Super User Daily: 2026-06-15
Today the feed splits cleanly between people getting real leverage and people getting laid off for it. Fable 5's brief window drove a wave of heavy, deliberate usage, half-a-billion-token build sprints, nine-pane parallel sessions, and honest reviews of where the top model actually unlocked the impossible, while the most-shared cases moved off coding entirely: bookkeeping run in the background, autonomous content businesses, lead inboxes triaged in your own voice, and Claude Code installed as a paid service for small businesses. The other half is anxiety, about token cost, about memory that rots, and about being the engineer who taught the whole team to replace him.
@ABesarovic [Claude Code]
https://x.com/ABesarovic/status/2065717812643959291
Built an AI agent in 30 minutes with Claude Code that triages and answers his team's inbound LinkedIn campaign replies. It only handles people who actually responded, skips spam and "not interested," and flags the hot buyers, then drafts a reply in each teammate's own voice using the full thread plus the lead's profile. The stack is an Aimfox webhook on new replies, Claude Opus drafting, and a Slack card with approve/edit/skip before anything goes back to LinkedIn. They now open Slack to about 50 ready-to-send drafts and approve the good ones in seconds, with a human always the last step.
@maestroslay [Claude Code]
https://x.com/maestroslay/status/2065887796204593661
Rebuilt a faceless financial-niche YouTube style (Microsoft-Paint stick figures) with Claude Code plus Higgsfield, no editing skills. He copied a title from a video doing 1.7M views, ran it through an AI script writer, transcribed the voiceover for timestamps, then gave Claude Code one master prompt: one deliberately-ugly paint image per timestamp. Claude Code wrote 98 image prompts, generated every frame through Nano Banana Pro for $7.84, and compiled them against the voiceover into a finished 10-minute video, voiceover synced, zero manual edits. He breaks down the full math: about $258/month to run it as a working YouTube business.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2065611863505920097
Produces vertical short-form video entirely inside Claude Code (on Fable 5) with no video editor at all. Instead of editing he "declares" the edit: text, color and animation are written into a script and re-running fixes them instantly. Narration is a clone of his own voice with scene durations auto-set to match the audio; subtitles, self-made numpy BGM and Slack/GitHub-style UI screens are all generated with Python, and the background is his own screen-recording sped up 30x and composited with ffmpeg. The video that introduces "tools worth connecting to Claude Code" was itself made with Claude Code plus those connected tools.
@helloyuki_ [Claude Code]
https://x.com/helloyuki_/status/2065666578759458905
Shares a concrete shift toward parallel Claude Code usage built on git worktrees and tmux. He says he needs to move to running several Claude Code sessions in parallel and is reworking his tmux setup first, highlighting a guide on Claude Code x tmux. The detail he most wanted was launching a file-picker from a tmux popup. A small but real workflow note from someone moving past single-session usage into managed parallelism.
@AlchainHust [Claude Code]
https://x.com/AlchainHust/status/2065685246167667139
During the 72 hours of Fable 5 access he spent almost the whole time iterating on an agent container called FanBox, burning 500M+ tokens in three days and shipping 20+ versions. He calls it Fable 5's parting gift, praising how natural and native it makes fusing agents with any terminal to boost Claude Code, Codex and Kimi Code. After Fable 5 was cut off he kept fixing legacy bugs and adding features with Opus 4.8 and K2.7. A concrete heavy-usage build log with real token volume and version count.
@jinglian [Claude Code]
https://x.com/jinglian/status/2065593828292100444
A pointed workplace story: a programmer earning 38k/month started using Claude Code late last year and maxed out his delivery, so his boss had him train the whole team over three months. Team efficiency doubled, the company realized the work he used to backstop could now be done by anyone, and they laid him off for being too expensive. The lesson he draws: the most endangered engineer is the one who only uses AI to write code, because writing code is getting cheap. Real value lives in understanding the business and translating needs into systems AI can execute.
@Suzacque [Claude Code]
https://x.com/Suzacque/status/2065593597626622233
A three-day Fable 5 log: a model that impresses more the more you use it, so he bought a Max 20x subscription just for the ten days it stays available. He describes being walked up Anthropic's pricing ladder day by day (Pro, then Max 5x, then Max 20x) and running Claude Code in parallel across nine panes on a large Mac screen. His best-use finding: consult it on problems you currently face and feed it things GPT-5.5 pro and xhigh couldn't solve, while saving the still-unsolvable for future models.
