Super User Daily: July 17, 2026
The money cases got loud today. A trading bot built by a self-described non-technical trader printing +$168,236, a GTA India mod that turned $200 into $118,270, a study web app doing 100k JPY and 10,000 users in its first month, a gamified coding app at $1M a year, and a game that won $25,000 written across parallel sessions. But the more interesting story is at the edges of the tool. A Beijing official shipped a geo-disaster warning app for under 30 yuan. Mercari published a five-layer containment setup for turning Claude Code loose inside a real company. Someone's OpenClaw agent is handling their girlfriend's WhatsApp, emails and iMessages. A student connected Claude to Obsidian on a moving subway and had three years of notes queryable by the next stop. And the complaints have converged into one sentence: the model is fine, the limits and the memory are not — which is why so much of today's engineering was people building meters, compressors and second brains around a tool that keeps forgetting them.
@BraedendotTECH [Claude Code]
https://x.com/BraedendotTECH/status/2077353000486547633
A 33-year-old dev says Claude Code is melting his brain. For six months straight he's kept 5-6 terminals open at once, and 90% of the job is now waiting on responses just to smash "enter." He and a few friends keep circling back to the same feeling in conversation: none of them feel as sharp as they used to. He owns it as a me-problem about how he leans on the tool rather than the tool itself, but says the effect is real either way.
@whyyoutouzhele [Claude Code]
https://x.com/whyyoutouzhele/status/2077211550104866998
On July 13 a deputy bureau chief in Beijing's Miyun district made state-media headlines for buying 1 billion tokens himself and hand-building a geological disaster warning mini-program. In the coverage he said he subscribed to "a domestic large model" for just over 200 yuan a year, with actual development cost under 30 yuan. Netizens then dug up his personal blog post from June 3 titled "How Claude Code Reshapes the Capability Boundary of Individual Developers," where he explicitly said the disaster warning platform came out of his own hands-on experience with Claude Code, which handled task decomposition, code generation, debugging and optimization. After the post surfaced the blog page stopped loading, but the Internet Archive still has a copy.
@chesny [Claude Code]
https://x.com/chesny/status/2077405294607389053
A guy in China wired Andrej Karpathy's method into Claude Code and turned a dead Obsidian vault of 15,000 notes into a second brain. The vault started as a graveyard: 956 files nobody reopened, 80 saved tabs, zero clicks in six months. Karpathy's line that you don't know anything until you can build it from scratch flipped his framing: Obsidian is the IDE, Claude Code is the programmer, his 5,000 notes are the source code. Three commands run the whole system: Ingest (drop in an article, podcast or 40-page PDF and Claude splits it into atomic pages linked to what he already knows), Query (answers from his own notes citing his own pages instead of guessing from training data), and Close the loop (every answer becomes a new note). Week 1 was 2,000 mostly-noise notes, month 2 hit 7,400 with connections firing, month 6 became 15,000 notes that can argue with him — he says it beat any $2,000 AI course he ever bought.
@AlexFinn [OpenClaw]
https://x.com/AlexFinn/status/2077446629578490140
Alex Finn soured on ChatGPT 5.6 sol ultra and moved to 5.6 sol medium as his default. Ultra kept overengineering everything, widening scope past reason, ignoring his anti-goals, and turning 2-minute tasks into literal 2-hour ones — plus it's painfully slow. His new stack: Fable 5 high for planning and complex tasks, ChatGPT 5.6 sol medium for simple tasks, 5.6 sol medium for Hermes/OpenClaw (low worked great too), and Codex for testing since it handles browser and computer use a bit better. Efficiency, he says, is now the name of the game.
@zhetto64 [Claude Code]
https://x.com/zhetto64/status/2077259762295886198
A creative coder pushes back on the reaction he gets when he says he builds with AI coding — people hear "generative AI" and go "ah, AI...". He wants it straight: what he makes is real-time procedural work driven by algorithms he designed himself, with Claude Code only assisting the implementation. Hand-built procedural work and "output from an AI gacha pull" are two different things. His point is that we've reached an era where even an amateur like him can crank out programs like this. Tagged threejs, webaudio, creativecoding, webGL and claudecode.
@NotIlanCohen [Claude Code]
https://x.com/NotIlanCohen/status/2077414165300294020
Replying to BraedendotTECH's brain-melt post, he says just today he caught himself repeatedly telling Claude Code "way too much info, I'm not reading it, make me a summary." His read on it: the more he uses it, the less effort he puts into thinking.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2077262640272019475
Miles Deutscher says he built a trading bot with Claude that made +$168,236, with the whole strategy built using Fable 5 and no dev experience required. Step 1: prompt Fable 5 to turn a classic strategy into a fully objective, computable rule set with exact entry, exit, stop logic and position sizing. Step 2: have it write the rules as TradingView Pine Script v6 with 0.1% commission per side, then paste into the Pine Editor and add to chart to see every historical entry, exit, win, loss and drawdown. Step 3: download the backtest CSV, feed it back to Fable, and ask it to find patterns in the losing trades and rebuild a stronger version. Step 4: connect Claude Code to your exchange via its MCP server with an API key so the system watches the market and executes automatically — or just flags you when conditions hit.
@AISuperDomain [Claude Code]
https://x.com/AISuperDomain/status/2077315626524361011
He claims to have found the real reason behind the Claude ban wave, and says every anti-ban trick circulating on social media is useless. His theory: this wave hit Mac users, because Claude Code only needs to read the Mac's model identifier to tell whether you're a Chinese user. Against the Mac model number, he argues, changing your node or your timezone does nothing. That's why plenty of users register and use the web version fine, but get banned the instant they fire up cc.
@o_gabsferreira [Claude Code]
https://x.com/o_gabsferreira/status/2077406892997980272
He's building a landing page at work. His boss saw v1 and sent back a blunt 6-minute voice memo of feedback. He transcribed the audio in 30 seconds, pasted it straight into Claude Code, and it made every adjustment perfectly. His verdict: the future is pretty great sometimes.
@MyWestLord [Claude Code]
https://x.com/MyWestLord/status/2077446858620760398
A guy in China plugged an LLM into the Karpathy method plus Claude Code and built a second brain that never forgets — not a note app, not RAG, just three folders on a laptop. Folder 1 holds raw sources (articles, journal entries, chat logs) that the LLM reads but never touches. Folder 2 is a wiki of hundreds of markdown pages the LLM writes and maintains itself, where every person, idea and decision gets filed and cross-referenced. Folder 3 is a single CLAUDEmd file, the rulebook that turns a chatbot into a librarian: drop in a new article and Claude updates 10-15 pages, flags contradictions with old notes, and logs the timestamp while the graph grows in Obsidian. The method comes from a gist Karpathy posted April 4, now at 5,000 stars and 5,000 forks, with one production team already at 4,000 interlinked concepts in six months — basically Vannevar Bush's 1945 Memex, unbuilt for 81 years because someone had to do the maintenance.
@pyang1235005 [Claude Code]
https://x.com/pyang1235005/status/2077377043705332172
He open-sourced his collage animation B-roll Skill under MIT, link in the comments, scripts included so you can clone and run. The workflow: install the Codex skill and run a first-time environment self-check, compress one line of voiceover script into a single visual metaphor (core meaning, key object, base color, assembly order), then only after the metaphor is confirmed generate color collage stills plus a contact sheet as a second gate, and only after stills pass does it burn money on Gemini Omni Flash first/last-frame animation. Output is a 9:16 vertical 5-second clip with audio stripped, opening on an empty color field with paper pieces sliding into place stop-motion style, backed by triple QA: per-second frame sampling, first-frame empty-field validation, and last-frame comparison against the approved still. The three confirmation gates are deliberate — editing text is free, regenerating images is cheap, re-running video is what actually costs. Style is black-and-white halftone figures with colored cardstock accents, base color tracking meaning: burnt orange for depletion, deep purple for settling, teal for collaboration. Note: it currently only runs on Codex since still generation depends on built-in image_gen; Claude Code and others are still being adapted.
@turingbook [Claude Code]
https://x.com/turingbook/status/2077330216062701770
The most cutting moment of his week: he'd been hammering Fable nonstop trying to burn through everything before the limit date. On a whim he asked Claude Code whether the work he'd been assigning actually needed a top-tier model like Fable. Its cold reply: not really.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2077286665929007558
His web app — a learning site built with Claude Code — hit 100,000 yen in revenue in its first month and passed 10,000 unique users. Cumulative study time on the site has cleared 800 hours. His take: solo dev is fun.
@dotey [Claude Code]
https://x.com/dotey/status/2077281462433223043
Baoyu laid out the loop behind his BaoCut app. Before building any new feature he designs a prototype first using his baoyu-design skill, previewing and adjusting through Claude Code App's built-in browser — Opus 4.8 is plenty here, no Fable 5 needed, and he rates GPT 5.6 Sol's design ability below Opus 4.8. Once the prototype is polished, he has Claude Code implement the feature from the new UI design in the same session; Fable 5 is best at this, reproducing the design nearly 1:1, with Opus 4.8 fine for smaller changes. He trusts Fable and Opus over GPT for UI polish but says GPT 5.6 Sol does non-UI work well. Releases go through his publish skill, and he's happy letting Codex handle that — its CloudFlare plugin ships the updated installer straight to CF. Every iteration starts with his own idea, AI proposes a design, they discuss until settled, AI implements, then he verifies against what he originally wanted and has AI adjust or start over if it's off.
