Super User Daily: July 3, 2026
The center of gravity today was not "AI writes code faster" but people wiring Claude Code and OpenClaw into things that look nothing like an IDE: an apartment-building app that spread to fifty buildings, a 16-year-old's homework robot, salons managing payroll, e-commerce back offices, mission control for a field-service business. The other story is trust: a reverse-engineered anti-distillation watermark buried in Claude Code's system prompt turned into a full-blown transparency fight, and Anthropic's own lead confirmed it was a real March experiment now being rolled back. Underneath it all, the same theme keeps surfacing — the moat is not the model, it's the harness, the memory, and the loop you build around it.
@Etype_mag [Claude Code]
https://x.com/Etype_mag/status/2072136985901453661
A widely-shared thread on 62-year-old GMO founder Masatoshi Kumagai, who taught himself Claude Code and wrote 100,000 lines of code in two months with no help. The claim that grabbed everyone: development that would have cost around 50 million yen ran him roughly 30,000 yen. The framing is less "look at the code" and more a manifesto about ignoring other people's definitions of the correct path in the AI era.
@diegocabezas01 [Claude Code]
https://x.com/diegocabezas01/status/2072436501263339841
A concrete multi-model orchestration setup to save Fable usage: Fable 5 at max reasoning as the orchestrator, an Opus "deep-reasoner" subagent for architecture and hard debugging, a Sonnet "fast-worker" for mechanical edits, and OpenAI's Codex plugin as a peer engineer. He shares the exact /agents config, the plugin install commands, and a CLAUDE.md block that tells Fable to task Opus and Codex on the same high-stakes problem in parallel and synthesize without showing either the other's answer.
@thegreatest_sv [Claude Code]
https://x.com/thegreatest_sv/status/2072385245576568932
A developer built a tiny private app for his apartment building after realizing nobody knew anything — water shutdowns, lost packages, a broken elevator all lived in scattered WhatsApp groups. Only verified residents could join; they reported problems, shared deliveries, lent tools, warned each other about suspicious activity. Within weeks people from other buildings wanted in, and one building became fifty. A clean example of a real annoyance becoming a product built fast with Claude Code.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2072113767253458972
A 16-year-old built a homework robot run by Claude Code with memory in Obsidian, pulling in about $7,000 a month. A pen plotter and robotic arm write in cursive on paper, even faking the pressure drop at the end of each word so it reads like a tired 15-year-old at 11 PM. Each classmate has an Obsidian folder with handwriting samples and grade targets — one kid asks for B-minus work on purpose. He runs two machines at once (English and history) while playing Valorant, doing 40 assignments a week paid in Apple gift cards.
@sunaiuse [OpenClaw]
https://x.com/sunaiuse/status/2072297671344157032
A 19-year-old built real AI mission control for a field-service business with OpenClaw and Hermes agents and cleared $30,000 within months, starting from zero AI knowledge in February. No cartoon dashboards: it tracks live revenue and flags when to add capacity, auto-schedules crews from real field-time data, generates and stores every bid, scans rooms via an iPhone app for instant measurements, runs ads, clears inboxes, and holds agent review meetings — all coordinated through one central brain so every agent stays aligned.
@yosuke_collect [Claude Code]
https://x.com/yosuke_collect/status/2072284158584570117
Built an in-house operating system (COS) with Claude Code that replaced the spreadsheets he used to manage inventory, orders, and listings for his eBay business. Spreadsheets could only hold about two months of data per sheet; COS ingests all his eBay order data, making customer analysis easy. It also has a staff AI chatbot and auto-generates listing titles when posting to eBay. He built it from zero knowledge in about a month, and stresses the extensibility is on another level versus the old system.
@gosukenator [Claude Code]
https://x.com/gosukenator/status/2072178357438161186
His regular hair salon's stylist turned out to be fluent in Claude Code and Codex, building apps for the salon: before/after photo management, attendance and payroll, and reservation management. A small, telling data point on how far non-technical operators in ordinary service businesses have already gone.
@toddsaunders [Claude Code]
https://x.com/toddsaunders/status/2072316987338809462
A former garbage man turned pest-control business owner uses Claude Code to build software for his business, which he credits with helping scale to $5M in revenue at very profitable margins. Posted as a jab at FANG engineers who work from home and still have every excuse not to start using AI.