@ScottyBeamIO [Claude Code]
https://x.com/ScottyBeamIO/status/2065895184530133288
Reports a 22-year-old student running a fully autonomous AI OnlyFans persona making $43,000+/month with no camera, filming, editing or human chatting. The stack: Claude Code writes every message, manages conversations and runs the content workflow; Flux generates every photo and thumbnail; ElevenLabs produces a consistent voice. He cites average revenue per subscriber around $34 and top fans spending nearly $2,000 on messages. A concrete non-coding revenue case with Claude Code as the conversation-and-workflow engine.
@DennisonBertram [Claude Code]
https://x.com/DennisonBertram/status/2065879414953455911
Shares a cost hack: run DeepSeek inside the Claude Code harness and it can still build and drive workflows at a fraction of Opus 4.6/4.7 cost and quality. He has it running 250+ subagents in a workflow doing adversarial reviews, "pennies on the dollar," and packaged the setup as a tool called Deep-Claude. A concrete account of swapping the model behind the Claude Code harness to cut spend while running large multi-agent review pipelines.
@insomnia_vip [Claude Code]
https://x.com/insomnia_vip/status/2065795417313325416
Claims one prompt plus Claude Code cloned an award-winning website that agencies charge $10,000 for, using only Higgsfield MCP plus Claude Code. The MCP handles everything automatically: Nano Banana Pro for images, Seedance 2.0 for video, and scroll-frame slicing for the cinematic effect, with no Figma, manual video generation or stitching tools. He puts the total cost per full website with custom images and video at roughly $1 to $2.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2065793686630457527
Automates NotebookLM from Claude Code (and Codex / Google Antigravity) via Skills. The flow throws multiple articles into NotebookLM creating one note per article (17 in this run), generates audio explanations with a custom prompt auto-filled, downloads the audio after a few minutes, converts it to mp3 with a script and files it into a dedicated folder, all driven by Skills instructions, leaving only the Spotify upload manual. A concrete content-pipeline automation built on the agent's skill system.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2065728301612630420
Describes a Beijing developer turning Claude Code into a $6,800/month UI business by feeding it a design system (Moonchild) so it stops inventing random colors, buttons and spacing on every screen. He rebuilt the same 10-screen app twice: without a system only 2/10 screens were on-brand, with Moonchild 9/10 came out clean on the first generation. Concrete numbers: six hours of restyling dropped to 40 minutes and 19 random hex colors dropped to zero, monetized by selling frontend cleanups at $1,500 each.
@Sprytixl [Claude Code]
https://x.com/Sprytixl/status/2065800095212183672
A vivid (if anecdotal) non-coding case: a German accountant plays World of Tanks all day while Claude Code runs the company bookkeeping in the background, reconciling transactions, auditing ledgers and generating financial reports 24/7. Overnight it processes what used to take him a full week, so by morning everything is formatted and ready. His salary reportedly went from $7,000 to $12,000 in two months because output doubled while hours stayed flat, and his boss just sees the most productive accountant in the company.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2065669087196295422
Two data points in one: a Korean PhD student parsed 7,944 Claude Code skills from GitHub and found 33% of them cost more tokens than using no skill at all. The same post uses a long-context model (Kimi, 262k context) to read the entire 500,000-token, 300-page SpaceX S-1 in one view, holding the whole filing at once to surface cross-section contradictions, like Starlink ARPU dropping from $99/mo in 2023 to $66 in Q1 2026, that section-by-section summarizing would miss.
@mtnleonardi [Claude Code]
https://x.com/mtnleonardi/status/2065900699183907022
A small but concrete customization: an anxious user who always wants to see how much quota is left built a custom Claude Code statusline in minutes. After a quick brew install jq, he just asked Claude Code via /statusline for a two-line bar: line one showing model, context percent and token count; line two showing session percentage (current and weekly) plus time remaining. An example of reshaping Claude Code's own UI with one plain-language request.
@gippp69 [Claude Code]
https://x.com/gippp69/status/2065864531553734926
Turns a $9,000 UI handoff into a 3D SaaS interface in about 30 minutes by having Moonchild build the design system first (tokens, spacing, cards, charts, navigation, typography), then letting Claude Code read that structure, logic and components through MCP and turn it into a working React app. A second Claude session watches for drift off the system. The result is 8+ on-brand screens from two tools and one design system, instead of paying an agency to align buttons by hand.
