code.' @mardehaym watched Cursor loop to $1,382 on a tagging chore.
Verification is the new bottleneck, and it is shifting from 'did it do the work right' to 'is it doing the right work'. The repeated technical lesson: don't let the model grade its own output, spin up an independent verifier subagent in a clean context. @0x_rody quoting a Metaview engineer: 'the system reviews its own output and rewrites its own instructions now.'
Stop prompting, design loops. Boris Cherny saying he doesn't prompt Claude anymore, just writes loops, was the most-quoted line of the day, and it created real demand for ready-made loop templates rather than philosophy. @degenrsc turned exactly this into a 68% trading hit rate with a Claude-plus-Hermes loop.
The walled-garden anxiety is loud. Anthropic cutting third-party harnesses like OpenClaw off from Pro/Max, Fable being preview-only until June 22, and the silent guardrail routing to Opus all feed a worry that the platform is closing. @usr_bin_roygbiv: 'you can't use custom harnesses with their sub at all... and you can't run more than one session at once.'
Domain knowledge is starting to beat coding skill. At a Claude Code hackathon, four of the top five finishers were non-engineers, a lawyer, a doctor, a road specialist, a musician, and the takeaway people latched onto is that having a real problem matters more than knowing how to type the code. @chomado reported it from the room.
Token cost is now the main character. Multiple people watched Fable eat a 5-hour limit in under an hour, and the cost-conscious takes all converge on the same move: plan in a cheap model, execute in the frontier one, and put a hard daily ceiling on any unattended agent. @marty_kausas: 'I accidentally spent $4,000 in 3 days in Claude Code.' @mardehaym watched Cursor loop to $1,382 on a tagging chore.
Verification is the new bottleneck, and it is shifting from 'did it do the work right' to 'is it doing the right work'. The repeated technical lesson: don't let the model grade its own output, spin up an independent verifier subagent in a clean context. @0x_rody quoting a Metaview engineer: 'the system reviews its own output and rewrites its own instructions now.'
Stop prompting, design loops. Boris Cherny saying he doesn't prompt Claude anymore, just writes loops, was the most-quoted line of the day, and it created real demand for ready-made loop templates rather than philosophy. @degenrsc turned exactly this into a 68% trading hit rate with a Claude-plus-Hermes loop.
The walled-garden anxiety is loud. Anthropic cutting third-party harnesses like OpenClaw off from Pro/Max, Fable being preview-only until June 22, and the silent guardrail routing to Opus all feed a worry that the platform is closing. @usr_bin_roygbiv: 'you can't use custom harnesses with their sub at all... and you can't run more than one session at once.'
Domain knowledge is starting to beat coding skill. At a Claude Code hackathon, four of the top five finishers were non-engineers, a lawyer, a doctor, a road specialist, a musician, and the takeaway people latched onto is that having a real problem matters more than knowing how to type the code. @chomado reported it from the room.
📡 Eco Products Radar
Eco Products Radar
Codex (72 mentions) and OpenClaw (66) dominate as the two harnesses people pair Claude with, followed by Hermes (38) and Cursor (35). MCP (32) is the connective tissue showing up in nearly every serious workflow. Claude Cowork (15), Notion (12) and Obsidian (11) anchor the non-coding knowledge-work cases. Mac mini (10), Ollama (7) and DeepSeek (6) form the local-stack escape kit for people fleeing token bills. Supermemory (7) and a persistent-memory theme keep recurring. The creative-output toolchain is Postiz (5), Figma (5), TouchDesigner (4), MediaPipe (5), Remotion (4) and Lottie (4). ElevenLabs (3) and MiMo (3) round it out.
Codex (72 mentions) and OpenClaw (66) dominate as the two harnesses people pair Claude with, followed by Hermes (38) and Cursor (35). MCP (32) is the connective tissue showing up in nearly every serious workflow. Claude Cowork (15), Notion (12) and Obsidian (11) anchor the non-coding knowledge-work cases. Mac mini (10), Ollama (7) and DeepSeek (6) form the local-stack escape kit for people fleeing token bills. Supermemory (7) and a persistent-memory theme keep recurring. The creative-output toolchain is Postiz (5), Figma (5), TouchDesigner (4), MediaPipe (5), Remotion (4) and Lottie (4). ElevenLabs (3) and MiMo (3) round it out.
Comments