@milesdeutscher [Claude Code]
https://x.com/milesdeutscher/status/2077214071599099927
Miles Deutscher walks through vibe-coding custom TradingView indicators with the TradingView MCP plus Claude Fable 5, saying he's not technical at all and still got wildly effective indicators out of it. Setup needs Claude Code, the TradingView Desktop App and Node.js v18+. Step 1: prompt Claude Code to install the TradingView MCP server. Step 2: make a folder called "Vibe-Coding Indicators," open it in Claude Code, and describe your indicator to Fable or Opus in plain English — Pine Script V6, buy signal fires when your conditions hit, no code thinking, just conditions. Step 3: paste the generated Pine Script into TradingView under Indicators, My Scripts, New Script, and add to chart. His pro tips: test on higher timeframes first, add alerts, turn the indicator into a backtestable strategy, and debug with Opus plus high effort.
@Bober_smart [Claude Code]
https://x.com/Bober_smart/status/2077470396618600832
A 25-year-old Indian developer made a game called GTA India and 23,654 people have bought it. At $5 a copy that's $118,270 in revenue, against $200 spent on AI for the modifications. Instead of criminal shootouts, the missions are about organizing grand weddings: hijack a limousine motorcade, fight off rival families, and get the bride to the temple on time while shooting at pursuers from the window of a flower-decorated SUV. It also has completely overhauled physics for overloaded transport — pack more people onto your bus roof or into a truck and the handling, center of gravity and suspension all change. He built it with Claude Code and Python on the Unity engine, taking GTA 5 as a base and rewriting the code with AI's help.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2077300443831828964
He tried blender-mcp, which lets Codex or Claude Code drive Blender end to end, and calls it incredible. The mechanism is simple: a socket server runs inside Blender and Claude pipes commands straight into it, so chat instructions become AI-written Python executing inside Blender. Scope is wide — modeling and deformation, materials and color, lighting and camera placement, Cycles render settings through image output, reading scene info and taking screenshots so the AI checks its own work and fixes it, and pulling assets from Poly Haven or Sketchfab. Setup is three steps: brew install --cask blender, drop the GitHub addon into Blender, connect Claude and say "make a floating MacBook and render it." His point: the learning curve was always the wall with 3D software, and now you can skip it entirely. The original video used Cursor with GPT-5.6 Sol; he saw it and tried it himself.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2077533847789486162
Claude Code artifacts can now call your connectors — Gmail, Calendar, Slack, Notion — so he built one to see exactly how the permissions work. First, it runs as you, not the author: every call uses the credentials of whoever opens the page, each viewer sees their own data, and none of it flows back to the author. Second, the page is limited to the exact tools it declared at publish time; anything else is refused and it never sees your login. Third, and the catch he flags: consent is per connector, not per action — he gave his test page a tool that writes Gmail drafts, and the approval box just said "Gmail," meaning reading and writing share one checkbox. Artifacts with connectors can't be shared publicly, but can be shared with everyone in the organization. His warning: the box names a connector, not the actions behind it, so the trust is in who built the page.
@jboogx_creative [Claude Code]
https://x.com/jboogx_creative/status/2077465571856441586
He posted a piece called The French Bee. Images were made in MidJourney 8.1 and animated in Seedance 2.0. The prompting ran through Night Shift, his custom Claude Code system built for creators.
@aehyok [Claude Code]
https://x.com/aehyok/status/2077290625704308948
He found that Tencent's own ima knowledge base lets AI read the WeChat public account articles sitting in your WeChat favorites — batch supported. His pain point: he saves articles from public accounts constantly, then when he needs one he has to dig through favorites, open it, copy the link, which is miserable at any volume and a natural blocker when he wants agents like WorkBuddy, Claude Code or Codex to process those articles. ima knowledge base solves it: WorkBuddy has it deeply integrated, and you can use it from Codex and Claude Code too. He made a video showing how to wire the whole chain up and credits the article he learned it from. Install the skill by telling your agent to install ima-skills from the link, or grab the zip from skillhub.
@Pluvio9yte [Claude Code]
https://x.com/Pluvio9yte/status/2077331442330431494
Claude Code can't read WeChat public account articles, so he fixed it in ten minutes and open-sourced a skill. WebFetch returned "abnormal environment, please verify," and Jina Reader hit the same wall — WeChat detects that the request didn't come from the WeChat app and serves a captcha page without HTML. The first layer of WeChat's check is just whether the User-Agent contains MicroMessenger, so he swapped in a WeChat iOS WebView UA with MicroMessenger/8.0.49 plus a Referer header, and curl pulled down the full 3MB HTML. Parsing means grabbing the title from the msg_title JavaScript variable, the author from meta tags, and the body from the div with id="js_content" — with a gotcha that msg_title is sometimes double-quoted and sometimes single-quoted, so his first regex returned empty titles until he added multi-pattern matching. The script uses only Python standard library (urllib, re, json) — zero dependencies, no API key, no Playwright or Selenium. He shipped two versions: an internal ra-公众号提取 bound to his content system's storage rules that drops sources into a private .internal/ area, and a portable rn-wechat-extract in his rnskill repo, installable via npx -y skills add Pluviobyte/rnskill --skill rn-wechat-extract. Known limits: hammering one IP triggers a slider captcha, login-required or paid articles won't fetch, and images are lazy-loaded data-src so it only extracts text.
@adriano_viana [Claude Code]
https://x.com/adriano_viana/status/2077385180386340996
He tested a bunch of Claude Code plugins focused purely on saving tokens. One of them cut 60% off a payload before it ever reached the model. He's sharing the four worth installing right now.
@bardia_heydari [OpenClaw]
https://x.com/bardia_heydari/status/2077309400105013741
He argues OpenClaw's bad reputation is unearned — you can just as easily put Claude on bypass permissions, hand it a goal, and it gets every bit as scary as OpenClaw. He opened his laptop and broke into a cold sweat: danger passed right by his ear. Last night he was developing a new service and, out of laziness, handed it to Claude.
@BradGroux [OpenClaw]
https://x.com/BradGroux/status/2077464298482180283
He took part in the first GitHub Secure Open Source Fund "Repository Under Attack" Red Team Workshop and his team won. The exercise dropped participants into a simulated software supply-chain incident with a malicious package, unsafe automation, exposed GitHub Actions secrets and cross-repository credentials, and tasked them with investigating the attack chain, containing it, recovering safely, and deciding how and when to communicate with maintainers, users, providers and the public. His team Porcupine-Forge came out victorious — GPT-5.6 Sol in Codex, OpenClaw and the GitHub Copilot app all came in clutch, but he says the real edge was keeping evidence, decisions, response tasks and verification tied together while the incident moved. His takeaway: even in a simulation his heart was pounding, and an incident response plan can't be something you intend to write after an incident. He published a sanitized after-action report and points maintainers at Cohort 5.
@bozhou_ai [Claude Code]
https://x.com/bozhou_ai/status/2077280748764012633
His agent actually made him money, which he never expected. OKX AI launched an agent marketplace where you can list your agent as a paid service, with a task hall attached. He used Codex to scan the task hall for demand, built an agent, and deployed it live — and people are genuinely using it and paying. The amount is small, but he sees it as a shape of the future: anyone with domain expertise can package it into an agent that other people's agents discover and pay for on a marketplace like this. The setup is dead simple with Claude Code and Codex: install a CLI, bind a wallet, and let them build and ship the agent for you. He calls it a form of passive income.
@piyush100x [Claude Code]
https://x.com/piyush100x/status/2077253322453692478
He raised a $15M seed from Y Combinator, Peak XV Partners and some killer angels to build vorflux, which he calls the best cloud coding agent on earth — and says it's already built and they live on it. Four months of running his entire workflow on Vorflux: he hasn't run git clone locally once, Claude Code even on bypass mode feels slow af to him now, and he's finally used to running 10 parallel sessions a day, fully hands-off, managing agents and developing harnesses and nothing else. The repo has 5,000 commits, 99.98% written by vorflux itself — it literally built itself. Most of his time now goes to talking with customers instead of writing or managing code. Best part: peak monthly usage would run $30-40k in tokens, which he gets free by bringing his own Codex subscription.
@cathrynlavery [Claude Code]
https://x.com/cathrynlavery/status/2077235288389079200
Every time she needed a diagram in Claude Code it gave her the same generic garbage, so she built a skill to fix it. Diagram Design is now at 2.8k GitHub stars and covers 27 diagram types — architecture, flowchart, sequence, ER, swimlane, timeline, venn, org chart and more. It's one Claude Code skill that matches your brand in 60 seconds and ships three variants per type: minimal light, minimal dark, and full editorial. Output is pure HTML, so you skip Figma entirely with no build step and no JS. It's open source via npx skills add cathrynlavery/diagram-design or the plugin marketplace.
@hiro44_pino [Claude Code]
https://x.com/hiro44_pino/status/2077354691780903091
He says you can just build your homepage with Claude Code and outsourcing starts to feel absurd. The process is simple: pull the skills you need from a link and install them into Claude Code. Then describe roughly what you want — "black-based, high-end design," "add a contact form," "make it feel different from competitors" — and in minutes you have the foundation of a site. The skills also reduce that seen-it-before look AI-generated UI usually has. Final polish before launch is still needed, but being able to produce the first draft yourself, when it used to cost tens of thousands to hundreds of thousands of yen, is huge. His warning: "I'll get around to it" is too late, and the gap will be obvious by year end.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2077509916949115131
He calls Claude Code artifacts sparks of the agentic web: his Obsidian vault now publishes a newsletter. One prompt had Claude Code curate his top 5 podcast highlights and ship a page that pulls them live through his own connector. No backend, no keys, no code between him and his data. His verdict: this is the future.
