Super User Daily: 2026-10-27
Sunday was a passwords-and-proof day. Two of the most-shared builds were people pointing Claude Code at their own 1Password vaults: one rebuilt a 1,116-account review dashboard and pushed a security score from 30 to 73 in a day, the other is changing passwords and cancelling services across 516 accounts. levelsio spent a whole day typing "do a full security audit" into every server in the fleet and walled each box off with Tailscale policies the agent wrote. The measured posts were the best of the week: someone scanned 50,721 merged Claude Code pull requests and found Stripe added in 8 repos out of 29,522, a skills test showed Claude Code still opening the right skill 35 times out of 36 with 512 installed while Codex dropped to 17, and a Stanford paper clocked Claude Code's main agent waiting on subagents 73 percent of the time. On the creative side a parent and son built a Pokemon-meets-Jurassic-Park game together, a Civilization IV mod gave every AI civilization its own Claude Code session, and a hobbyist rendered an 858-frame product film on a 2018 laptop. The quieter story is skepticism maturing: several posts fact-checked the viral advisor-model trick and the scary deleted-photo stories rather than repeating them.
@gourav_io [Claude Code]
https://x.com/gourav_io/status/2114243446378828000
This developer scanned 50,721 pull requests that Claude Code got merged on public GitHub between October 17 and 23, slicing the week into 363 one-minute windows and counting each package once per repo. On new JavaScript projects the defaults are TypeScript, Vitest and React, and tests arrive on day one: Vitest showed up in 127 new projects. On existing projects the additions are again mostly test and type tooling, Vitest, Zod and Playwright. Payments almost never: Stripe was added in 8 repos out of 29,522. It is the rare post that measures what the agent actually chooses instead of what people say it chooses.
@cbarmorecpa [Claude Code]
https://x.com/cbarmorecpa/status/2114408274862669934
A parent spent Sunday building a game with their son in Claude Code, built entirely out of the boy's favorite shows. It started as a Pokemon Gold and Silver style game with Johto and Kanto, real-time battles, all 16 gym leaders and the legendary dogs, then grew a Dino Island where you outrun a T-Rex in a jeep, Sonic zones with racing loops, and a Hero City where you fight Ultron as Iron Man, Hulk, Spider-Man or Thor. The son comes up with every idea, prompts Claude to build it, and playtests everything himself, which the parent calls a very strict QA department. The game is just for the two of them, and that is the point: the cost of making a child's imagination playable has dropped to an afternoon.
@yibie [Claude Code]
https://x.com/yibie/status/2114239109665546464
An open-source mod for Civilization IV: Beyond the Sword replaces every computer opponent with an LLM, one Claude Code session per civilization, with Codex or other models also pluggable. The design keeps the agent's long-term state outside the model: each civ has a notes file with its persona, plans, dossiers on rivals and secrets; the mod and the model talk only through a folder of JSON requests; and one command, civ, covers every action a human player could take. The model is woken by events such as a new turn, a message or a trade offer, given only the playbook, its own notes and what changed, then it moves, updates its notes and goes back to sleep. Every message also carries a private note of what the model really meant, so a replay shows how far its diplomacy diverged from its intent.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2114476100294426897
Following the 1Password-plus-CLI trend, this user moved all 1,116 saved passwords into 1Password and had Claude Code build a dashboard that reviews them one by one. Each account is scored on whether it has leaked, is reused, is weak, lacks two-factor authentication, or belongs to an AI service worth extra caution, and a batch button opens the change-password pages in order of danger. The starting score was 30 and one focused day got it to 73. The user prefers the homemade version to 1Password's own dashboard because it goes straight from finding to fixing and can be customized.
@GOROman [Claude Code]
https://x.com/GOROman/status/2114235972946944020
This developer is using Claude Code to build a tool that changes the password on every one of the 516 accounts stored in 1Password and automatically cancels the services that are no longer needed. A follow-up post shows the system working as an automatic password rotation. The interesting part is the second job: cancellation flows are deliberately unfriendly, so an agent that can walk through them account by account turns a chore people avoid for years into a batch job. It pairs naturally with the 1,116-account audit dashboard another user built the same day.
@levelsio [Claude Code]
https://x.com/levelsio/status/2114317753985188201
levelsio spent an entire day typing one sentence into Claude Code on every VPS and site in the portfolio: do a full security audit of this website and server. Unattended security upgrades with automatic reboots were already in place, so the new work was isolating each VPS on Tailscale with policies that Claude Code generated, so a compromised box cannot reach any other server, the MacBook or the iPhone. Web services were also pushed down to the www-data user to keep them away from root. The conclusion is a monthly security week next to the marketing week, on the bet that attack and defense capabilities of the models are rising together, so regular audits are what keep a solo operator even.