@taiyo_ai_gakuse [Claude Code]
https://x.com/taiyo_ai_gakuse/status/2065682972863520922
Points out a lesser-known trick: even without Fable 5, you can call a slide-template library over MCP inside Claude Code and easily produce Fable-5-quality slides. He demonstrates it at a live event, showing concrete output. A useful note that MCP-connected templates can extend Claude Code into high-quality presentation generation without needing the top model.
@itarutomy [Claude Code]
https://x.com/itarutomy/status/2065708143980495078
Explains and uses COLLEAGUE.SKILL, a departed-colleague knowledge-handoff system (18.5k GitHub stars) that distills a person's judgment rules into Markdown instead of imitating them via a prompt. Feeding it traces (chat logs, design docs, code-review comments) auto-generates two files: work.md capturing expertise, judgment criteria and review priorities, and persona.md capturing communication style and timing. Both install directly into Claude Code or Codex, generate patches from natural-language feedback, and support version control and rollback. A transparent alternative to black-box persona prompts.
@Av1dlive [Claude Code]
https://x.com/Av1dlive/status/2065747876005937416
Highlights claude-obsidian, a free Claude Code plugin that turns the agent into a self-maintaining wiki of your notes, inspired by Karpathy's idea of letting an LLM maintain a linked wiki of your sources. Setup is two commands, then open Obsidian and Claude Code in the same folder and type /wiki. Your notes become queryable by Claude and get richer every time you read something. A concrete non-coding knowledge-compounding application.
@TNGeography [Claude Code]
https://x.com/TNGeography/status/2065665579953037660
A brief but genuine first-person account: for the last two months his office gave software dev teams Claude Code access with a custom IDE, and his working style is now completely different, a 180-degree paradigm shift from two months earlier. He adds a side observation about tech sovereignty, noting his country is unprepared. Short on detail but a real testimony of how team-wide Claude Code adoption changed day-to-day engineering.
@Yuchenj_UW [Claude Code]
https://x.com/Yuchenj_UW/status/2065831626115072263
First-person account of using Omnigent, Databricks' open-source meta-agent for orchestrating a swarm of agents, in production before it was open sourced. He used to run the same task with both Codex and Claude Code then pick the better output; Omnigent instead lets them collaborate, debate and converge. He especially likes real-time collaboration, where you can invite people into a session to watch, steer and send commands, and notes it was built by a lean team in six weeks.
@_xjdr [Claude Code]
https://x.com/_xjdr/status/2065847603968036922
A detailed harness-level model comparison: running Moonshot's K2.7-Code inside Claude Code with his custom inference harness works surprisingly well. Compared to his K2.6 (which he fine-tuned for the harness) and his recent Fable runs, K2.7 stacks up favorably: where K2.6 felt very Opus-like, K2.7 is more terse, more argumentative and overall smarter for his use cases. He praises Moonshot's post-training and says continuing to use K2.7 as-is instead of his K2.6 fine-tune is the highest endorsement he can give.
@maarcoofdezz [Claude Code]
https://x.com/maarcoofdezz/status/2065738693042647488
Moved from loose AI tools to an orchestrated agent team using the open-source self-hosted Paperclip. Before, he opened Claude Code, wrote a prompt, waited, opened another terminal, and lost two hours acting as coordinator between AIs while context exploded. Now he leads a team of 14 agents (a CEO, a CTO and tech team, QA and security engineers), each with a role, budget, limits and objectives, that delegate, wait for reports and review each other's work, with self-improving AI loops set up via one npx command.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2065773939520119040
Argues that builders running only Claude Code get lapped by those running Hermes alongside it, and lays out the split: session-based agents (Claude Code, Cursor, Codex) win focused coding sessions, while persistent agents (Hermes) win everything around coding. He runs four specialized Hermes agents 24/7 on a $5 VPS (Chief of Staff, Head of Research, Head of Content, Head of Finance), each with its own SOUL.md identity and memory that writes its own skills after every task.
@leploutos [OpenClaw]
https://x.com/leploutos/status/2065761496592744740
Made 4,000 euros last month installing Hermes and OpenClaw agents for French-speaking SMBs. His model: charge a setup fee to configure the agent to the client's tools via API, their rules and constraints, then bill for maintenance, updates, monitoring, workflow tweaks and hosting, while the client pays their own LLM/API usage so he carries no token-cost risk. The real value, he says, is adapting the agent to actual workflows, wiring it to Slack, building useful prompts and skills, and making it usable by non-technical staff.