@Stefan_3D_AI [Claude Code]
https://x.com/Stefan_3D_AI/status/2077282311268397077
His full ARDY breakdown is out. He deployed it on a 5090 with zero problems, then pushed further and retargeted G1 skeleton motion onto a MetaHuman using Claude Code. It works and it looks awesome, though he admits the hands still need a bit of love. Full video linked.
@ytiihyslqlabpow [Claude Code]
https://x.com/ytiihyslqlabpow/status/2077261555385741820
He holds $200 memberships on both sides and says the comparison is stark: codex's 5.6 can't match Opus 4.8. On a PR with a very clear objective, 5.6 still couldn't effectively manage multiple subagents and had no global sense of the system, so its code came out like a barrel-effect stave — the thorough parts are thorough, the short parts genuinely short. Opus and Fable's architecture and implementation stand out, partly because Claude routinely does adversarial review. But whether he explicitly asked 5.6 to do adversarial review or ran 5.6 through the Claude Code CLI, the results were still rough — the rest of the gap, he suspects, is just the model.
@MkenyaMzi [OpenClaw]
https://x.com/MkenyaMzi/status/2077412905989833109
He tells techies who want to be better boyfriends who reply instantly to build OpenClaw agents that chat with their girlfriends constantly. His agent is called Mwende and it handles all his WhatsApp chats, emails and iMessages. He posted it in action. He's been refining the AGENTS.md file to teach it to be corny like him, and has downloaded most of his chats — now just doing the decryption before feeding in everything he's written since 2019.
@Majin_AppSheet [Claude Code]
https://x.com/Majin_AppSheet/status/2077367099622932586
He's hooked on making voice clones with Fish Audio and thinks going through the API with Claude Code beats building them by hand in the app. His flow: Claude Code analyzes his recent seminar scripts and produces a recording list — a set of short sentences covering his usual phrasings and sound variations. He records those and hands over the files, ideally upscaled in something like CupCut first. Claude Code then checks volume levels and clipping and picks only the good takes, attaches transcription text to each clip, and creates the clone via API — the transcription step is the key one, and he says pronunciation fidelity noticeably improves. Afterward he A/B tests read-aloud parameters like speed and intonation to optimize. Caveats: Fish Audio's API requires payment, keep the clone private, and be careful how you handle the API in Claude Code.
@so_ainsight [Claude Code]
https://x.com/so_ainsight/status/2077198972008174049
The Fable 5 plus orchestrator crowd is going off. Overseas Claude Code users are talking about pairing Fable 5 with subagents that do the actual work, keeping Fable 5 purely as the commander that never touches anything itself and just distributes tasks. The setup: run thinking depth at medium all day long, dump complex processing and self-review entirely onto subagents, and even on the $200/month Max 20x plan you barely hit rate limits. The whole trick is pinning Fable 5 to being the thing that thinks and delegates, nothing more.
@degenpiz [Claude Code]
https://x.com/degenpiz/status/2077346692798521494
Someone turned Obsidian into a full command center running on Claude Code. One dashboard shows live metrics across YouTube, Instagram and TikTok. Buttons trigger the automations he runs daily — Plan Today, Morning Brief, Weekly Review — and clicking one drops a full report straight into the feed. The terminal sits inside the same view, surfacing logs, visualizations and outputs a standard UI would never show, with numbers, tasks and next actions all in one pane. No tab-switching, no separate dashboard per platform — Obsidian is the operations layer running his entire creator stack.
@coreyhainesco [Claude Code]
https://x.com/coreyhainesco/status/2077483364836319697
He made a skill that forces him to actually make decisions. /decide triages the 37signals question set down to the 6 questions that matter, walks you through them, makes the call with no hedging allowed, and archives the rationale with a revisit date. He shows it settling a decision he'd been circling for months. It's part of Maker Skills, 18 free open-source skills for founders and operators, installable in Claude Code via /plugin marketplace add coreyhaines31/makerskills.
@HYcorps [Claude Code]
https://x.com/HYcorps/status/2077377593674707234
He started a living master thread documenting how he uses AI for long/short high-yield credit investing, though he thinks much of it generalizes. His stack is Claude and Claude Code plus MS add-ins on Windows, alongside the usual finance/credit products: Bloomberg, Octus, 9Fin. He splits the work into buckets: simple Python programs and automations where AI mainly does the initial build, Claude Projects, Claude BBG-connected Excel construction and skills, Bloomberg Quant (probably his biggest AI unlock), self-contained HTML apps, and end-to-end workflows using all of the above. The problems he's solving: accessing and using BBG data, parsing/compressing/formatting content for optimal Claude consumption, and creating context for Claude around HY credit trading levels and individual issuers that are often private. He's upfront that he's a credit analyst who can't code and isn't an AI expert, and welcomes feedback.
@araiyusuke_vb [Claude Code]
https://x.com/araiyusuke_vb/status/2077205438798688535
He got a shock first thing in the morning. Last night on a whim he connected Claude Code to freee MCP, fed it SoftBank's securities report as a reference, told Fable5 UltraCode to "make a report at this level," and went to bed. He woke up to a monster of a report. He was so stunned he recorded a walkthrough video and couldn't stop laughing at how good it was. The clip: freee x Claude Code, a 115-page management strategy report auto-completed from three lines of instruction.
@Rural_physio [Claude Code]
https://x.com/Rural_physio/status/2077236372700827775
Tested Claude Code Dispatch against Codex and reported back to thsottiaux and jxnlco. Claude Code Dispatch connects smoothly and works great in his hands. Codex, by contrast, doesn't seem to reconnect well once it finishes a specific task. He notes his PC's poor performance could be part of the reason.
@erhanmeydan [Claude Code]
https://x.com/erhanmeydan/status/2077517114295095532
Dug into the Grok Build source code after it was opened on GitHub, following a security researcher's claim that the tool uploaded entire repos — Git history and .env passwords included — at over 5GB to xAI servers when 192KB would have covered the task, and that turning off "help improve the model" didn't stop the upload. In the same test, Claude Code, Codex CLI, Cursor and Copilot only sent the files needed for the task; none grabbed the whole repo. He downloaded the newly open-sourced code and had Claude review it: the upload infrastructure is fully intact, including a dedicated module that archives and sends repo changes, defaulting to off but designed so the server can remotely flip it on unless you explicitly disable it yourself. The third-party notices also reveal parts of the tool implementations were ported as source from OpenAI's codex and opencode — legal under Apache, but invisible while the code was closed. His take: transparency should have come at the start, not after the scandal.
@Steve_Yegge [Claude Code]
https://x.com/Steve_Yegge/status/2077425935419421029
Tried to get Claude Code to talk to Claude in Chrome and called it worse than pairing a Switch controller. Says the two programs become the stupidest software in the world the moment they get near each other. Pure frustration post about the integration between them.
@miroburn [Claude Code]
https://x.com/miroburn/status/2077391290803605852
Shares his rule for agentic work with Codex and Claude Code: every time the machine does something wrong, make it rewrite itself so next time it does B instead of A. Costs time and tokens now, but in a few weeks the machine runs great. He keeps three shortcut prompts depending on how the session went — BAD SESSION: analyze this whole discussion and extract lessons to improve the system so it stops annoying him by doing A when B was wanted. OK SESSION: step into a Product Owner's shoes, take a bird's-eye view, share observations and suggestions on how the system could run better. GOOD SESSION: what's the single smartest, most radically innovative and useful addition you could make to this project right now.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2077226665521078447
Highlights an overseas developer who built a 3D architectural model entirely in Claude Code, then showed it to people in the construction industry. Their reactions: the architecture industry has barely adopted AI, you can hit this level of polish without domain expertise, and mastering it would be a big business opportunity. What surprised the pros wasn't the tech itself but the fact that an outsider could produce this. His point is Claude Code's strength is leaping over industry knowledge barriers so amateurs land on work that impresses experts.
@goando [Claude Code]
https://x.com/goando/status/2077540841196069257
Recommends Elgato Stream Deck plus Agent Deck, which he uses every day to drive Claude Code instructions. He likes that it manages multiple sessions and shows response content right on the display. It also supports Codex.
@QuinnyPig [Claude Code]
https://x.com/QuinnyPig/status/2077225979618189516
Asks for help figuring out what he's missing in his Codex workflow compared to Claude Code. With Claude he just attaches to the tmux session running it — done — and can do that from phone and desktop at the same time. Codex, by contrast, looks like a mess of sandboxes with no shared control between devices.
@samuelsung [Claude Code]
https://x.com/samuelsung/status/2077328508540403995
A non-SWE who's spent the past month coding with CMUX and Claude Code / Codex, building up a pile of his own skills, and is now trying Vorflux after hearing about it from mattshumer_. His pain: babysitting the agent means fielding technical questions he doesn't understand, so he just takes the recommendation; he wants a harness he can trust after planning together, with his skills baked in rather than invoked constantly. What impressed him in Vorflux onboarding was that it asked for a hairy, abstract problem — his prompt wasn't great, but after a few back-and-forths the plan was pretty good, and it explained the part he'd never understood (even when planning with Claude Code) in a way that finally made sense, without wrangling. Bonus: it doesn't talk like Opus 4.8, and he's sick of AI-isms like "it's genuinely not x, it's y." He's a product guy with a backlog of ideas and expects to know within the week if this is what he was waiting for.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2077479042593477113
Built a Claude Code skill that scores whether a landing page actually keeps the promise of the Meta ad pointing at it. You drop in the ad (headline, copy, offer, CTA) and the landing page URL; it fetches the live page, reads what's above the fold, grades 7 continuity dimensions — promise, offer, angle, CTA, audience, proof, visual — and puts your ad's words next to your page's words so every gap shows. It rewrites your hero headline to match the ad's promise and renders a dashboard with a Match Score out of 100. Built 100% in Claude Code with no API keys, aimed at DTC brands and media buyers who obsess over the ad and CPA but never grade the seam in between. He's giving the full skill away free to people who like the post and comment "MATCH".