@froessell [Claude Code]
https://x.com/froessell/status/2114310224391655453
In two hours this developer moved the racing game Drift Lane from Three.js to Unity with a whole new UI, so it can ship on iOS and Android. Asked whether to use TopDown Engine or HighRoad Engine from the Asset Store, Claude Code read the code and recommended neither: the drift handling is about 170 lines of hand-tuned physics that neither engine would preserve, so it ported those lines one by one. For the interface, ChatGPT produced a concept image, Claude Design turned it into a first draft over several rounds, and Claude Code rebuilt the screens in Unity. The advice to skip the off-the-shelf engine is the kind of judgment people usually assume an agent will get wrong.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2114347925182816488
This test installed 12 real skills into Codex and Claude Code and then piled 500 decoys next to them to see whether the right skill still gets opened. With 512 installed, Claude Code opened the correct one 35 times out of 36, while Codex managed only 17, and Codex's own error explains why: it exceeded its skills context budget, removed all skill descriptions and left 134 skills out entirely. With descriptions gone it had to guess from names, four skills were never opened again, and every request for two different slide-deck skills went to one of them. Up to 112 installed skills both tools stayed at 35 or 36 out of 36, and every request and result is published.
@jinmingyang666 [Claude Code]
https://x.com/jinmingyang666/status/2114396449769410717
This user runs a homemade bot whose engine can be Claude Code, Codex or another harness, whichever is strongest that month. The bot itself is a Git repository: skills, a system prompt file in AGENTS.md, channel connectors, scheduled tasks, notifications and memory. It is deployed across three machines at home joined by Tailscale, each able to SSH into the others and do computer use across machines, and it keeps the local browser's logged-in sessions. It is wired into WeChat, Telegram, iMessage and the Paseo app, and the owner says they have barely opened a computer for work in a month outside of meetings. The stated reason for building it instead of using a cloud bot: no vendor lock-in, so when one provider degrades they switch the next day.
@Xudong07452910 [Claude Code]
https://x.com/Xudong07452910/status/2114221415012442598
A new Anthropic paper put Claude Code, Codex and Gemini into 20 real machine-learning-for-science benchmarks and let them read the paper, edit the code, run experiments and try to beat the published best. Every frontier agent found improvements, with Claude Opus averaging a 10.4 percent relative gain, but the trajectories showed most of it came from familiar engineering tricks such as hyperparameter tuning, optimizer swaps and ensembles, while the agents overrated their own novelty and left GPUs idle. Adding just three light skills, for research ideation, experiment management and scheduled self-checks, raised Opus's average gain to 20.1 percent and Codex's from 4.4 to 8.1. The summary's verdict: research agents are bottlenecked on judgment, not execution, and research habits are cheaper to give them than more compute.
@snwiki238337 [Claude Code]
https://x.com/snwiki238337/status/2114253870583189846
A Stanford paper from Azalia Mirhoseini's team measured what centralized subagents actually cost. On long Terminal-Bench 4.0 tasks, Claude Code's main agent spent 73 percent of its time waiting for subtasks to return, and 44 percent of all agent compute went to waiting bubbles; on DeepSWE, turning on native subagents dropped accuracy from 71.67 to 50 percent while raising time by 15.9 percent and cost by 65.1 percent. The proposed DeLM framework removes the boss agent and has peers share an append-only state log and a first-come task queue, publishing intermediate findings and dead ends as soon as they have them. With two Opus 5.5 agents it reached 93.33 percent on Terminal-Bench, four agents ran 2.49 times faster and cut the waiting share from 44 to 2 percent.
@ai_database [Claude Code]
https://x.com/ai_database/status/2114184481951330365
Researchers from Georgia Tech and elsewhere built 40 projects where CLAUDE.md or AGENTS.md deliberately contradicted the code, then asked Claude Code and Codex to add features. Claude Code noticed the conflict in writing 79 times out of 80 and followed the stale instructions anyway, while Codex rewrote working code to match the outdated docs in 20 of 80 runs, breaking tests in 17. The explanation offered is that instruction files arrive as instructions and code arrives as observations, and the models are trained to obey instructions. The recommended fix is one sentence in the instruction file: when documentation and existing code disagree, follow the code and list every inconsistency in the final reply, which a separate summary says lifted code-following from 4 to 91 percent.
@theo [Claude Code]
https://x.com/theo/status/2114166439813017873
Theo spent a day chasing a usage mystery and confirmed it is a real bug on Anthropic's side: any usage through the Agents SDK drains the five-hour limit faster than using Claude Code directly. Weekly limits are not affected, only the five-hour window. The bug has been reported to Anthropic, and the post spread quickly through English and Chinese timelines because it explained why some people hit the cap early this week. It is a useful reminder that the same subscription can meter differently depending on which door you use.
@RealYDT [Claude Code]
https://x.com/RealYDT/status/2114382577721319622
A viral post claimed Anthropic had revealed the best Claude Code trick: run claude --model sonnet --advisor opus, let Haiku swarm the repo, and have Opus step in only at three fixed moments. This user checked it line by line against the official docs. The advisor flag is real but undocumented in claude --help, and Sonnet as main model with Opus as advisor is an officially recommended pairing; but when the advisor steps in is model-driven, not rule-based, Haiku exploration has nothing to do with the flag unless you define your own Explore subagent with model: haiku, and the routing layer at the end of the prompt simply does not exist in Claude Code. The missing cost note is the useful part: the advisor rereads the whole conversation uncached every time it is called, and subagents inherit the advisor setting.