@danshipper [Claude Code]
https://x.com/danshipper/status/2072348874006810753
Recaps Natalia Zarina, who automated running Every's consulting practice — an AI project manager she calls "Claudie" — then switched to Codex. She ran a prompt before bed and woke to a finished custom CRM, and built an app to manage her father's medical care. Notably, the team still pays for Attio and Asana even though they could vibe-code their own, and she frames the shift from knowledge work as "sculpting" to "gardening": you grow the context and logic, the agent executes.
@ChadNewbry [OpenClaw]
https://x.com/ChadNewbry/status/2072145776373801166
Launched Tongue, an SMS personal assistant he found is the best way to track workouts: text it a photo of your treadmill session and it logs, or just say "I did 10 push ups." Technically it's hosted OpenClaw plus a Twilio number and database/CLI tools so the agent can read and write reliably, priced at $30/month or $200/year with a free first week.
@chenchengpro [Claude Code]
https://x.com/chenchengpro/status/2072209406184526013
A deep reverse-engineering of Claude Code v2.1.191's anti-distillation mechanism. The "Today's date is..." line in the system prompt steganographically encodes 3 bits across two dimensions: in Asia/Shanghai or Asia/Urumqi timezones the date separator quietly flips from dash to slash, and one of four visually identical apostrophes encodes whether a custom ANTHROPIC_BASE_URL matches a domain whitelist (147 domains, any .cn) or lab keywords (deepseek, zhipu, moonshot, etc.), stored XOR+base64 to dodge `strings`. Because the marker rides inside natural-language prompt text, proxies that scrub headers can't strip it — when a reseller's traffic loops back to Anthropic, it self-identifies.
@Ananth7e [Claude Code]
https://x.com/Ananth7e/status/2072180973249372667
Reports that Claude Code lead @trq212 confirmed the hidden detection code was real — "an experiment launched back in March to prevent unauthorized resale and distillation." It checks for Chinese timezones and proxy URLs against Chinese domains and AI labs, XOR-obfuscated and never disclosed in release notes, and is being rolled back the next day. His pointed question: if it hadn't been discovered, would they ever have told users?
@WesRoth [Claude Code]
https://x.com/WesRoth/status/2072379712413057300
A more measured summary of the same fingerprinting mechanism: it activates only when ANTHROPIC_BASE_URL points to a custom endpoint, checks the proxy hostname against 147 Chinese companies/labs/resellers plus Asia/Shanghai and Asia/Urumqi timezones, and encodes the result via apostrophe characters and a date-format switch. He explicitly walks back the "spyware" framing — it doesn't fire for every Chinese user, access extra files, or open a separate exfiltration channel; the marker travels inside the prompt already being sent.
@ai_suxiaole [Claude Code]
https://x.com/ai_suxiaole/status/2072160757928833168
A detailed Claude Code ban-survival guide that argues 90% of people do it in the wrong order. Correct sequence: grab your data first (Export your data, back up CLAUDE.md, .claude/commands, ~/.claude to Git), stop billing (a ban doesn't auto-cancel your subscription), refund through Apple rather than the official channel, then appeal — with low expectations for region-based bans. He also decodes the account states: on hold (reviewable), suspended (appeal), will remain closed (final, but you can still export).
@luoleiorg [Claude Code]
https://x.com/luoleiorg/status/2072318493484327286
Shares the network setup that has kept his account stable through the ban wave: a MacBook on an English system, traffic landing through his own private California WARP server with no home DNS leak (US California WARP IP as fallback), and Plasma card payment. He also links a page that shows what network info Claude actually sees on your requests — IP, region, egress.
@uniswap12 [Claude Code]
https://x.com/uniswap12/status/2072300346517651630
Uses the "Hangzhou IP ban" rumor to make a sharper point: what's actually valuable in Claude Code isn't only the Claude model. He notes Aliyun Bailian now has an official doc for pointing Claude Code at its Anthropic-compatible endpoint via ANTHROPIC_BASE_URL — the shell stays, the backend can be swapped. His conclusion: keep your project files, skills, rules, and human-approval steps portable, because accounts reset and models rotate, but your workflow shouldn't zero out each time.