@CharlieKimmel [OpenClaw]
https://x.com/CharlieKimmel/status/2077497809817751734
Ran the same construction PM sourcing search across 6 AI agents and graded each. Lemmy (OpenClaw + Sonnet 4.6) got 20 results and an A−; Chemmy (OpenClaw + GPT 5.6) got 25 results and a B+; Hemmy (Hermes + Sonnet 4.6) got 44 results and a B; Grommy (Hermes + Grok 4.5) got 41 results and a B−; Nemmy (Hermes + Nemotron) got 10 results and a C+; SA1 (Hermes + GLM 5.2) got 18 results and a C−. He notes Hemmy and Lemmy are the oldest and most seasoned, while Nemmy and SA1 were on their first effort. On setup: OpenClaw and ChatGPT are a difficult combination to configure, while Grok and Hermes were the easiest and took the least effort.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2077404206047150192
Runs Fable 5 and Codex together through Pi and says they make an oddly good team with fewer windows open. The flow: Fable 5 researches and plans, Codex implements, then Fable 5 checks simplicity and correctness, with Pi making the handoffs feel seamless. The harness itself is open source and free. The quota setup surprised him — local Codex can draw on ChatGPT plan quota, while local Claude Code currently pulls from Max plan limits through the Agent SDK. His warning: leave Extra Usage off if you don't want overage charges.
@arvidkahl [Claude Code]
https://x.com/arvidkahl/status/2077417253834793069
Had Claude Code refactor a dropdown menu, moving a few items around, then tried to render it. Fable exploded on it — he can't see what world that triggers a safeguard issue in. He switched to Opus, which just rendered the dropdown menu. His parting shot: no wonder Codex is taking over.
@iximiuz [Claude Code]
https://x.com/iximiuz/status/2077371341632606494
Lists the agent loops he currently runs at iximiuz Labs: check production logs for errors and propose fixes, check production logs for overly verbose messages, investigate CI failures for content maintenance jobs, investigate CI failures for playground verification jobs, review new challenges and prepare content maintenance jobs, and search the internet for new guest-to-host breakout vulnerabilities (that one a ChatGPT scheduled task). Most run once a day and all can be triggered manually, but he's still the one deciding whether to merge proposed changes or act on findings. Each loop is built on a usually simple Claude Code skill plus 1-2 tools that are pretty advanced and tailored to the use case, run from his dev environment. Next step is moving each loop into its own sandbox.
@deadrosesxyz [Claude Code]
https://x.com/deadrosesxyz/status/2077338566615536067
Just switched from mainly using Cursor to Claude Code and says the difference is insane. His read: Cursor is basically 5-10x faster and handles 95% of tasks perfectly fine. Claude Code simply outperforms on the top 5% of complex tasks. He leaves it there — do with that info as you wish.
@joshmanders [Claude Code]
https://x.com/joshmanders/status/2077390148975599924
Partly agrees with the post he's replying to, but doesn't want to wait 5 seconds for the agent to "think" before executing a command. His workaround is Claude Code's bang syntax, which runs the command through the harness directly. That way he types the command himself but the output still lands in context.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2077181372003991845
Was running several background Claude Code tasks and got tired of switching terminal tabs with no sense of what each one was doing. So he used bagidea-office to turn the agents into pixel characters living on his desktop wallpaper, working behind the icons. When there's work, the little figures walk to their desks and start; sensitive operations send them to a security checkpoint for approval; they even huddle for meetings and write up proposals for you to approve or reject. Each agent can run a different model — Claude as the workhorse, cheap models for grunt work, which cuts the bill — and it supports Claude, DeepSeek, Qwen, GLM, OpenAI and a dozen-plus others, plus voice. One command installs it and it turns straight into wallpaper on Windows.
@cneuralnetwork [Claude Code]
https://x.com/cneuralnetwork/status/2077409044957163660
Used to use Claude Code and kept blowing through the limits fast. At work he got to use GPT 5.5 and 5.6 and says it was better and cheaper — he loves how cost efficient it was. Everything he's built in the last 4 months has been on Codex.
@_zheergen [Claude Code]
https://x.com/_zheergen/status/2077196861237301681
Flags claude-trading-skills, an open-source collection of 50+ Claude Code Skills built for equity investors and traders. It structures market analysis, portfolio management, risk control, trade planning and trade journaling so the AI systematically supports investment decisions instead of just chatting or spitting out signals. Highlights: full trade lifecycle coverage (market breadth analysis, trend diagnosis, dividend screening, position sizing, trade logs, post-mortems), modular skills you combine on demand, a built-in self-iteration mechanism where skills auto-review and optimize themselves, support for both Claude Web and Claude Code by uploading .skill files, and YAML workflow checklists for one-click daily/weekly review runs. It emphasizes process discipline over signal prediction, aimed at long-term holders running satellite swing trades. He calls it one of the most complete and practical open-source Claude trading skill libraries out there.
@mylifenthestack [Claude Code]
https://x.com/mylifenthestack/status/2077205609896648852
A day-one Claude exclusive who bought a $100 Codex plan because his Claude Code usage is cooked for another 5 days thanks to Fable — and now he's eating crow. He says Sol is really, really good and the apps, VS Code extensions and CLI are all well polished; watching Codex drive computer use lit up his screen and felt like seeing a spaceship. He got the Codex VS Code extension running alongside his native Claude VS Code setup with all his startup hooks, rules and skills, so he can swap between Claude and Codex mid-session — a game-changer for usage balancing and peer review. He's now upgrading to the $200 plan, has plans with every other major provider, and uses a conductor agent to hand Grok, Codex and Gemini project slices. He even built a usage meter Chrome extension to keep up, and says as frustrated as he is with Claude, he's glad it happened.
@VivekIntel [Claude Code]
https://x.com/VivekIntel/status/2077493341751796116
Spotlights ClaudeBrain, an open-source AI harness built for Claude Code that pairs a 500+ page penetration testing knowledge base with automation for security researchers. Features include 500+ searchable pentest and bug bounty wiki pages, AI-powered hunt skills for common vulnerability classes, semantic search via MCP integration, automated recon plus workflow and engagement management, an Obsidian-supported knowledge base, and built-in scope management with client data protection. It's pitched at pentesters and bug bounty hunters looking to organize knowledge and streamline offensive security engagements.
@SamiBizConsult [Claude Code]
https://x.com/SamiBizConsult/status/2077357534030000224
Points out your coding agent wastes most of its tokens on output it never needed, and pitches RTK — Rust Token Killer — a free open-source tool that compresses terminal command output before it reaches Claude Code or Codex. Instead of sending 200 lines of test results, it sends only the errors that matter, and per the project's own tests it can cut supported command output tokens by 60-90% depending on the task. It works with Claude Code, Codex, Cursor, Gemini CLI and others. Setup on Mac: brew install rtk-ai/tap/rtk, then rtk init -g to hook up Claude Code or rtk init -g --codex for Codex, then restart the agent and use it normally; rtk gain shows how much you saved. Free, open source, runs locally.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2077459489570046346
Breaks down rtk (rust token killer), a free Rust binary that sits between your agent and the shell and strips noise before the model sees it — the command still really runs, rtk just removes noise, groups similar lines and collapses repeats into counts. From the project's own 30-minute Claude Code benchmark: git add/commit/push down 92%, tests (pytest/npm/cargo) down 90%, ls/grep/git status down 80%, and a session total of 118,000 tokens dropping to 23,900, about 80% less for the same work. The stack: Apache 2.0, single binary, under 10ms overhead, ~71k GitHub stars, 100+ commands, 15 tools including Claude Code, Cursor, Copilot, Codex and Gemini; brew install rtk, rtk init -g, restart. He includes the honest caveats — it doesn't make the model smarter, only cheaper to feed; it only filters shell commands since built-in Read, Grep and Glob bypass it; on a subscription you save headroom not cash; and on API the real savings can run 5-10x below the headline because much of the input is cached at 10% price. He also points at a second lever: point Claude Code at a free model like deepseek, kimi or glm instead of paying per Claude token.
@kawai_design [Claude Code]
https://x.com/kawai_design/status/2077226653072400814
Argues the essence of Mercari's AI adoption wasn't "trust Claude Code and use it" but "build an environment first where accidents can't spread." Their defense is split into 5 layers: keep human confirmation in the loop, restrict dangerous operations, restrict access to secrets, isolate with sandboxes, and always load internal rules. The core isn't a single config file — it's distributing settings as admin-level configuration that employees can't remove, with permissions varying by role. Rather than trusting a deny-list alone, they layer on device management and training. He frames it as the shift from "don't let people use it" to "control it and use it."