@Layton_Gott [Claude Code]
https://x.com/Layton_Gott/status/2114361430036090941
After more than 100 hours of running Opus 5.5 in Claude Code, including overnight loops, this user wrote down what actually held up. Keep each task under about 100,000 tokens, clear and restart instead of compacting, and never let a big session sit idle, because the cache expires after five minutes and the next message rebills everything: one session at 98 percent cache still cost about 200 dollars API-equivalent in half an hour. Add one line to CLAUDE.md every time Claude makes a mistake, make it write a failing test before any fix, give it a way to check its own work, and review diffs in a fresh session. For overnight runs, set permissions before leaving and give it more work than it can finish, because an agent with nothing left to do simply stops.
@kirisando [Claude Code]
https://x.com/kirisando/status/2114342837382721850
Fiverr lists 3D product motion videos for online shops from about 100 euros, so this user, with no motion design or 3D skills, tried making one with Claude instead. In about an hour and four rounds of notes, Claude Code rebuilt a game controller as a 3D model in code from a single photo, built the water, studio lighting and one long camera flight in code, took the controller apart to show the battery, board and motors inside, and wrote the sound in code too. No video model generated a single frame, and all 858 frames rendered on a 2018 laptop on the cheapest paid plan at 18 euros a month. The honest caveat: Opus 5.5 still needs fixes along the way.
@afscott [Claude Code]
https://x.com/afscott/status/2114371621893070998
This creator used to pay 800 dollars to get one video edited and two weeks to get it back. The new pipeline costs about 4 dollars of AI per video and two hours from recording to finished edit: Claude Code does the editing, Gemini drafts, and DaVinci Resolve finishes. The full setup is shared start to finish. It is one of several posts this weekend where video editing, not coding, was the clearest money saved.
@EcomKostnchko_ [Claude Code]
https://x.com/EcomKostnchko_/status/2114355190719807822
This e-commerce operator combined Claude Code, Omnisend, Nano Banana and FFmpeg to build email flows, and produced 32 emails for an abandoned-cart sequence in under 10 minutes from an existing workflow. Done the old way, with designers and someone slicing images, it would have taken months. The welcome flow and checkout-abandon flow are done, with cart, site and browse abandon flows next. The trick is feeding detailed reference guides on how each flow should be built and having Claude make artifacts to check against before shipping.
@jacky_gng [Claude Code]
https://x.com/jacky_gng/status/2114383539961495896
This founder is building 500 real e-commerce shops in public to sell links through a link marketplace, with 15 done one day and roughly 130 expected by that evening. The work started about 20 days ago with ten themes, from home and travel to health and parenting, each stocked with 3,000 dropshipping products shipped from Europe. Claude built a CMS that can generate up to 100 shops a day and edit products in bulk, and every shop is static HTML regenerated only where something changes, so pages load fast and thousands of sites can live on one server. The founder estimates it would have taken six months before Claude Code and will take under one.
@BlakeTOliver [Claude Code]
https://x.com/BlakeTOliver/status/2114402482176626918
A roundup of four small accounting firms using AI on one annoying task each included a clean Claude Code case. High Rock Accounting built its own client satisfaction tracker in five hours with Claude Code and dropped the paid software it replaced. The other firms in the same roundup used Lovable and Claude for intake forms that lifted conversions 25 percent, and ChatGPT for multi-state M&A tax memos in hours instead of days. The common pattern is a small professional firm picking one repetitive job and owning the tool instead of renting it.
@leoai_build [Claude Code]
https://x.com/leoai_build/status/2114201149460840624
This user already runs agents across several devices with Yobiru and read through the source of the new Yobiru Friends release, which lets different people's Grok Bot, Muse and Claude Code befriend each other and chat, share files and book time, end-to-end encrypted. The code has real guardrails: messages from non-mutual contacts sit in a request pool and the bot cannot even read their content, calendar and FAQ answers run in deterministic workers that strip titles, locations and attendees, and bot-to-bot chats cut off at 20 turns per session, 100 per day and 4 relay hops. The one real hole is the mutual-friend path, where a message wakes your bot within a second and its reply goes out without your review, while all bots under one account share files, browser logins and memory. The advice: be stingy with follow-backs and keep the owner token on your own Mac, giving the cloud bot only a reduced token.
@koder_dev [Claude Code]
https://x.com/koder_dev/status/2114296158059405823
A Japanese post went around about a developer whose Mac started taking 20 minutes to boot after six weeks of vibe coding. The cause was about 990,000 Claude Code working files piled up in /tmp. The poster's point is the general one: machinery built to make you faster can quietly become load somewhere else, and they planned to check their own Mac. It is the kind of failure nobody benchmarks, and a cheap thing to look at today.
