Super User Daily: 2026-11-16
Friday, November 14, was the day Claude Code desktop got custom themes through mods, and the best builders skipped the wallpaper and turned the mod system into a dashboard for the agent itself: a usage bar with live cost and limits, a radar that lights up on rm -rf or a read of .env and can block it. The other strong thread was hard numbers on who does the work. One solo builder logged 56 agent hours for every hour of their own across 1,977 runs, Garry Tan measured Fable 5.2 at 2.8 times the cost per solved task of Opus 5.5 on the same coding set, and a Japanese creator says Claude Code edits now save close to a million yen a month in outsourced video work. Non-coding use kept growing up: a marketer running a 1,500-person event list from Claude Code started grading software vendors on whether they work headless, a four-and-a-half-hour Blender film cost about $20 in rented GPUs, and a person whose glasses were stolen at an airport shipped a finder app. The lesson of the day: the people getting the most out of Claude Code are measuring it, not decorating it.
@Dawid_AI_ [Claude Code]
https://x.com/Dawid_AI_/status/2121469503154589871
A solo builder on day 25 of a 30-day sprint posted the honest before and after: the planned product, a desktop agent harness for small business owners, still has no finished plugin, because a quick task manager for agents ate the month and became the project, now called SwarmUp. Claude Code, Codex and other CLI agents pull from one queue of Markdown task files in a git repo, the orchestrator caps every run by time, tokens and plan allowance, reviews go to a different provider, and nothing merges until the builder's tests pass. The log from October 12 to November 11 is the best number of the day: 194 tasks closed, 1,977 agent runs (1,448 Claude Code, 529 Codex), 229.8 agent hours, 61.9M tokens and 4 hours 5 minutes of human time, about 56 agent hours per human hour. The rules that made it work are short: one task at a time, nobody reviews their own work, money, accounts, publishing and deletes stay with the human.
@rileybrown [Claude Code]
https://x.com/rileybrown/status/2121642720045261185
A two-minute 4K film about Clawd was made entirely in Blender from one prompt on Fable 5.2 inside Claude Code, in about four and a half hours. Claude wrote every asset as Python, five characters and props and seven full sets, plus every camera move across 52 shots, about 12,500 lines, and ran seven copies of itself to build the sets in parallel. It then rented 89 GPUs on RunPod, 36 at a time, to render 3,220 frames for a little over $20. The sharpest takeaway is about permissions, not prompts: the same prompt on a machine without the RunPod skill came out far worse, because a Claude Code that knows it can rent compute plans more ambitiously.
@KathleenLBooth [Claude Code]
https://x.com/KathleenLBooth/status/2121602030300975132
A marketing leader spends much of the day inside Claude Code, connected to the CRM, Slack, Google Drive, enrichment providers and project management, and logs into fewer tools every month. This week the job was an automation mapping the team's warm connections across 1,500 people to invite to an event, and the platform's API throttled at about 10 calls a minute, so the fix came from DMing someone at the vendor, who sped it up the same day. The buying checklist changed as a result: can it run headless with open APIs and a real MCP server, is that access on the plan actually being bought, and what does it do that could not be built in a weekend. A marketer asking a vendor to raise an API rate limit is the clearest sign of who the new software buyer is.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2121470943314788696
A maintainer of claude-code-templates rewrote the database-architect agent and tested it instead of just announcing it. The agent now reads the existing schema and migrations, flags locking risks and breaking changes, handles rolling deployments with lock_timeout and CREATE INDEX CONCURRENTLY, and tests migrations on a temporary database before finishing. In one test Claude got an 80M-row table and ten minutes before deployment: the same prompt scored 6 out of 10 without the agent and 10 out of 10 with it, and Claude picked the agent on its own in all 8 tests. The finding is precise: Claude was already good at schema design, and the agent's value is entirely in migrations that could break a running application.
@garrytan [Claude Code]
https://x.com/garrytan/status/2121464508225053153
Garry Tan ran 52 coding tasks in Claude Code 2.1.322 and came back with a plain recommendation: keep Opus 5.5 as the coding default. Fable 5.2 solved about as many tasks, 82.7% against 84.6%, but cost 2.80 times as much per solved task. The caveat is honest too: Opus already solves most of these, so telling the two models apart will need harder tasks. In a week when much of the timeline argued about model routing in the abstract, this is one of the few posts with a measured cost per solved task behind the advice.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2121567676741091784
A Japanese creator says that for explainer-style YouTube videos that do not need heavy craft, Claude Code or Codex alone is now enough: jump cuts, captions, sound effects, background music and animation all run automatically. The concrete result is close to one million yen a month in outsourced editing that no longer gets paid. The workload is also getting bigger, not smaller, with overseas vlogs of more than 500 source files now handed to the agent. Video editing keeps showing up as the clearest non-coding job where Claude Code replaces a paid vendor rather than a hobby.