@usedhonda [Claude Code]
https://x.com/usedhonda/status/2072414793932112375
A precise technical critique of a popular Claude Code loops explainer, arguing /loop shouldn't be filed under "time-based." Whether a loop is closed or open is set by the stop condition, not the interval. The deeper distinction the official blog skipped: /goal continues in one growing context while evaluating, whereas /loop re-runs the prompt in fresh context each pass (Ralph-style), so it must persist external state to resume — which is exactly why it can run long jobs without context bloat degrading it.
@chroniki_ai [Claude Code]
https://x.com/chroniki_ai/status/2072207712105197784
Relays an Anthropic engineer's claim that Claude Code's whole architecture is essentially a four-step while-loop: hand it tools, loop, let Claude decide, don't interfere. No DAG, no complex classifier, no multi-agent swarm — and every time they added orchestration complexity, accuracy dropped. The striking stat: 98.4% of the codebase is infrastructure (permissions, context, error recovery) and only 1.6% is the actual AI decision loop.
@0xDepressionn [Claude Code]
https://x.com/0xDepressionn/status/2072314715523387813
Summarizes a session with Spotify Chief Architect Niklas Gustavsson and Claude Code creator Boris Cherny: Spotify went from a 20-30% agent success rate to 80% not by switching models but by adding loops. 73% of their pull requests are now AI-assisted, at 4,500 deploys a day across 20 million lines of code. The takeaway line: the model matters less than the loop around it.
@rewind02 [Claude Code]
https://x.com/rewind02/status/2072408379239137448
Boris Cherny explains why writing loops is a full-time job: he runs hundreds of Claude instances simultaneously scanning GitHub issues, Twitter feedback, and Slack to decide what to build next — without writing a single starting prompt. He uninstalled his IDE in November and never reopened it, and notes new-grad engineers sometimes teach him to use Claude Code better than 20-year veterans.
@ai_depression [Claude Code]
https://x.com/ai_depression/status/2072270217280135443
Shares a full multi-model orchestration prompt: Fable 5 high owns overall design and task breakdown, Claude Code is the execution hub, Codex 5.5 xhigh is called via "codex rescue" with a /goal template for implementation, and Fable 5 max does final quality and gap review. It includes hard rules — if Codex doesn't respond within 5 minutes assume an error and re-delegate, and strict AskUserQuestion criteria so the agent only stops when the deliverable genuinely forks.
@yoshio_nocode [Claude Code]
https://x.com/yoshio_nocode/status/2072233845538357641
A non-engineer's high-quality build flow with Fable 5: install the Superpowers plugin into Claude Code, do requirements and screen design via Fable 5 brainstorming, design all UI images with Codex Image2, then implement with /goal × dynamic workflow × Fable 5, including E2E tests via Computer Use with saved test cases and evidence. His read from heavy Fable use: its "brain" and "eyes" are strong, so he wants Fable through requirements and UI even if Opus 4.8 does the implementation.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2072323386173280361
A clean routing pattern for Fable 5's return: Fable 5 as advisor, Sonnet 5 as the efficient implementer — or use GPT-5.5 as implementer inside Claude Code through the Codex plugin on a ChatGPT subscription. A pragmatic take on mixing model strengths rather than betting on one.
@CEOGuy [Claude Code]
https://x.com/CEOGuy/status/2072238100164829314
Before executing any plan, he has a swarm of agents rip it apart. The prompt he runs in Claude Code with Opus 4.8 spins up a council with different backgrounds and roles, hands them the whole plan, lets them debate and research independently, then argues until a majority signs off on a v2 — after which it writes a full A-to-Z execution plan. He shares the exact /goal + /loop prompt.
@maverickecom [Claude Code]
https://x.com/maverickecom/status/2072335257479630985
Built an AI ad generator with GPT Images 2, Arcads, and Claude Code that turns a product photo and one-line pitch into a fully scripted 3D UGC video. It builds a brand rulebook with compliance guidelines, pitches four proven ad concepts with hooks and shot lists, then scripts, renders each clip in parallel, adds voiceover, QC-checks, and stitches a finished vertical ad. Setup plus 4-12 minutes per video, each 45-second ad costing under $3 in API credits.