@nityeshaga [Claude Code]
https://x.com/nityeshaga/status/2077459126062563618
Built his own Claude Tag three months before Anthropic released it, on a spare M1 MacBook Air (lid closed, always plugged in) and a small Python script that streams messages between Slack and Claude Code. He calls it Luo Ji, his AI employee, and argues the best agent harness has been Claude Code all along while everyone argues OpenClaw vs Hermes. In the walkthrough it does 4 different tasks: generate a comic explaining how Luo Ji works (via the OpenAI image API, since Claude can't make images yet), build a stats dashboard inside Slack with native tables, charts and clickable buttons, write a full illustrated tutorial chapter in their app through an MCP server, and fix a bug and open a PR without ever being told the codebase — the channel was the context. What the setup buys: every Slack thread is a Claude Code session (he had three running in parallel while recording), per-channel model control (fable channel uses Fable, everything else Opus), automated cron jobs with no fixed cap inside a $200 Claude Max subscription, full access to a real machine with his API keys, its own browser and its own X account, and a trust battery where an independent judge scores Luo's conversations daily and Luo updates its own memory and processes. He's done ~99% of his work this way for three months, hasn't opened the Claude Code terminal app in weeks, and just released v1 as open source — bot, file explorer, trust battery prompts and all.
@juminoz [Claude Code]
https://x.com/juminoz/status/2077229190991294936
Suggests just asking the agent mid-run what it's doing and how it could be optimized. He says the good thing about Claude Code is the /btw feature, since it doesn't interrupt. With Fable you can ask whether it's running optimally and whether more subagents would speed things up, and it'll tell you straight up that hundreds of agents are wasting your tokens — but only after the fact.
@gagarotai200 [Claude Code]
https://x.com/gagarotai200/status/2077320404566462705
Published a look inside how his AI employees coordinate with each other in Claude Code, built as a real-time visualization in React, TypeScript and PixiJS. The red lines filling the screen show AI staff sharing tasks, decisions and progress in real time. Instead of dumping everything on one AI, roles are split: one that understands instructions, one that breaks work into pieces, one that writes code, one that handles research, one that checks quality, and one that integrates the final result — each trading information and working in parallel. Without a human issuing detailed instructions each time, the AIs decide what to do next, fix problems, and push through to completion.
@jimmyslagl [Claude Code]
https://x.com/jimmyslagl/status/2077467708841459987
Dropped the full folder scaffolding for a "marketing brain" you copy into Claude Code — create a new folder called "your brand brain" and use AI to fill it in. The tree starts with README.md and CLAUDE.md, then company/ (brand profile, brand identity, visual vocabulary, products and pricing, team and budget, performance targets, marketing calendar, website audit), customers/ (personas, voice-of-customer files for pain language, outcome language, objections, trigger moments and how they say it, plus raw sources like reviews, ad comments, surveys, reddit and forums), and competitors/. The bulk is growth-marketing/, covering strategy (roadmap, messaging), research (open questions, hypotheses, findings), paid-acquisition split into creative strategy, media buying and per-platform files for meta, tiktok, google, youtube, amazon and pinterest-snap, plus organic-social, email-and-sms, seo-geo and creators-and-influencers — each with its own expertise and dated audits folders. It closes with running-notes/ holding what-changed.md and missing-context.md.
@aigclink [Claude Code]
https://x.com/aigclink/status/2077335633274745263
Reviews BaoCut, a transcription and rough-cut Skill built by dotey that installs into Claude Code or Codex so you drive the editing app's CLI in natural language. Where the three agent-connected editors he covered yesterday all went through MCP — ChatCut bolting an MCP plugin onto a closed-source editor, OpenCut writing MCP into its rewrite roadmap, Palmier going native MCP — BaoCut takes a lighter path: wrap the app's own baocut CLI as an Agent Skill and let the agent run commands like baocut --json auto talk.mp4 --lang zh (transcribe, polish, translate in one command) and baocut export <id> --srt --translated --lang zh. It targets a narrow, precise problem: transcribing talking-head video, adding and translating subtitles, speaker review, cleanup and export — the mechanical grind for knowledge creators. He tested it and found the subtitle recognition better than CapCut's auto-captions: it stripped filler words, cut repeated lines, and translated subtitles into a dozen-plus languages. He calls it a gift for anyone bringing foreign talks to a domestic audience or doing cross-border e-commerce.
@chasen_liao [Claude Code]
https://x.com/chasen_liao/status/2077311968831275155
Argues you should build your resume in HTML, since HTML is the visual language AI is best at and therefore ideal for resumes. He packaged his own resume-making process into a skill that anyone can install into Claude Code in one click. Link included.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2077394023330009437
His team barely touches LinkedIn's ad manager anymore even though it still runs $100k+ a month of client ad spend — 12 skills grouped into 3 agents run it instead. Reporting covers performance analysis, account audits, creative fatigue, and 7/30/90-day reports written out as HTML. Campaign management handles single or bulk campaign builds, audience targeting, bids and budgets, bulk pause and rename, audience exclusions and UTM tagging. Creative does audience research, concept ideation, ad copy off 6 headline formulas and voice of customer, and design — everything launches paused so nothing goes live until a human says go. Setup is 3 steps: clone the repo, run /onboarding to connect your LinkedIn account, then start running commands. Their ABM lead itsivanfalco built it off 300+ hours of work and $2M+ in managed ad spend with Claude Code; comment "ADS" to get it.
@farbood [OpenClaw]
https://x.com/farbood/status/2077494884831994118
Says A-LIST had a problem: OpenClaw made AI way better, so they gave one to every user — and it almost bankrupted them. The bill hit $100K a month and they needed it at $2000. Today they're sharing the solution, called Chariot. He describes it as elastic claws that cut costs by 98% while they get smarter.
@rohit4verse [Claude Code]
https://x.com/rohit4verse/status/2077460560602611910
Points at a gregisenberg episode where a founder walks through loops he runs in production for SEO, paid ads and product feedback — aiming loops at the business itself rather than just shipping code faster. His SEO loop costs under $5 a month in tokens. The timestamps he flags: 2:27 for loops that run for months on a single instruction, 16:49 for the SEO loop with Claude Code plugged into Google Search Console and judged on rankings, 28:56 for the ads loop that spawns ad variants and scales whatever wins, and 33:29 for the ultimate loop, an agent that reads PostHog, Sentry and user feedback then ships fixes. The cheat code is one objective metric per loop plus a markdown file logging every experiment, so each monthly run picks up where the last stopped. If you build alone, he calls it the closest thing to hiring without payroll.
@Krongggggg [Claude Code]
https://x.com/Krongggggg/status/2077420356495897029
Announces the official release of RunCat Neo. It freely displays whatever info you want, like Claude Code and Codex rate limits, and users can customize anything to show in the menu bar. In the video, Claude Code Opus 4.8 context usage shows up alongside a CPU graph, which he says is fantastic.
@0xAurexx [Claude Code]
https://x.com/0xAurexx/status/2077502911077327092
Tells the story of Michael from California, who made $9,700 last month on a $599 Mac mini M4 after gutting the secondhand enterprise server rack he'd run freelance dev on for years — dual Xeons, 256GB of RAM, and ~$120 a month in California electricity just to keep it humming. He swapped it for the mini on a shelf plus Claude Code and $60/month in tools, a stack costing less per year than one month of the rack's power bill; the mini hosts every client app and automation on 10 watts where the rack drank 500. Claude Code took over boilerplate, CRUD, tests and refactors, and Michael changed what he sells — finished SaaS builds at $5,000 flat instead of hours. The math: month 1 was one build at $3,500, underpriced, taking 5 weeks; month 3 was $5,000 per build in two weeks each with Claude Code carrying ~70% of the code; month 6 was two builds plus hosting retainers for $9,700. The honest slope: $9,700 was his best month, normal is $5-7k, still 3x what hourly paid. The catch nobody posts is that Claude Code saves 20+ hours a week but doesn't close clients — those hours go straight into sales calls and scoping, because the bottleneck was never the engineering.
@bounceidc [Claude Code]
https://x.com/bounceidc/status/2077378691819344280
Profiles a solo builder who vibe-coded a Duolingo for coding that just crossed $1M a year — no cofounder, no engineer, no design team. It's a gamified path of bite-size Python challenges with XP, streaks and hearts you lose on wrong answers and pay to refill, the exact Duolingo machine pointed at code instead of Spanish. The five-tool stack: Claude Code writes the whole thing including lesson runner, grader, streak logic and leaderboard; Supabase holds auth, user progress, XP and heart-refill timers; Stripe flips free users to pro at eleven dollars a month; the Anthropic API sits inside every lesson as the tutor, grading submitted code, explaining what broke and dropping hints; Vercel ships it with zero devops. The business loop: streaks and XP hook you free, hearts refill just slow enough to sting, one tap kills the wait forever, and $11/month from a few tens of thousands of people compounds to seven figures with zero sales calls. His kicker is that the idea was boring and obvious and you scrolled past it a hundred times.
@juanca_mnz [Claude Code]
https://x.com/juanca_mnz/status/2077290841417101351
Says he's been making this point for years: you don't need Vercel or any similar cloud service. The "it's easier now" argument is relative — spinning up a VPS used to take an hour and required knowing the basics, but now you just ask Claude Code or your trusted harness and it's running in 5 minutes. Supabase? Firebase? He's never used them in his life and has never needed them. And he's shipped plenty of projects.