@zagundo [OpenClaw]
https://x.com/zagundo/status/2065727296619835707
Upgraded his OpenClaw memory by installing Supermemory, cutting MEMORY.md from 11,706 to 3,105 characters. The new architecture uses MEMORY.md as a boot primer only (identity, session rules, project pointers) while Supermemory holds the actual semantically-indexed memory, recalled dynamically instead of loading a giant file every context window. Concrete setup: Supermemory runs fully local on port 6767 with its own 106MB embedding model, data persisting in workspace/.supermemory, kept alive by launchd, with the plugin's autoCapture and autoRecall both on so every conversation is indexed.
@trevin [OpenClaw]
https://x.com/trevin/status/2065888997721936239
Set up the Viktor Slack agent for a team and calls it the best out-of-the-box Slack agent he has tried after many. He notes Hermes and OpenClaw can get far more custom but the tinkering gets out of control in a company context with different permission levels, so Viktor fits teams and Hermes fits solo. Concrete result: up and chatting in Slack in under five minutes, then another 30-45 minutes wiring up first workflows for Linear status, Granola meeting recaps and a Sentry integration.
@MPxbt [Claude Code]
https://x.com/MPxbt/status/2065917841111491067
A trader open-sourced a tool (just scripts plus config) that lets Claude Code watch TradingView charts live by hooking into TradingView Desktop via Chrome DevTools, so Claude sees real-time candles, levels and price action. It works: Claude reads a live 15-minute Dollar chart, maps supply/demand zones, draws them and analyzes structure as the market moves, with no indicators or paid Discord signals. You clone the repo, set watchlist/bias/risk in JSON and run one script to get an on-demand chart analyst with eyes on the screen.
@egocgp [Claude Code]
https://x.com/egocgp/status/2065736435294937281
Building his own AI-agent orchestrator platform, one per business unit, by forking Paperclip with a remote UI plus browser window. The concrete stack: SQL queries into Claude Code hosted on Hetzner, noVNC to control the browser, every possible MCP connector, Supabase databases compartmentalized per business unit, and Telegram/PWA to talk to his agents from his phone, with an agentic dashboard on Netlify or Tailscale. He concludes he doesn't really need OpenClaw or Hermes for this self-hosted multi-business setup.
@aye_dreee_an [OpenClaw]
https://x.com/aye_dreee_an/status/2065801920204165241
Vibecoded a 3D Hong Kong-style Mahjong game with single player and online multiplayer that also supports real-time voice and video chat, made entirely through Discord chat with his OpenClaw agent using GPT-5.5 as the primary model. A concrete shipped product built end-to-end via an agent over a plain chat interface, no IDE in the loop.
@erukiti [Claude Code]
https://x.com/erukiti/status/2065635385158607304
An honest first-person Fable 5 review: it's not omnipotent and can even become useless for various reasons (especially Claude Code's own bugs), but it's unmistakably smart, comparable to or beyond the intelligence-shift feeling when GPT-5, even GPT-4, first appeared. Concretely, an experiment that was impossible no matter how hard he tried on Opus 4.8 or GPT-5.5 actually worked on Fable 5, which made its shutdown genuinely painful.
@conao_3 [Claude Code]
https://x.com/conao_3/status/2065815666310238330
His single most valued task in two days of Fable 5: a memory-dream SKILL that periodically reorganizes the memory hierarchy to remove duplication, contradictions and staleness, recreating Anthropic's managed-agents Dreams feature in a manual git-based environment. He shares the full skill, including a four-phase procedure (Mine, Consolidate, Dedup-and-Resolve, Prune-and-Index), dedup rules, a checklist and trigger conditions, and notes it applies whether you use a custom memory system or Claude Code's built-in memory files.
@israfill [Claude Code]
https://x.com/israfill/status/2065868713895829991
Describes agent-reach, a trending open-source tool (23k stars) that gives an AI agent free real-time access to Twitter/X, Reddit, YouTube, GitHub and 10+ platforms via direct parsing with no API keys or billing. Setup is a pip install plus connecting it as a tool, and it works with Claude Code, Cursor, aider and other frameworks. Positioned as zero-marginal-cost web access for research agents, content radar and competitive intel instead of paying $100+/mo for platform APIs.
@dtcprophet [Claude Code]
https://x.com/dtcprophet/status/2065885792246157401
Connected Claude Code to a Shopify theme via GitHub and calls it magical for ecommerce. He scrolls his live site on mobile as a user would, talks through fixes and updates, and Claude makes them and deploys easily. It also audits and optimizes the codebase, removing dead CSS, old scripts and sections and doing speed work, and is rebuilding a clunky paid auto-add-to-cart app natively. He's non-technical and figured out the whole setup in under an hour, replacing a 7-14 day Shopify dev-shop turnaround.