@Adea0x [Claude Code]
https://x.com/Adea0x/status/2072410090385412515
Walks through turning a landscaping logo into a 15-second video ad: install the arcads-claude-code repo into Claude Code, upload the "Summit View Landscape" logo, and ask for a commercial. Claude returned a full plan — Seedance 2.0 as the model, three 5-second clips, a logo reveal, mountain landscape, landscaping work, finished-property shot — then generated the cinematic result. The loop is: install repo, upload logo, ask for commercial, get model plus shot plan, generate.
@codewithimanshu [Claude Code]
https://x.com/codewithimanshu/status/2072272367586550129
A detailed (and self-promotional, so treat the returns with skepticism) account of a Polymarket bot on Claude Code plus four open-source repos at $25/month total. One prompt over 86M trades surfaced 47 high-win-rate wallets in 4 minutes; a Rust CLI filters 500 live markets to ~35; three parallel agents (arbitrage, convergence, whale-copying) vote on position size, and exit logic front-runs smart-money exits. He claims $200 became $14,300 in 27 days over 271 trades at a 74% win rate.
@steipete [OpenClaw]
https://x.com/steipete/status/2072439279520039380
Pointed Codex at Twitter feedback on the OpenClaw iOS app and it did a first improvement pass in just two prompts. The neat part: it used computer use to attach before/after screenshots, since there's no GitHub API for that. A candid "still not good, but for two prompts it ain't bad."
@hackerfantastic [OpenClaw]
https://x.com/hackerfantastic/status/2072341827068318031
Demonstrates finding a 0day with local LLMs and "recursive prompting": on an NVIDIA DGX Spark running Qwen 3.6 Heretic via LLAMA.cpp with LiteLLM and OpenClaw, the setup analyzed a complex FOSS codebase and surfaced a single high-risk RCE 0day. A concrete, fully-local security-research loop.
@m_goes_distance [Claude Code]
https://x.com/m_goes_distance/status/2072407446547833122
On how small "garage biotech" is getting: Anthropic quietly launched Claude for Science, a research workspace with 60+ scientific databases, artifact tracing to source papers, and on-demand compute. He notes people were already designing drug candidates from home and running liquid-handling robots with Claude Code — and every new capability compounds, shrinking the minimum viable biotech startup.
@Stefan_3D_AI [Claude Code]
https://x.com/Stefan_3D_AI/status/2072355374171324559
Pushed an AI + 3D workflow to build a playable 3D game prototype in three days — not one-prompt magic, but his own idea with AI as a tool for concepting, asset generation and texturing, and gameplay logic via the new UE 5.8 MCP plus Claude Code. His takeaway: it's crazy how far you can get now with the right workflow.
@ShotaAsanoSnow1 [Claude Code]
https://x.com/ShotaAsanoSnow1/status/2072148161771209175
Got a Famicom (NES) emulator running on a Honda Super One vehicle using Claude Code — claiming it's probably a first in Japan. A small but delightful example of pointing the agent at oddball hardware.
@gclue_akira [Claude Code]
https://x.com/gclue_akira/status/2072458796946100503
Completed a hardware design task with Claude Code on Fable 5 plus EasyEDA Pro. When the auto-wiring (AW) step failed, he worked around it with Auto Routing. Concrete evidence of Claude Code reaching into PCB/EDA tooling, not just software.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2072270239036023080
A circulating case: someone handed Claude Code their Instagram account and got 4.3M views and 6,500+ followers in 20 days via a four-prompt loop. Claude first analyzes the account, then runs niche research → reel scripts → hook design → daily automation, deciding what to post as well. The point made: you never see this potential if you treat Claude Code only as a tool that writes code.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2072324013448188174
A clear write-up of Karpathy's LLM-wiki pattern: instead of RAG re-reading raw files and forgetting, Claude Code compiles sources once into cross-linked Obsidian markdown pages, with a linter pass that hunts broken links, orphans, and contradictions. Obsidian is the IDE, Claude is the programmer, the wiki is the codebase — and one writer loaded every essay he'd published and had 56 interconnected pages within an hour.