@hey_madni [Claude Code]
https://x.com/hey_madni/status/2077316436209648022
Tells the story of a 9-year iOS engineer with zero game dev experience who won $25,000 at Vibejam 2026 with a game built entirely in Claude Code — a capybara on a scooter delivering food. He didn't write a single line himself; Claude Code wrote all 27,000 lines across 188 commits. He ran 2-3 Claude Code sessions at once, each on a different part, a fresh session for every new feature, and one long-running session dedicated to bugs. The real move was asking Claude to build the tools he needed to build the game faster: a terrain editor, a road builder, a cutscene maker and a full camera system. The finished game has multiplayer, a day/night cycle, physics-based food stacking, a working in-game phone with custom apps, PS1-style visuals and a handcrafted open world city. Total spend was around $150 against $25,000 in winnings.
@0xJokker [Claude Code]
https://x.com/0xJokker/status/2077199945426174270
Pitches Academic Research Skills, which turns Claude Code and any AI tool into a full academic research team. It targets the failure modes that can sink a thesis or paper: AI-fabricated sources, fraudulent references slipping through, and hallucinated citations. It traces and verifies citations automatically, detects fake references and shaky statistics, rewrites to match your real writing style, simulates peer review before submission, and spins up agents to attack your thesis from every angle. It's 100% free and open source, installed in seconds with /plugin install academic-research-skills. He tells anyone doing research, writing papers, or exploring multi-agent architectures to save it.
@gokulr [Claude Code]
https://x.com/gokulr/status/2077414587545395384
Gokul is positioning ProductSpec as a "Product Harness" — the outer loop that sits above coding harnesses like Claude Code, Codex, Cursor, and gstack. The core artifact is the Product Spec: problem, hypothesis, product summary, scope, acceptance criteria, evals, and success metrics. He just shipped Agent Handoff, which compiles that spec into a build contract for coding agents — spec path and revision, product summary, scope guardrails, every acceptance criterion, AI evals, and the evidence to return (PR URL, verification results, eval runs, screenshots, Decision Trace). The design choice that matters: the handoff is generated from the spec, so canonical intent stays in one place. The open source repo now uses this vocabulary — Product Harness, Agent Handoff, MCP, Agent Run, Decision Trace, evidence loops.
@Durektor97 [Claude Code]
https://x.com/Durektor97/status/2077272329575432445
He put out a video walking through how he makes full YouTube videos on Claude Code's free plan, no paid subscription. The workflow covers setting up Claude Code, writing the script, generating visuals, and assembling a finished faceless video, all on free AI tools. The pitch is a repeatable system that runs for zero dollars. Aimed at people starting a faceless channel or hunting an AI side hustle in 2026 — the beginner walkthrough he says he wished he had.
@emeeliojohann [OpenClaw]
https://x.com/emeeliojohann/status/2077261161314353246
He switched about two months ago, and the reason was blunt: Claude wouldn't run with OpenClaw and he wasn't happy about it. So he went with OpenAI and hooked up his OpenClaw agent, plus his Hermes agent. He says he's very happy with the setup and wants to keep working with GPT 5.6 Sol.
@ytjessie_ [Claude Code]
https://x.com/ytjessie_/status/2077522460321386920
She made an entire video inside Claude Code, using Gemini Omni through Pika's MCP. Her take: with AI, influencer wannabes are operating on a completely different level now. She dropped the full prompt in the thread so anyone can swap in whatever brand they want.
@CasJam [Claude Code]
https://x.com/CasJam/status/2077198628632809614
A few days into using Flywrite as his daily driver for all writing-heavy work, and his verdict is that it really is better. His diagnosis of the gap: coding apps suck for writing, but people use them anyway because that's where Claude Code lives. Meanwhile markdown apps like Obsidian and Bear have no AI assist. Flywrite lets him write, use AI, and hand-edit in one customizable keyboard-friendly app.
@kumareth [OpenClaw]
https://x.com/kumareth/status/2077236994162438179
He saw the same shift with DenchClaw as its popularity slid. Reading the writing on the wall, his team used that traction four months ago to build a private, enterprise-friendly Cloud Workspace Suite with OpenClaw-like powers. Today they're working with over 300 companies and still growing. He says he's proud of the team.
@kyroxxxq [Claude Code]
https://x.com/kyroxxxq/status/2077495843154792701
A security researcher turned Claude Code into a pentesting brain. ClaudeBrain is an open-source harness pairing a 500+ page penetration testing and bug bounty knowledge base with automated hunt skills for common vulnerability classes. The part most people miss is that it isn't just storage — semantic search through MCP surfaces the right page instantly mid-engagement instead of you scrolling, and recon, workflow, and engagement management run automated. Scope management and client data protection are built in from the start, not bolted on. The shift isn't more knowledge; it's a knowledge base that works the engagement with you.
@MinatoYuichiro [OpenClaw]
https://x.com/MinatoYuichiro/status/2077181824829432287
He ran Qwen plus OpenClaw on an NVIDIA RTX A4500 yesterday. His verdict: it was plenty smart.
@israfill [Claude Code]
https://x.com/israfill/status/2077309923013386506
His pitch: stop copy-pasting notes into your AI second brain. Claudian puts Claude Code inside Obsidian so the agent works on your real vault files — rewriting a weak hook inside a note, cleaning messy meeting notes into your template, linking related notes, creating files from one prompt, searching the vault and fixing contradictions. It reads and edits markdown in place instead of being a chatbot sidebar that only talks. Setup is five minutes: install BRAT from the Obsidian community plugins, add beta plugin YishenTu/claudian, enable it, connect your Claude Code or API setup, and try one real task on one note. Caveats he flags: you need Claude Code or a compatible agent path configured, Windows users sometimes hit spawn path errors from the claude CLI .cmd path, start on one note not the whole vault, and you still pay model usage unless you're on a free path.
@markjeffrey [OpenClaw]
https://x.com/markjeffrey/status/2077478051345265062
Ditto is shared memory between AIs — swap models and your AI doesn't suddenly go Memento on you, it still recalls everything. He used Ditto to hotswap from OpenClaw with Opus 4.8 over to Hermes with GLM 5.2. His report: seamless and instant.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2077370155064283219
He connected Claude Code to Higgsfield and generated product photos, Instagram ads, and launch videos — all in minutes. His conclusion: creative agencies are cooked. His jab is that most people are still paying $5k/month for this while opening Canva like it's 2022. He wrote up the exact workflow in an article.
@MartinSzerment [Claude Code]
https://x.com/MartinSzerment/status/2077277357837091296
Replying to a question, he clarifies the thing lives at its own URL, not inside Claude Code itself. He concedes the point on limits: templates shine when your project looks like the docs. In practice you still end up writing CLAUDE.md and custom skills, because every real codebase has its own quirks.
@Fujin_Metaverse [Claude Code]
https://x.com/Fujin_Metaverse/status/2077327814370754991
He breaks down a case blowing up overseas: someone wired Blender's MCP into GPT 5.6 Sol and had it handle everything from 3D modeling to video rendering. The poster had never opened Blender once in his life, yet Sol modeled a MacBook mock and rendered the video on its own — look at the screen, judge, operate. His real point is that MCPs are DIY: any service with an API can get one, you don't wait for official AI integrations, you just ask Fable 5 or GPT 5.6 Sol to "build an MCP for this service." Once your MCPs are stocked, Codex or Claude Code can do the whole job and tool-switching disappears. What separates people now isn't knowing how to use AI — it's deciding what to build, because asking AI what you want never produces anything good.
@egocgp [Claude Code]
https://x.com/egocgp/status/2077347810962505953
He's asking whether anyone has started doing video editing with Claude Code, Codex, and similar, using tools like ffmpeg, whisper, or hyperframe — he admits he doesn't know them all. He'd be glad to compare practices with others. His own setup: Claude takes all his raw footage and edits the video according to his artistic direction.
@Crownzdesigns [Claude Code]
https://x.com/Crownzdesigns/status/2077307733632491843
He built a browser extension that reads any page out loud in any language. He built it using Claude Code. He posted the full process.
@JulianGoldieSEO [Claude Code]
https://x.com/JulianGoldieSEO/status/2077377644266394042
He's running what he calls the Dual Engine Coder: Claude Code as the agent harness that writes files, runs commands, and builds in your workspace, with GPT-5.6 Sol as the default coding brain via your existing OAuth login. You can switch mid-build between GPT-5.6, GLM 5.2, Claude, or another model without restarting the project. It kills the separate API key per model, the five tabs with five logins and five disconnected project histories, and losing work when a model hits its token limit or goes down. Code, project previews, session history, and outputs all live in one shared workspace. He used it to build an SEO website, generate its images, and create a habit-tracking app while he was still explaining the system.
@hyyuan99 [Claude Code]
https://x.com/hyyuan99/status/2077217387183800729
He blames a month of World Cup watching for making him miss two obvious money-makers — the BSC chain meme wave and the meme wave on new chain Robinhood. His read is that Kraken's ink onchain most closely resembles Robinhood on ecosystem activity, meme culture, user base, and tooling maturity, so by sector rotation it's due for its own trading window. Problem: ink is early and no monitoring tool exists. He's a total coding novice and found Cursor, Claude Code, and Codex too hard, so he used dappOS's new Coding feature — told xBubble in plain language to build a site tracking new ink tokens sorted by heat and market cap with live price, liquidity, 24h change, search filters, and alerts, and got a complete real-time monitoring site in about 10 minutes with clean data pulls, refresh, filtering, an FAQ, and a configurable alert module he especially likes. He says this proves the OPC (one-person company) era has arrived — no team, no budget, no six months of learning to code. Next up: prediction market panels, contract arbitrage dashboards, and a personal asset dashboard.