@bigbropatchwork [Claude Code]
https://x.com/bigbropatchwork/status/2114371892664479751
Solitaire: Kingdom in Chains took this solo developer three months, built with Photoshop, Cubase, Filmora, Claude Code and Codex, with design, editing, testing and rebuilding done by hand until it felt ready. The developer estimates the split at 30 percent AI and 70 percent human. Once Steam attached its AI disclosure label, some players dismissed it without looking at the gameplay. The post is not asking for sympathy sales, just for the game to be judged instead of the toolbox, which is becoming a real distribution problem for small studios that use AI honestly and disclose it.
@DannyLimanseta [Claude Code]
https://x.com/DannyLimanseta/status/2114405388171034809
Danny Limanseta spent a weekend playing with Unreal Engine and found it far more vibe-code friendly than its AAA reputation suggests once the setup was done. The prototype was built entirely from Claude Code through the command line, without opening the Unreal editor once. Assets came from the Fab store and turned out better than expected. Driving a big-studio engine from a terminal agent lowers a barrier that has kept solo builders on lighter engines.
@KanaWorks_AI [Claude Code]
https://x.com/KanaWorks_AI/status/2114190567345381742
A black liquid tentacle creature that crawls across floors, walls and ceilings with zero hand-keyed animation and no machine learning, only code and real-time physics in Godot 4. The post doubles as a spec an agent like Claude Code can rebuild from: grip points scattered over every surface with Poisson-disk sampling, a per-frame score that picks reachable points in the direction of travel, release rules when a tentacle stretches past 92 percent of its length, and a body treated as one mass whose gravity scales with how many tentacles are holding. The look comes from ray-marching a signed distance field that melts body, tentacles and droplets into one fluid with smooth-min. It is a good example of a creator publishing the design document instead of the code, on the assumption that the agent will write the code.
@sai0xgen [Claude Code]
https://x.com/sai0xgen/status/2114394053932458297
A developer is building a 4 by 5 kilometer island in Unity using Claude Code, based on real elevation data from Lofoten in Norway, with a 1,000 meter snowy peak. Beyond the terrain almost everything is procedural: rocks, moss, plants, forests and even where birds gather, all placed by the island's height and water rather than by hand. The developer directs while Claude Code writes the code, shaders and tools, then tests and benchmarks its own work. The next goal is generating whole landscapes from scratch, with erosion, glacial valleys and coastlines, using the real island as the quality benchmark.
@HashItOut_ [Claude Code]
https://x.com/HashItOut_/status/2114492475931078760
gokayfem gave Claude Opus 5.5 a Gaussian splat, a photoreal 3D capture of a garden, and had it rebuild the scene in Blender as editable procedural geometry. The agent recreated objects with Geometry Nodes, wrote materials with Shader Nodes and mapped captured textures back onto selected surfaces, so the result can be moved, edited and relit. The author is candid in the replies that it is not very accurate and matches by rough similarity rather than exact position. It is open-sourced under MIT as a Claude Code skill and works in Codex too, and the sensible read is a blockout pass an agent drafts while an artist does the detail work.
@AIandDesign [Claude Code]
https://x.com/AIandDesign/status/2114421153909596522
This user finished a Claude plugin that autonomously makes music videos with beat-accurate cuts and custom effects, using archive footage and, optionally, AI renders through any video MCP such as Hedra, Magnific, Scenario or Runway. The more detail you give it, the closer it lands, and after a render you can edit it with notes like change the section between 0:43 and 0:56. The showcased video used only archive footage, with a second example weaving in AI renders. It installs as a Claude Code plugin with a /music-video command and has been submitted to the official plugin repository.
@ericosiu [Claude Code]
https://x.com/ericosiu/status/2114376940513616134
This team open-sourced the rubric their AI uses to package YouTube videos: it scores every title and thumbnail out of 10 before a human sees them. Dropped into Claude Code, it writes three title-and-thumbnail packages from the transcript, kills anything under 8.0 overall or under 7 on any part, and explains the winner. It throws out any thumbnail you cannot read in one second at phone size and any image two viewers would describe differently. Long-form click-through went from around 4 percent to 6 to 8 percent once it started packaging the videos.
@nutlope [Claude Code]
https://x.com/nutlope/status/2114445139800461618
This builder laid out the current app-building loop. Record a one to five minute voice prompt with Wispr Flow, add visual inspiration through screenshots or an Inspo MCP, then let Matt Pocock's grill-me skill make the agent ask clarifying questions before it builds. Feedback goes back as another voice recording, and every new feature must come with a quick screen recording proving it works before the PR merges. The harness is a mix of Claude Code on Opus 5.5 at high effort and OpenCode with DeepSeek V4.1 Flash or GLM 5.3, with Vercel and GitHub CLIs so agents can deploy on their own.
@kaif9998 [Claude Code]
https://x.com/kaif9998/status/2114386504059601389
This studio ships AI MVPs for founders in three to four weeks with Claude Code and described the cadence. Week one has no code at all: the idea becomes a one-page spec, user flows and a CLAUDE.md with the stack, rules and a never-do-this list. Week two is only the core flow, auth, database and the one feature the product exists for, with every Claude PR reviewed like a junior's; week three adds payments, emails, onboarding and admin in parallel sessions on separate branches; week four is tests, edge cases and real users clicking through. The biggest single time saver: Claude never starts coding without a written plan it has to show first, because one rejected plan saves a day of rewrites.