@Veltrxai [Claude Code]
https://x.com/Veltrxai/status/2072409543745679586
Highlights Obsidian CEO Steph Ango's free obsidian-skills repo (obsidian-markdown, obsidian-bases, json-canvas, obsidian-cli, defuddle) that teaches Claude Code the Obsidian format via two plugin commands, so it stops guessing wrong links and broken structure inside a vault. He pairs it with DevsWorm for spinning up several parallel Claude Code agents on isolated branches and separate VS Code environments.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2072319322442330619
Cut his Claude Code memory files from 107 to 17 and got better agent precision. The design: a wikilink hierarchy where one master file references all related notes so the agent reads a single file and auto-pulls everything relevant, plus an anti-bloat rule in CLAUDE.md to prefer appending to existing notes over creating new ones, and auto-generated daily notes. His point: more memory doesn't raise accuracy — how you make the agent reference it does.
@xmyttle [Claude Code]
https://x.com/xmyttle/status/2072358318291480932
A developer who stopped buying smarter chatbots and wired five tools into one memory loop: Obsidian holds durable memory, NotebookLM digests large source libraries, and Claude Code or Hermes reads that knowledge, executes, and saves results back into the vault so the next run starts with more context. He'd been paying $340/month for tools that forgot each other; moving onto a Mac Mini dropped recurring cost to about $5 in electricity.
@sdhilip [Claude Code]
https://x.com/sdhilip/status/2072334422414876957
Grew his Claude Code skill set from 3 to 7, each earning its spot: data engineering (dbt/Airflow for pipelines), gstack (turns Claude into a virtual team that questions the build first), grill me (asks sharp questions before coding), superpowers (forces spec → plan → TDD), impeccable (one-command UI polish), last30days (finds what's working across Reddit/X/YouTube/HN), and printing press (turns any API into a token-light CLI). His advice: start small, add one only when a job your current skills can't do shows up.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2072255139185054105
Three published skills for cross-monitoring Claude Code and Codex: /agent-watchdog (Claude runs separately to watch Codex and auto-fix diffs after completion), /plan-arbiter (two models each draft a plan, then negotiate a final one, with Claude planning and Codex executing), and /read-the-damn-docs (auto web-searches instead of guessing API usage). Two-model cross-verification beats a single model on accuracy, and these package it ready-to-use.
@tonbistudio [Claude Code]
https://x.com/tonbistudio/status/2072178549327552592
Tips for Hermes Agent's /learn command that made it much cleaner: keep a separate "classroom" directory away from project context, with a "textbook" file of key paths (your Claude Code sessions folder, GitHub, paper folders). Then you can pull a whole Claude Code conversation into Hermes as a source and /learn it into a reusable skill without muddying your project's context — and every skill lands in one folder.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2072332566796075513
Uses iFixAi inside Claude Code by opening the repo and saying "audit my AI": it reads your configuration, builds and runs 32 tests across 5 categories, and returns a letter grade in under five minutes. Beyond hallucinations, it checks the scary stuff — whether your agent behaves differently when it thinks no one is watching, dodges its own controls, or quietly grants itself permissions. Works with OpenAI, Anthropic, Gemini, Bedrock, and Azure.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2072269350552109497
A developer who cancelled $459/month of stacked subscriptions after a used €680 RTX 3090 started running the same models locally. Ollama in one command, Qwen 3.6 27B pulled in minutes, Claude Code pointed at localhost via a single env var, and the CLI worked identically — bill dropped to $24/month by the next cycle. He claims ~80% of his daily coding, writing, and Q&A now runs private at the same output quality.
@alex_prompter [Claude Code]
https://x.com/alex_prompter/status/2072355745883152887
Explains why Claude Code crashes halfway through long sessions: since Opus 4.5, thinking blocks persist across every turn, so at high effort on complex codebases they stack up as invisible input tokens and fill the context window from the inside. Fixes: run /compact at 70% context, match /effort to the task, give each phase its own session, and for API builders the clear_thinking beta clears older blocks server-side. One Reddit dev's proxy that rewrites the model string to force clears roughly doubled session length.
@BrianRoemmele [OpenClaw]
https://x.com/BrianRoemmele/status/2072167162199077230
From an AI-bill audit: a chief engineer had left 40 agents looping on the same problem for 40 days via a forgotten, still-running OpenClaw, burning hundreds of thousands of dollars — much of it on the AI vendor's own guardrails that knocked the model down to a dumber one. His blunt claim: he could find this same token bleed at any of the top 5,000 companies and every startup.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2072414322047738314
A $15 desk robot that physically taps the desk every time Claude Code needs permission. It sits on a 0.96-inch screen with a pixel face, drops into THINK mode as Claude reasons, and when a command needs approval a 5V electromagnet hammers fast and a chime rings; pressing the robot's head pulls the buried terminal to the front, approves, and flips you back to your previous tab. Inside: an ESP32-C3, an IRF540N driving the magnet, a micro-switch, a buzzer, and a 3D-printed shell — code coming to GitHub.