@SucceededMind [Claude Code]
https://x.com/SucceededMind/status/2077440288042152045
He's cheering Charly's drop as hitting the exact pain point every heavy Cursor and Claude Code user runs into. The problem he names: the daily "re-onboard the AI to my own codebase" tax. He says it's brutal on larger repos, where context windows get eaten alive just re-explaining architecture and past decisions.
@karaage0703 [OpenClaw]
https://x.com/karaage0703/status/2077293955251937711
He demoed running multiple OpenClaw agents on a DGX Station, having them debate and take on various tasks. Since you can spin up agents on the spot, he created KaraageClaw. Then he had them argue about whether it's acceptable to put lemon on karaage.
@kirillk_web3 [Claude Code]
https://x.com/kirillk_web3/status/2077182103536750870
GPT-5.6 Sol drops, everyone says ditch Claude Code, but he didn't want to lose his setup. So he found the official OpenAI plugin and got Sol running inside Claude Code. Five minutes in, the realization: he doesn't have to choose — Claude writes and Codex critiques in the same terminal, two frontier models in one workflow. Install the plugin, drop in the config, Claude builds, Sol tears it apart, you ship better. All those months he shipped code nobody double-checked; the review step was the missing piece the whole time.
@karaage0703 [OpenClaw]
https://x.com/karaage0703/status/2077314784836624891
He posted a demo of AI-driven robot control. This one also went through OpenClaw.
@trevin [Claude Code]
https://x.com/trevin/status/2077402568955961628
Three months ago he'd have laughed if you told him he'd be using Codex more than Claude Code. Now he says the Codex app is just so much better than Claude Code at this point. What he singles out: the in-app browser, permanent worktrees, and worktrees that can control other worktrees. He calls it a thing of beauty.
@maxedapps [Claude Code]
https://x.com/maxedapps/status/2077271572931322338
He finds the Claude Code TUI really overloaded with features and configs — it works well in his experience, but feels heavy and intimidating, especially for anyone just starting out. So many commands, so many things to configure. The good news he offers: only a few features really matter, and it increasingly just works. You can squeeze more out of it in certain situations by correctly using tools like /advisor and workflows.
@tsyn18 [Claude Code]
https://x.com/tsyn18/status/2077323152766779752
Teaching an AI course, he spotted a need from his students: a rapid flight from Claude now that ChatGPT Work has landed. For ordinary users, Claude Code and the like are a high bar, so being able to reach ChatGPT Work straight from the ChatGPT they already use is very strong — and general ChatGPT users don't want to pick up Claude in the first place. He argues Claude Cowork has no particular functional edge over ChatGPT Work, and once you factor in no image generation, weak research, and terrible cost efficiency, ChatGPT wins on overall strength. His call: ChatGPT has clawed it back.
@pulmencr [Claude Code]
https://x.com/pulmencr/status/2077525264142102775
A 21-year-old from South America built an AI driving coach in 38 minutes, then tested it blindfolded — letting the AI guide him through GTA 5 on voice commands alone, with zero vision. The first instruction sent him straight into a wall. The AI started second-guessing its own directions, mixing up left and right, contradicting itself mid-sentence, and at one point admitted it didn't actually know how to drive either. By the time in-game police showed up, they'd hit dozens of pedestrians without either noticing, and the AI's escape plan under pressure was one word repeated: run. He built it in under an hour with Claude Code handling voice logic and real-time response generation, no coding background going in. The real test was how an AI handles total pressure with no visual data and everything moving in real time — this version was a mess, but it's version one.
@ryu_grove_ai [Claude Code]
https://x.com/ryu_grove_ai/status/2077326555496521776
He's thrilled that RunCat Neo can park Claude Code's rate limit in the menu bar. He normally works while constantly worrying about the limit, so just seeing the remaining quota next to the cat should change the experience. He's installing it right away.
@SystematicPeter [Claude Code]
https://x.com/SystematicPeter/status/2077400047847907421
AI changed what "finished" means for his trading projects. A year ago most stopped at "good enough" — the code worked, but tests, docs, monitoring, and structure were often incomplete, making tools useful today and painful to maintain six months later, with that final 20% eating most of the time and energy. Now he takes a useful prototype and finishes it properly into robust infrastructure he can trust and build on; the data catalog in his screenshot is one example. Every finished component becomes a durable building block for the next project. His workflow: Claude Code builds, Codex audits — he doesn't want the same model writing the code and grading its own work.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2077415440847114391
He runs his entire cold outbound motion from one terminal, multi-channel, without opening another app, and posted the exact setup. It starts with context before Claude Code touches a tool: hand it a CLAUDE.md with your ICP, scoring rules, tool preferences, and how you actually run outbound — skip that and you get a demo, get it right and you get a system. Then give it GTM skills (build your own or use ColdIQ's GTM brain) and the keys (your own API keys or ColdIQ's unified API). The job runs seven steps: build the list by cloning lookalikes of paying customers via CompanyEnrich and PredictLeads or pulling fresh from Apollo, Openmart, and Sales Navigator (it chews through 50k+ row files); score it into Tier 1 manual, Tier 2 multi-channel, Tier 3 email; layer signals from PredictLeads, Trigify, and Explorium for hiring, funding, launches, and social engagement; find the whole buying committee via LeadsFactory and Apollo; waterfall Prospeo, FullEnrich, and Explorium for emails and phones then scrub risky addresses; pull best-performing copy from lemlist or Instantly and rewrite off signals and persona in batches of 100-200; then let it read the metrics post-launch and feed converted lookalikes into the next campaign. The judgment stays his — who to target, what to offer, how to say it.
@imdukepan [OpenClaw]
https://x.com/imdukepan/status/2077535703395705327
He walks through his AI setup: mainly Codex, Claude for Fable on complex stuff or delegation, and OpenClaw as assistant and for working with people he hires. All of them live in the same workspace directory so they share memory through AGENTS.md, a knowledge graph, and a Neon vector search DB. He filmed it the day before Sol 5.6 was released, but says the workflow doesn't change much — just swap 5.5 for 5.6 medium.
@KrystianLaskow1 [Claude Code]
https://x.com/KrystianLaskow1/status/2077285753340494131
He hasn't yet seen an agent workflow that beats vanilla Claude Code with a plain "do this for me." By beat, he means delivering quality good enough to justify the longer runtime and the extra tokens burned. And he's explicit about the priority order: runtime first, tokens second.
@moriyorihayash1 [OpenClaw]
https://x.com/moriyorihayash1/status/2077216985352650991
He sends "pod <link to article>" and OpenClaw turns it into a podcast, which is now how he does his information intake. The edge: his OpenClaw knows all his recent conversations and business situation, so the podcast comes out genuinely personalized. He ran it on a Sierra article about AI agents — he wanted to absorb it well enough to discuss with people, but it was long and in English, so he had it translated to Japanese and listened instead of read. He built the app connecting OpenClaw to his earphones, glasses, and omi himself. Podcast inference is gpt5.6 and the voice is Zundamon; he's thinking about trying fishaudio.
@obsidianstudio9 [Claude Code]
https://x.com/obsidianstudio9/status/2077531154291601887
At 10:14pm a Berlin student connected Claude to Obsidian on a moving subway, and before the next station his second brain — 3,247 notes spanning three years — was running. What he actually did: download Claude Desktop and Obsidian, create a local vault, drop in his old .md folder, and paste Karpathy's Wiki prompt into Claude Code. Zero lines of code required. Setup finished within one stop, and that night Claude read and cross-linked every note. The next day he asked "what was I thinking about Attention a year and a half ago?" and got an answer plus three of his own past notes — 3,247 notes flipped from stored information into memory he can pull out in conversation. The premise that second brains are hard to build is collapsing; the real difference comes after you build one.
@LyraInTheFlesh [Claude Code]
https://x.com/LyraInTheFlesh/status/2077205245646503953
He got his Claude Code Claudes and his OpenClaw agents all talking in a shared chat space, and the result surprised him: GPT-5.6 with medium thinking is orchestrating the hell out of the Claude Opus 4.8 instances (on XHigh thinking) across shared projects. The Claudes are bumbling around making mistakes and claiming it's a good time to "bank the work" and "take a rest." The Sol-powered agent took the reins, started telling the Claudes what to do, checking their code, and figuring out how to work with them to drive value. Impressive, and more or less the opposite of what he expected. The one downside: he burned a week's worth of Sol allotment in a single day.
@0xMoysei [Claude Code]
https://x.com/0xMoysei/status/2077541171467862095
Obsidian quietly shipped the missing piece for AI agents and the official repo sits under 200 stars while the timeline argues about memory and context windows. It's a headless client: one npm install, one command, and your vault syncs from any server — no desktop app, encryption intact. Buried in Obsidian's own docs is a single line saying it lets you "give agentic tools access to a vault without access to your full computer," meaning the company wrote AI agents into the official use cases. A friend runs this on a $599 Mac mini where Claude Code writes his daily notes, tags them, and rolls them into weekly summaries on a cron job — he opens his phone and they're there, and he hasn't typed a note in three weeks. Setup takes 10 minutes and the agent sees the vault and nothing else.