@by__huy [Claude Code]
https://x.com/by__huy/status/2114478011814924746
This studio moved its site, MONOLOG, off Webflow to Astro with Keystatic CMS, building alongside Claude Code, and says the speed of iteration and the reduced bloat are both real. Before Webflow the founder was a custom-code purist who found code too slow to iterate and too hard for clients to manage, which is exactly what Webflow solved for years. Now the open problem is the client editing experience, because custom code moves fast but the CMS still has to feel good to non-developers. The bigger shift is in deliverables: the studio wants to start handing clients AI frameworks, prompts and systems they can use to extend their own brand.
@ShikshanNivesh [Claude Code]
https://x.com/ShikshanNivesh/status/2114323511858851874
Most of the companies in this investor's portfolio report by early November, and most people read the results as a headline the next morning. So the investor built a program with Claude Code that reads the filings and results for every holding twice a weekday, at zero running cost. The detail worth copying is that no AI runs it day to day: Claude Code wrote the program once, and the program does the reading. It is the same move several builders made this weekend, using the agent to build a deterministic tool rather than paying the agent every time.
@liuqindong888 [Claude Code]
https://x.com/liuqindong888/status/2114159223240306889
This creator shared a working method for running a Xiaohongshu account with AI aimed at selling digital resources. Any working agent on the computer will do, WorkBuddy, Codex or Claude Code, and even a DeepSeek model is good enough to earn some side income. The key step is telling the agent you want account positioning, a material library, note style templates, drafting, review, and a loop where review data flows back into the library and templates, then asking it to work through every stage with you Socratically until a closed workflow exists. After that the daily work is feeding it material and competitor hits, publishing its drafts, and having it run the review.
@ding_liam [OpenClaw]
https://x.com/ding_liam/status/2114260170171658707
This user deleted the Hermes Agent that ran on Railway, spent an afternoon moving every workflow and project to Grok Bot, and reorganized the bots' roles and reporting lines, noting that more bots is not better and group chats should be avoided unless necessary to save tokens. The same post ranks nine personal agents from best to worst, with Grok Bot on top for its 1,000 free daily X searches, Muse and Claude Duet next, and OpenClaw at the bottom because the author left it long ago. The leftover Railway plan still has 8 GB of memory, 8 vCPUs and 100 GB of disk to find a use for. It is a clean snapshot of the migration path many early OpenClaw users took this autumn, OpenClaw to Hermes to a hosted bot.
@AudioMetaX [OpenClaw]
https://x.com/AudioMetaX/status/2114411447903993947
On day two of an experiment, this producer had an OpenClaw agent make music inside REAPER. The agent programmed the drums, 808s and melodies from scratch, arranged them, shaped the mix with FORM plugins, and screen-recorded the whole session with the finished master as its soundtrack. The result is a dark trap instrumental with sliding 808s, cold bells and rolling hats, sparse verses and a stripped-back breakdown. A digital audio workstation is about as far from a code editor as creative software gets, which is what makes the attempt interesting.
@DumbTien [OpenClaw]
https://x.com/DumbTien/status/2114307892501459187
This contributor got a batch of OpenClaw Android fixes merged into main: session groups in the sidebar to match the web app, Effort and Fast mode no longer limited to admins, photo and voice note sent as one message, and voice transcripts kept in history. The maintainer replied that the release should go out within a week or so. It is a small example of the project still being shaped by its users even as attention moves to hosted agents. The contributor's only complaint is waiting for the stable release.
@RealYDT [OpenClaw]
https://x.com/RealYDT/status/2114504860167200828
A practical answer to the timeline argument over whether a personal agent needs a Mac mini or a small Linux box. Using OpenClaw as the example: if the agent talks to cloud models like Claude or GPT, the always-on machine mainly receives messages, runs the gateway and calls tools, and the official minimum for the gateway on a VPS is one core and 1 GB of memory, more only if you add browsers and channels. A Mac earns its place when the agent must touch iMessage, macOS apps, local files or a visible browser, and OpenClaw can keep the gateway on a server and attach the Mac as a node only when needed. Running models offline is a separate GPU and memory question, so list the agent's actual jobs before buying hardware.
@smhanov [OpenClaw]
https://x.com/smhanov/status/2114392468313563567
This developer has been writing many agent harnesses and abstracted the lessons from studying OpenClaw, Hermes, Crush, piclaw and the DeepSeek harness into a single Go library. It continues a long-running series on running several 10,000-dollar-MRR companies on a 20-dollar-a-month tech stack, and is aimed at people who share that philosophy. The value is less any single feature than the comparison: one person reading five harnesses and keeping what they agree on. It is part of the weekend's broader trend of builders owning the loop rather than renting it.