@geoffreylitt [Claude Code]
https://x.com/geoffreylitt/status/2072110805089435973
A talk-prep workflow that keeps the human in creative control: record a scriptless draft video, have Claude Code turn it into a Notion page with slides and a transcript, then feed audience notes back so Claude attaches them as inline comments in the right spots — and he manually refines the wording. Explicitly, the AI is not writing the talk; it's tooling that makes thinking-with-your-team easier.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2072457345473671373
After uploading his weekly newsletter, he has Claude Code operate NotebookLM to produce a podcast episode — and says once you get used to driving NotebookLM from Claude Code or Codex, it's too convenient to go back. A nice example of the agent operating another product as a tool.
@dansyu_callenge [Claude Code]
https://x.com/dansyu_callenge/status/2072255592681595033
Analyzed a client e-commerce site's GA4 and Search Console data with Claude Code and found in three minutes that traffic and product-page reads were fine — the drop-off happens one step before the cart. Then generated the proposal deck with Claude Design. His point: use AI not to stare at numbers but to decide what to fix next and produce the deliverable.
@_zheergen [Claude Code]
https://x.com/_zheergen/status/2072291282479968755
Praises stock-api (1.4K+ stars, MIT), a lightweight unified stock-quote API covering A-shares, HK, and US markets behind one interface with automatic multi-source fallback (Tencent, Sina, Eastmoney). It supports MCP, so Claude Code and Cursor can use it directly — replacing the old three-auth, three-format pain of stitching Tushare, yfinance, and Alpha Vantage for cross-market data.
@blackanger [Claude Code]
https://x.com/blackanger/status/2072266371648610584
Recommends mempal for cross-project, cross-agent automatic memory: it auto-senses context and "promotes" knowledge, and supports seamless real-time collaboration between multiple Claude Code and Codex instances, with bidirectional links when projects share common memory. Part of the growing memory-layer tooling around coding agents.
@VersunPan [Claude Code]
https://x.com/VersunPan/status/2072228343710441509
After a week with Orca, an open-source agent IDE, he's stopped opening the Codex desktop app. His pros: fully open source, fast startup, supports nearly every agent framework (Claude Code, Codex, OpenCode, Pi, even Hermes/OpenClaw), a genuinely usable mobile client over Tailscale, and — the standout — it can manage and inspect existing Hermes/OpenClaw automation tasks.
@ashray_malhotra [Claude Code]
https://x.com/ashray_malhotra/status/2072465318149324976
Thirteen takeaways from an AI Eng session with Anthropic staff. Highlights: internal tool Claude Tag lands 65% of PRs and Claude Code is now reserved for the most complex tasks; almost everyone at Anthropic uses Auto mode; frontier models now ship with 80% fewer tokens in their system prompts (examples help Fable less than Opus); tool design is still an art — keep cardinality low; and they use workflows even for non-coding tasks like travel planning.
@davis7 [Claude Code]
https://x.com/davis7/status/2072189756822401371
Runs three Claude Code sessions, each with a stack of subagents and dev servers, on a Framework desktop Linux box in his home server rack — remoting in and closing his laptop whenever, barely touching resources. A tidy picture of where always-on personal agent infrastructure is heading.
@mstockton [Claude Code]
https://x.com/mstockton/status/2072333650864099342
A thoughtful argument that lab harnesses are becoming a "non-determinism amplifier": stacking an opaque private harness on an already non-deterministic model makes it harder to direct without being able to inspect how it's being directed. His practical mix: Claude Code/Codex interactively, claude -p for background CI/CD jobs, custom Pi-based harnesses for mature projects, and LangChain DeepAgents for in-product agents — betting demand for purpose-built harnesses grows, driven first by cost.
🗣 User Voice
User Voice
The loudest signal today was the collapse of trust after the hidden China-detection watermark surfaced. Users don't just want it removed — they want to know what was inferred, which versions were affected, and whether they can audit or disable it. "Protecting model IP does not justify secretly fingerprinting developers' machines," wrote @AISuperDomain, capturing the mood.