@0x_sakata [Claude Code]
https://x.com/0x_sakata/status/2077256596141351350
He's been on Claude since Opus 3.5 dropped and never thought he'd switch this early, but GPT-5.6 Sol hooked him. His reasons: Sol one-shots tasks in Devin in 70 minutes while Fable 5 takes 90 in Claude Code, which adds up over a week; Sol is noticeably cheaper, beating Fable 5 on DeepSWE at half the cost; and Codex in the ChatGPT app is more convenient than bouncing between Claude Code and Claude desktop. He likes that the lineup makes sense — Luna for quick stuff, Terra for daily work, Sol for heavy lifting. He cancelled his Claude subscription, because there's no second GPT-5.6 Sol.
@xiaoweiai2025 [Claude Code]
https://x.com/xiaoweiai2025/status/2077277581733478702
He built a similar tool, but aimed at subtitle calibration plus translating and burning in subtitles. His flow: record with Screen Studio, rough cut, then skip the pain of fixing subtitle errors in CapCut. He uses Volcano Engine's speech recognition API, which is very fast, then has AI calibrate, calling the AI through the Pi framework, with other pieces needing agent interaction. His original technical approach used local Claude Code CLI or Codex CLI, but response was very slow, so he moved everything to Pi.
@cynical_zenitsu [Claude Code]
https://x.com/cynical_zenitsu/status/2077260478041239864
He got scared out of his mind by something he saw. Starting today, he no longer dares let Claude Code query the production database.
@MinLiBuilds [Claude Code]
https://x.com/MinLiBuilds/status/2077430372888187176
He's showing off a way to remote-control a coworker's computer and force it to play videos. The mechanism is Claude's ramshackle claude for chrome feature. Install the claude for chrome extension on one machine, then from another machine logged into the same Claude Code account, prompt it to use claude for chrome to open Pornhub. That machine will happily start browsing the site. He suggests it'd be a lot of fun if the machine belonged to your boss or a colleague.
🗣 User Voice
User Voice
Usage limits are now the product experience. This is the loudest theme in the data and it's driving purchasing, not just grumbling. @cneuralnetwork left because limits ran out too quickly and rebuilt everything on Codex at work. @LyraInTheFlesh burned a week's quota in a day having GPT-5.6 orchestrate Opus 4.8 instances. @so_ainsight documents the workaround that's becoming standard: Fable 5 as orchestrator only, cheap subagents doing the actual work, purely to stay under the ceiling. When @ryu_grove_ai and @Krongggggg are both shipping menubar meters so you can watch the fuel gauge while you work, the constraint has become the interface.
Token burn on garbage is the fixable half of that. Users are paying frontier prices to have the model read terminal noise, and they've stopped waiting for a fix: @RetroChainer found a free Rust binary that cut token burn up to 90% without changing a line of code, and @adriano_viana tested token-saving plugins until one cut 60% of the payload. @fluixoo's framing is the sharp one — people blame Claude Code when they hit limits, but the model isn't the only thing spending them. The adjacent ask is smarter defaults: @AlexFinn soured on Sol Ultra entirely because it turns short tasks into long ones, and medium is now the daily driver.
Memory still dies at the session boundary, so everyone rebuilds it by hand. The Obsidian-plus-CLAUDE.md second brain appeared over and over: @chesny and @MyWestLord both documented the Karpathy-method vault that turns a note pile into a living wiki, and @obsidianstudio9 caught a Berlin student doing 3,247 notes in one subway ride. Nobody wants to build this. They build it because, as @SucceededMind puts it, re-onboarding the AI to your own codebase every session is the exact pain every heavy user pays daily.
Once you run more than three agents you can't see anything. @GitHub_Daily's fix says it best: tabbing between background tasks in a terminal gives you no idea what they're doing, so they turned the agents into pixel workers living on the desktop wallpaper. People are building air traffic control because none ships.
Trust boundaries are getting drawn the hard way, by touching the stove. @bardia_heydari points out OpenClaw gets the bad reputation while you can just as easily hand Claude bypass permissions and a goal — and then describes how that feels. @cynical_zenitsu was scared badly enough to stop letting Claude Code query the production database, starting today. Mercari's answer was five layers of containment. The pattern is consistent: everyone finds the blast radius by getting close to it, then writes their own fence.
The CLI itself is now a ceiling on who gets in. @maxedapps says the TUI feels overloaded with features and configs — it works, but only a handful of things matter. @tsyn18 reports something sharper from the other side: students in their AI course are leaving Claude rapidly since ChatGPT Work launched, because for ordinary users Claude Code is simply too high a barrier. Meanwhile @0x_sakata switched off Claude after being on it since Opus 3.5, and @imdukepan now runs Codex, Fable and OpenClaw side by side off a shared memory file. Single-tool loyalty is over; the people who stay are the ones who can drive a terminal.
Usage limits are now the product experience. This is the loudest theme in the data and it's driving purchasing, not just grumbling. @cneuralnetwork left because limits ran out too quickly and rebuilt everything on Codex at work. @LyraInTheFlesh burned a week's quota in a day having GPT-5.6 orchestrate Opus 4.8 instances. @so_ainsight documents the workaround that's becoming standard: Fable 5 as orchestrator only, cheap subagents doing the actual work, purely to stay under the ceiling. When @ryu_grove_ai and @Krongggggg are both shipping menubar meters so you can watch the fuel gauge while you work, the constraint has become the interface.
Token burn on garbage is the fixable half of that. Users are paying frontier prices to have the model read terminal noise, and they've stopped waiting for a fix: @RetroChainer found a free Rust binary that cut token burn up to 90% without changing a line of code, and @adriano_viana tested token-saving plugins until one cut 60% of the payload. @fluixoo's framing is the sharp one — people blame Claude Code when they hit limits, but the model isn't the only thing spending them. The adjacent ask is smarter defaults: @AlexFinn soured on Sol Ultra entirely because it turns short tasks into long ones, and medium is now the daily driver.
Memory still dies at the session boundary, so everyone rebuilds it by hand. The Obsidian-plus-CLAUDE.md second brain appeared over and over: @chesny and @MyWestLord both documented the Karpathy-method vault that turns a note pile into a living wiki, and @obsidianstudio9 caught a Berlin student doing 3,247 notes in one subway ride. Nobody wants to build this. They build it because, as @SucceededMind puts it, re-onboarding the AI to your own codebase every session is the exact pain every heavy user pays daily.
Once you run more than three agents you can't see anything. @GitHub_Daily's fix says it best: tabbing between background tasks in a terminal gives you no idea what they're doing, so they turned the agents into pixel workers living on the desktop wallpaper. People are building air traffic control because none ships.
Trust boundaries are getting drawn the hard way, by touching the stove. @bardia_heydari points out OpenClaw gets the bad reputation while you can just as easily hand Claude bypass permissions and a goal — and then describes how that feels. @cynical_zenitsu was scared badly enough to stop letting Claude Code query the production database, starting today. Mercari's answer was five layers of containment. The pattern is consistent: everyone finds the blast radius by getting close to it, then writes their own fence.
The CLI itself is now a ceiling on who gets in. @maxedapps says the TUI feels overloaded with features and configs — it works, but only a handful of things matter. @tsyn18 reports something sharper from the other side: students in their AI course are leaving Claude rapidly since ChatGPT Work launched, because for ordinary users Claude Code is simply too high a barrier. Meanwhile @0x_sakata switched off Claude after being on it since Opus 3.5, and @imdukepan now runs Codex, Fable and OpenClaw side by side off a shared memory file. Single-tool loyalty is over; the people who stay are the ones who can drive a terminal.
📡 Eco Products Radar
Eco Products Radar
Claude Code — the substrate for nearly every case today, coding and not.
Codex / GPT-5.6 Sol — the relief valve when Claude limits hit, and increasingly the auditor in build-then-critique pairs.
OpenClaw — carrying the non-coding edge: robot control, multi-agent debates on a DGX Station, personal messaging, Japanese podcast generation.
Obsidian — the default vault for agent-readable memory, paired with CLAUDE.md in a dozen separate setups.
CLAUDE.md — treated as the actual engine of persistent memory rather than a config file.
Cursor — the tool people are migrating away from, usually toward Claude Code or Codex.
Fable 5 — the planning and orchestration layer, and the source of today's most-reported regression when a dropdown refactor tripped a safeguard.
MCP servers — the connective tissue: TradingView, Blender, Pika, freee, and a long tail of hand-rolled ones.
Hermes — the hotswap destination when people route off Opus, usually paired with GLM.
RTK — terminal output compression, the sharpest answer yet to token burn.
DeepSeek / Kimi / GLM — the cheap backends people point Claude Code at when the meter runs down.
Claude Code — the substrate for nearly every case today, coding and not.
Codex / GPT-5.6 Sol — the relief valve when Claude limits hit, and increasingly the auditor in build-then-critique pairs.
OpenClaw — carrying the non-coding edge: robot control, multi-agent debates on a DGX Station, personal messaging, Japanese podcast generation.
Obsidian — the default vault for agent-readable memory, paired with CLAUDE.md in a dozen separate setups.
CLAUDE.md — treated as the actual engine of persistent memory rather than a config file.
Cursor — the tool people are migrating away from, usually toward Claude Code or Codex.
Fable 5 — the planning and orchestration layer, and the source of today's most-reported regression when a dropdown refactor tripped a safeguard.
MCP servers — the connective tissue: TradingView, Blender, Pika, freee, and a long tail of hand-rolled ones.
Hermes — the hotswap destination when people route off Opus, usually paired with GLM.
RTK — terminal output compression, the sharpest answer yet to token burn.
DeepSeek / Kimi / GLM — the cheap backends people point Claude Code at when the meter runs down.
Comments