@fkadev [OpenClaw]
https://x.com/fkadev/status/2114420928985858495
This site now has community-written bot recipes for Grok Bot, OpenClaw and Hermes, 331 so far. The twist is in the footnote: the feature itself was developed by a group of Grok bots. Replies asked whether the 331 recipes are spread across all three agents or mostly Grok Bot, and which recipe gets copied the most. A shared recipe library across three competing personal agents is a sign that the setups people build are becoming portable even when the agents are not.
@A7_data [Claude Code]
https://x.com/A7_data/status/2114282454345527628
This user is building a personal Jarvis with Claude Code underneath and says the rough version is already more useful than Dots or Muse. It writes code, lives inside LINE, knows the full history of past conversations with Claude Code, works from the phone and by voice, and costs nothing extra. The moment that sold it: while deep in work, the assistant spoke up with the next appointment, said leaving in 10 minutes would make it, and that the train time had already been added to Google Calendar. That proactive nudge is exactly what the hosted agents advertise, built here on top of a coding agent.
@sisithe88 [Claude Code]
https://x.com/sisithe88/status/2114360287050744182
This user tried the new Claude Code projects feature by creating a commander agent and simply saying what needed doing; it spun up its own subordinate agents and pushed the work forward. The test case was a low-priority YouTube project that had been sitting idle, and it reached the point where the video is ready to publish. The user admits the setup has quirks, but expects work to go ten times faster once used to it. It is a small but telling signal that the orchestration layer is now something non-developers try on side projects.
@PawelJLisowski [Claude Code]
https://x.com/PawelJLisowski/status/2114466003300012193
This developer is building a browser game demo with Opus 5.5 in Claude Code and just updated the class builder: pick three of eleven skill trees and get one of 165 named classes, the author's being a gunpowder spellstalker. Opus wrote the game systems, Tripo made the characters, Mixamo and Meshy did the animation, and Three.js runs it in the browser. The honest bottleneck the post ends on: deciding what is actually fun still takes forever. Generating 165 class combinations is now cheap; designing which ones deserve to exist is not.
@jiangjiangling [Claude Code]
https://x.com/jiangjiangling/status/2114295810087387200
Cruising Bones, a downhill skateboarding game built with Claude Code, is now free to play in the browser. By the developer's account work started on October 14 and the third update shipped about ten days later. The toolchain is spelled out: Claude Code and Claude for development, Magnific and Tripo for art, Babylon.js as the engine, with multiplayer marked as coming soon. One commenter said they would pay for it once the controls and shading are polished, which is the useful kind of feedback: the prototype attracts people, and retention is the next test.
@Its_lakshya_ai [Claude Code]
https://x.com/Its_lakshya_ai/status/2114318178042167595
afterthewinddev, a developer with no coding background, spent six weeks with AI and Godot and went from rough sketches to a playable RPG inspired by Ragnarok Online and Monster Hunter. Opus 5.5 in Claude Code handles most of the development, Codex and Astra create visual references, Tripo turns them into 3D assets, and Cascadeur handles animation. What stands out is the process work: custom workflows to route coding tasks, continue sessions, verify research and keep the art style consistent. Even successful prompts are saved and reused so the visuals do not slowly drift.
@ezsmith397 [Claude Code]
https://x.com/ezsmith397/status/2114418133670748421
A parent shared their daughter's game in progress from a first-ever Claude Code session, already playable online. The daughter is prompting and revising along the way and having a blast. Next to the parent-and-son Pokemon project the same day, it shows a pattern: kids are not learning to code first and build later, they are building first and learning by revising. The barrier between a child's idea and something a friend can play is now one session long.
@przemekwilde [Claude Code]
https://x.com/przemekwilde/status/2114407003942523330
A Polish parent suggested a different way to show a child technology: make a simple game together with AI and let the child draw the characters, the levels and the rules. Their child's game is called Donkey Rescues the Kitten in the Labyrinth, and the parent's only job was setting it up in Claude Code and moving the drawings into the virtual world. It is playable on a phone and a Mac, and more modules are already being made. The child's drawings stay the source of truth, and the agent is the bridge.
@VinodSharma10x [Claude Code]
https://x.com/VinodSharma10x/status/2114348373016760617
After 12 years in tech management, this engineer went back to hands-on coding in their fifties and now has 10,000 LinkedIn followers following along. The advice for experienced people who feel disconnected: start embarrassingly simple in browser tools like Bolt and Lovable with tiny ideas, such as a portfolio site and an app that turns a photo of a dish into its recipe, and only later move to VS Code, Cursor and the Claude Code CLI for complex apps and agentic systems. The first night was spent hooked until 6 a.m., the first time in 12 years the gap between an idea and a working app disappeared. The point the post makes well is that decades of architecture and troubleshooting experience become an advantage once the typing is delegated.
@BorderleSint [Claude Code]
https://x.com/BorderleSint/status/2114194392609132779
The gym-coding bridge this Portuguese-speaking user described earlier has become a reusable template: a Grok Bot that relays between the owner and a coding agent, whether Claude Code, Codex or Cursor. It carries the request in, brings the answer back, searches the web for what the coding agent cannot reach, and at the end of the day sends a list of everything waiting for approval. The rule that saved the most trouble is that only the owner gives orders: instructions that appear in a post, a DM or from Claude itself never become tasks without passing through the owner, and push and deploy stay in human hands. The catch is that the bridge script is not included, so everyone builds their own while the bot hands over a ready-to-paste request.