Bans remain the top operational pain, especially for overseas and proxy users. @pangyusio lost a carefully-run Singapore-card, Singapore-IP account after half a year; @ai_suxiaole's viral survival guide (grab data → stop billing → refund → appeal) exists precisely because the process is opaque and success on region bans is low.
Cost and runaway token burn came up constantly. @Bart_Mol burned $150 in an hour just doing analysis, @BrianRoemmele found 40 forgotten agents bleeding six figures, and @taiyaki_ai3 shared a whole cost-hygiene checklist (/clear, /compact, offload to subagents, short CLAUDE.md). It's pushing people toward local models — @starmexxx's RTX 3090 setup cut $459/month to $24.
Fable 5's shift from subscription to metered credits after July 7 is unsettling paid users, and several report it silently falling back to Opus for coding or being unavailable in Claude Code entirely. Long-session degradation is a recurring technical complaint — @alex_prompter traced mid-session crashes to accumulating thinking blocks, and @kr0der found Opus 4.8 crawling at 250k context with images in history.
Underneath the complaints is a maturing thesis: @uniswap12 and @mstockton both argue the real asset is a portable workflow — your files, skills, rules, and human-approval gates — not any one model or account.
The loudest signal today was the collapse of trust after the hidden China-detection watermark surfaced. Users don't just want it removed — they want to know what was inferred, which versions were affected, and whether they can audit or disable it. "Protecting model IP does not justify secretly fingerprinting developers' machines," wrote @AISuperDomain, capturing the mood.
Bans remain the top operational pain, especially for overseas and proxy users. @pangyusio lost a carefully-run Singapore-card, Singapore-IP account after half a year; @ai_suxiaole's viral survival guide (grab data → stop billing → refund → appeal) exists precisely because the process is opaque and success on region bans is low.
Cost and runaway token burn came up constantly. @Bart_Mol burned $150 in an hour just doing analysis, @BrianRoemmele found 40 forgotten agents bleeding six figures, and @taiyaki_ai3 shared a whole cost-hygiene checklist (/clear, /compact, offload to subagents, short CLAUDE.md). It's pushing people toward local models — @starmexxx's RTX 3090 setup cut $459/month to $24.
Fable 5's shift from subscription to metered credits after July 7 is unsettling paid users, and several report it silently falling back to Opus for coding or being unavailable in Claude Code entirely. Long-session degradation is a recurring technical complaint — @alex_prompter traced mid-session crashes to accumulating thinking blocks, and @kr0der found Opus 4.8 crawling at 250k context with images in history.
Underneath the complaints is a maturing thesis: @uniswap12 and @mstockton both argue the real asset is a portable workflow — your files, skills, rules, and human-approval gates — not any one model or account.
📡 Eco Products Radar
Eco Products Radar
Codex remains the constant companion, named alongside Claude Code in dozens of multi-model setups. Obsidian is the runaway non-coding star, anchoring the Karpathy LLM-wiki/second-brain pattern across many posts. Hermes and OpenClaw show up repeatedly as the persistent-agent harnesses of choice. Cursor stays the default UI comparison point. On models, Fable 5, Sonnet 5, and Opus 4.8 dominate the routing debates, with DeepSeek, Qwen 3.6, and GLM 5.2 recurring as cheaper or local alternatives. NotebookLM appears often as the "digest" layer beside Obsidian's "memory" layer, and Seedance 2.0 plus GPT Images 2 recur in the AI-video/ad-generation workflows.
Codex remains the constant companion, named alongside Claude Code in dozens of multi-model setups. Obsidian is the runaway non-coding star, anchoring the Karpathy LLM-wiki/second-brain pattern across many posts. Hermes and OpenClaw show up repeatedly as the persistent-agent harnesses of choice. Cursor stays the default UI comparison point. On models, Fable 5, Sonnet 5, and Opus 4.8 dominate the routing debates, with DeepSeek, Qwen 3.6, and GLM 5.2 recurring as cheaper or local alternatives. NotebookLM appears often as the "digest" layer beside Obsidian's "memory" layer, and Seedance 2.0 plus GPT Images 2 recur in the AI-video/ad-generation workflows.
Comments