@akshay_pachaar [Claude Code]
https://x.com/akshay_pachaar/status/2065779599745995235
A practical map of the top Claude Code CLI integrations and what each unlocks: GitHub so the agent reads and writes issues, PRs, Actions and releases; HuggingFace to pull a base model, train and push the fine-tuned version from the terminal; Bright Data for live search and scraper generation; Stripe to forward live webhooks and run real payment events; Playwright to click, fill forms and run UI tests in a real browser; E2B sandboxes to safely run agent-written code; plus CodeRabbit, Slack and more. A concrete reference for extending Claude Code beyond the terminal.
@shushant_l [Claude Code]
https://x.com/shushant_l/status/2065720874783932730
Walks through building an AI second brain to offload an overloaded memory, with a detailed architecture: a structured vault for foundation and daily captures, a memory layer, and an AI agent layer connected via MCP. The system captures everything for you and lets you retrieve it on demand, turning scattered notes into a queryable knowledge base. A concrete personal knowledge-management build oriented around AI agents rather than another note app.
🗣 User Voice
User Voice
Token cost is still the loudest theme: people route plans to Opus or GPT-5.5, code with cheaper models like K2.7 or DeepSeek, and obsess over context windows, with @starmexxx showing a third of public skills cost more tokens than they save.
Memory that doesn't rot is the recurring infrastructure ask: @zagundo wants a slim boot primer plus dynamically-recalled indexed memory instead of one bloated file, and @itarutomy wants distilled, transparent, version-controlled knowledge instead of stale notes pointing at deleted files.
Parallelism needs better ergonomics: @helloyuki_ is reworking tmux and worktrees just to run several Claude Code sessions at once without losing track.
Design taste has to be injected: @gippp69 shows Claude Code goes from 2/10 to 9/10 on-brand screens only once you feed it a real design system, otherwise it invents random colors and spacing.
And there's genuine grief over model deprecation: @erukiti and @AlchainHust both describe Fable 5 unlocking things nothing else could, making its shutdown sting.
Token cost is still the loudest theme: people route plans to Opus or GPT-5.5, code with cheaper models like K2.7 or DeepSeek, and obsess over context windows, with @starmexxx showing a third of public skills cost more tokens than they save.
Memory that doesn't rot is the recurring infrastructure ask: @zagundo wants a slim boot primer plus dynamically-recalled indexed memory instead of one bloated file, and @itarutomy wants distilled, transparent, version-controlled knowledge instead of stale notes pointing at deleted files.
Parallelism needs better ergonomics: @helloyuki_ is reworking tmux and worktrees just to run several Claude Code sessions at once without losing track.
Design taste has to be injected: @gippp69 shows Claude Code goes from 2/10 to 9/10 on-brand screens only once you feed it a real design system, otherwise it invents random colors and spacing.
And there's genuine grief over model deprecation: @erukiti and @AlchainHust both describe Fable 5 unlocking things nothing else could, making its shutdown sting.
📡 Eco Products Radar
Eco Products Radar
Fable 5 - the short-lived top model that drove most of today's heavy-usage stories before being cut off
Claude Code - the default harness across nearly every case, from bookkeeping to UI to trading charts
Codex / Kimi K2.7 / DeepSeek - the cheaper coding models people swap behind the harness to cut cost
OpenClaw / Hermes - the persistent self-hosted agents people run 24/7 alongside session-based tools
MCP - the connector layer behind slides, websites, TradingView and design-system workflows
Higgsfield / Nano Banana Pro - the image/video generation stack powering faceless video and website cloning
Moonchild - the design system repeatedly credited with making Claude Code's UI output sellable
Paperclip - the open-source self-hosted multi-agent orchestrator behind several agent-team builds
Fable 5 - the short-lived top model that drove most of today's heavy-usage stories before being cut off
Claude Code - the default harness across nearly every case, from bookkeeping to UI to trading charts
Codex / Kimi K2.7 / DeepSeek - the cheaper coding models people swap behind the harness to cut cost
OpenClaw / Hermes - the persistent self-hosted agents people run 24/7 alongside session-based tools
MCP - the connector layer behind slides, websites, TradingView and design-system workflows
Higgsfield / Nano Banana Pro - the image/video generation stack powering faceless video and website cloning
Moonchild - the design system repeatedly credited with making Claude Code's UI output sellable
Paperclip - the open-source self-hosted multi-agent orchestrator behind several agent-team builds
Comments