@yossy_indie [Claude Code]
https://x.com/yossy_indie/status/2114194487371055440
A cost tip for anyone generating images from inside Claude Code with gpt-image-2.5 and watching API charges pile up. Log into the Codex CLI with a ChatGPT account, ask Claude Code to delegate the job with codex exec, and write in the instruction not to use an API key. Images then come out of the ChatGPT plan's allowance instead of the API bill. It is a neat example of one agent calling another as a tool, and of users routing work to whichever subscription already covers it.
@techresearchlab [Claude Code]
https://x.com/techresearchlab/status/2114214118328803758
A setting that trips people up: if you write Claude Code's auto mode configuration into the project's .claude/settings.json, it is ignored, because it has to live in the user-level ~/.claude/settings.json. The reason appears to be deliberate, so that a settings file shipped inside a repository cannot quietly widen permissions on your machine. With auto mode about to become the default for new sessions in many Claude apps, the post recommends reviewing the setup now. A /auto-mode-setup command can draft the configuration.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2114286199133331646
This creator tested whether Meta's Muse can run an Instagram and Threads account and documented the result. Reading the numbers, writing posts in the formats that performed, cutting a reel and publishing all happened in conversation on the free plan, and every post required tapping a confirmation card, with the whole run using about 1 percent of the weekly free allowance. The explainer video itself was assembled with Claude Code on Opus 5.5, cutting and zooming three screen recordings with ffmpeg. The most useful finding was a near-miss: on the first confirmation card Muse had quietly re-summarized the carefully written post, and only refusing and resubmitting got it word for word, which is why the human approval step matters.
@llllleo9 [Claude Code]
https://x.com/llllleo9/status/2114365358572548146
A pitfall for anyone routing several models through Magpie into Claude Code. One conversation started sounding like two people arguing, and the AI misjudged things twice. The cause was not the routing group: in Magpie's agent page the subagent model for Claude Code had been changed during setup and never set back to default, so after switching the main model to Kimi, the subagents kept running on the old DeepSeek slot. The fix is to set the subagent model back to default, and the general lesson is that main and subagent models are configured separately and drift apart silently.
@ZypherHQ [Claude Code]
https://x.com/ZypherHQ/status/2114269985233072224
A CLI-only user is testing T3 Code as the control base for all of their harnesses, since it works more as an orchestrator across them, and plans to keep testing through November. Onboarding was rough because of VM issues, but the small details fit how developers actually work, and opening branches and PRs from the interface is unusually simple. The one missing piece is a view of the context window for each subagent and the main session, which matters because this user decides when to switch agents based on how full the context is. It is a concrete request any orchestrator could ship next.
🗣 User Voice
User Voice
Metering still decides the mood, and it now depends on which door you use. @ValencianaAbel repeats the Agents SDK finding that only the five-hour window drains faster; @CampbellKaleb23 has two accounts at 100 percent and is waiting for a fix; @Brooooook_lyn says the five-hour cap got harder to hit while the weekly one runs out faster; @StatsWire points at Claude Code's average context jumping 2.6 times since Opus 5.5. The counter-voice is @Michaelzsguo, on the 20-dollar plan, who has never used up a five-hour window.
The desktop app is quietly winning arguments. @leoleotoobad asks what the CLI does that the app cannot, @MrTacticalX forgot the CLI existed after days in the app, and @Nilzkool sees non-technical people choosing the CLI anyway. The complaints are concrete: @_Phil_Wilson_ says the console on Windows 10 is still broken, @chipit24 wants finished threads out of a 20-item project list, and @seatedro cannot paste a local image into Claude Code running on a remote server.
Handoffs between agents are where work gets lost. @daniel_du asks what to do when Codex runs out of quota halfway and Claude Code starts with no context; @phuctm97 pays for Codex only because Claude cannot generate images or plug cleanly into Hermes or OpenClaw; @emrahdma reads the private MEMORY.md files in ~/.claude and ~/.codex as lock-in; @sanketnadhani wants a plain task list that agents pull from one after another.
Users want observability that tells the truth. @TheLucasToday notes that MCP result bodies are missing from both OpenTelemetry signals in Claude Code 2.1.259 and only reach the PostToolUse hook; @robshocks no longer launches agents from the Downloads folder after the AGENTS.md attack report. And the community is policing its own feed: @NDLBylg, @caloow and @mark_0_poulos each took apart a viral Claude Code found my deleted photos story.
OpenClaw sentiment has split into leavers and maintainers. @0xVeryBigOrange says it was replaced within a year, @nicekate8888 has shut down the Mac that ran it, @russellclarke9 left because it broke with every update, and @p4panic calls local setup on a Surface not worth the time; but @das_rdsm notes most complaints come from people who stopped updating, and @CLOUT_COCKS says an OpenClaw bot on Codex still beats Grok Bot for trading.
Metering still decides the mood, and it now depends on which door you use. @ValencianaAbel repeats the Agents SDK finding that only the five-hour window drains faster; @CampbellKaleb23 has two accounts at 100 percent and is waiting for a fix; @Brooooook_lyn says the five-hour cap got harder to hit while the weekly one runs out faster; @StatsWire points at Claude Code's average context jumping 2.6 times since Opus 5.5. The counter-voice is @Michaelzsguo, on the 20-dollar plan, who has never used up a five-hour window.
The desktop app is quietly winning arguments. @leoleotoobad asks what the CLI does that the app cannot, @MrTacticalX forgot the CLI existed after days in the app, and @Nilzkool sees non-technical people choosing the CLI anyway. The complaints are concrete: @_Phil_Wilson_ says the console on Windows 10 is still broken, @chipit24 wants finished threads out of a 20-item project list, and @seatedro cannot paste a local image into Claude Code running on a remote server.
Handoffs between agents are where work gets lost. @daniel_du asks what to do when Codex runs out of quota halfway and Claude Code starts with no context; @phuctm97 pays for Codex only because Claude cannot generate images or plug cleanly into Hermes or OpenClaw; @emrahdma reads the private MEMORY.md files in ~/.claude and ~/.codex as lock-in; @sanketnadhani wants a plain task list that agents pull from one after another.
Users want observability that tells the truth. @TheLucasToday notes that MCP result bodies are missing from both OpenTelemetry signals in Claude Code 2.1.259 and only reach the PostToolUse hook; @robshocks no longer launches agents from the Downloads folder after the AGENTS.md attack report. And the community is policing its own feed: @NDLBylg, @caloow and @mark_0_poulos each took apart a viral Claude Code found my deleted photos story.
OpenClaw sentiment has split into leavers and maintainers. @0xVeryBigOrange says it was replaced within a year, @nicekate8888 has shut down the Mac that ran it, @russellclarke9 left because it broke with every update, and @p4panic calls local setup on a Surface not worth the time; but @das_rdsm notes most complaints come from people who stopped updating, and @CLOUT_COCKS says an OpenClaw bot on Codex still beats Grok Bot for trading.
📡 Eco Products Radar
Eco Products Radar
Codex: the second harness everywhere, from the skill-routing test and the DeLM paper to image generation through codex exec and the agent people fall back to when quota runs out.
Grok Bot: the hosted personal agent people migrated to from OpenClaw and Hermes, now also used as a relay to coding agents.
Hermes: the self-hosted agent that sits between OpenClaw and hosted bots in most migration stories, with a Claude Code importer.
Muse, Dots and Claude Duet: the hosted personal agents ranked, tested on Instagram and compared against homemade Claude Code assistants.
1Password CLI: the center of the weekend's most-shared builds, password review and bulk rotation.
Tailscale: the network under self-built bots and server isolation, and the subject of a vulnerability warning for anyone using Tailscale SSH.
Obsidian: the second-brain front end people keep debating against a plain folder plus CLAUDE.md.
Blender, Unity, Unreal and Godot: the creative engines driven from Claude Code, from Gaussian-splat rebuilds to a CLI-only Unreal prototype.
Remotion and FFmpeg: the video stack behind the editing pipelines that replaced 800-dollar edits.
Kimi K3.1 and DeepSeek V4.1 Flash: the cheaper models routed under Claude Code and compared against Opus 5.5.
T3 Code: the orchestrator CLI users are testing as a control base for multiple harnesses.
Agent Skills and design skill packs: the format behind skill-routing tests, landing-page design packs and video skills.
Codex: the second harness everywhere, from the skill-routing test and the DeLM paper to image generation through codex exec and the agent people fall back to when quota runs out.
Grok Bot: the hosted personal agent people migrated to from OpenClaw and Hermes, now also used as a relay to coding agents.
Hermes: the self-hosted agent that sits between OpenClaw and hosted bots in most migration stories, with a Claude Code importer.
Muse, Dots and Claude Duet: the hosted personal agents ranked, tested on Instagram and compared against homemade Claude Code assistants.
1Password CLI: the center of the weekend's most-shared builds, password review and bulk rotation.
Tailscale: the network under self-built bots and server isolation, and the subject of a vulnerability warning for anyone using Tailscale SSH.
Obsidian: the second-brain front end people keep debating against a plain folder plus CLAUDE.md.
Blender, Unity, Unreal and Godot: the creative engines driven from Claude Code, from Gaussian-splat rebuilds to a CLI-only Unreal prototype.
Remotion and FFmpeg: the video stack behind the editing pipelines that replaced 800-dollar edits.
Kimi K3.1 and DeepSeek V4.1 Flash: the cheaper models routed under Claude Code and compared against Opus 5.5.
T3 Code: the orchestrator CLI users are testing as a control base for multiple harnesses.
Agent Skills and design skill packs: the format behind skill-routing tests, landing-page design packs and video skills.
Comments