@Ekko237 [Claude Code]
https://x.com/Ekko237/status/2121641744785633434
A fan of the discontinued Eyezmaze GROW series built its successor with Claude Code and Godot: Island Evolution, a deserted island with eight things to place, fire, fish, trees, a boat, water, wind, people and stone. Each placement levels up what is already there, so order matters, and the wrong order means wind blows out the fire or waves sweep the people away. The builder then had every one of the 40,320 possible orderings played out and found that only 3 reach the full 27 points. Brute-forcing the whole solution space of a puzzle game is the kind of balance check a solo developer would never have done by hand.
@manoj_surya_ [Claude Code]
https://x.com/manoj_surya_/status/2121548186896760988
A B2B growth marketer of about a decade learned to code and built MarketHQ with Claude Code, a tool for being found in search, communities and AI answers at once, after checking several founders' brands and finding the AI usually recommended their competitor. The product went into private beta in late September, the builder ran 10 demos with founders and marketers before launch and rebuilt parts of it from that feedback. On November 13 the first paying customer arrived at $59.40 a month, 40% off for the first ten. The useful detail is the order: interviews and rebuilds before revenue, not a weekend build followed by silence.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2121452059409142116
A builder who runs Claude Code, Codex, Goose, Aider and OpenHands at the same time built Campfire for the layer nobody else is funding: governance across several agents on one codebase. Each agent works in its own git worktree, risky tool calls go to a permission vote where the majority approves but any single deny blocks, with a 30-second timer so nobody stalls the run. Sessions can be forked at any message, replayed at speed, or run as a race across every backend with the results diffed, and a cost dashboard tracks spend per agent. The pitch is a sharp one: billions of dollars are betting on a single agent winning, while serious users already run five and need someone to settle disagreements between them.
@ewind_dev [Claude Code]
https://x.com/ewind_dev/status/2121407618300432743
The team behind ThreeRT explained why they built a renderer: impressive agent-made games mostly depend on a local machine with Claude Code, Blender or Unreal and a strong GPU, and moving that into a serverless sandbox breaks frame capture. Without a GPU the agent's screenshots fall back to software rendering, a slightly heavy shader hangs the run, and a black screen gives the agent no useful feedback, which may be why agents write poor WebGPU. ThreeRT is meant to let the creative agent loop run on a Cloudflare Worker and call a GPU render service only when it needs a frame. While optimizing it, Opus pulled off real-time full path tracing beyond expectations, so that part was split out and open-sourced, with a macOS client of 30MB against Unreal's 100GB editor.
@vishalsingh2972 [Claude Code]
https://x.com/vishalsingh2972/status/2121604612147089509
A Spotify engineer cut Claude Code token use by about 90% on bulk-reading tasks by making Claude do less, not by picking a smarter model. Two cheaper agents on Gemini Flash, run through Spotify's Portal, take over the grunt work: one reads large files and hands Claude a short summary, the other writes tests, configs and boilerplate from existing patterns. The first attempt used CLAUDE.md to tell Claude when to delegate, and Claude simply ignored it, so the final plugin, Shunt, uses hooks that block whole-file reads over 350 lines, scripts that route work to the cheap agents, and skills for what to do next. The trade-offs are stated plainly: delegated calls take 10 to 30 seconds, so small tasks are not worth routing, and debugging and precise edits stay with Claude.
@LargitData1 [Claude Code]
https://x.com/LargitData1/status/2121626161822552421
A developer took an action game first made with Codex, Mo Ren (Ink Blade), and expanded it with Claude Code to test the claim that Fable 5.2 is better for games. The new version runs to ten chapters across eight regions, 11 enemy types, and 40 models and motion files instead of 5, plus treasure collection and a map system. The biggest change is the art: the hero and enemies moved from code-posed figures to real skeletons with motion-capture animation, and the signature move now pulls in for a close-up, gathers ink, then chains a spinning slash into a landing seal. The builder says every part of it, art, 3D and music, came from telling Claude Code the task, and the game is playable from a link.
@scarlett_bouya [Claude Code]
https://x.com/scarlett_bouya/status/2121424320090235217
A Japanese solo operator connected Claude Code, Codex, Grok, Antigravity and Blender inside Muse, all within existing plans and free tiers, and moved from doing everything in sequence to directing a parallel staff. For blog articles Claude Code mechanically checks every link and pricing table against the source while Codex rereads the text, a fix that came from a past incident where an AI writing tool published an outdated price. For backtesting investment methods, Claude Code writes the test code, Codex reviews only for look-ahead bias, and the final decision stays human, and the first round already decided whether to adopt the first method. The honest limits are listed too: 3D characters were rejected as not good enough yet, and payments and contracts stay out of the agents' hands.
@f0rgent [Claude Code]
https://x.com/f0rgent/status/2121621941916147959
The night Claude Code got mods, one user told it to build a radar that watches its own agent. The panel sits next to the chat and shows every tool call, every file and every command live, and anything dangerous, rm -rf, a force push, a read of .env, lights up red. A Guard Mode goes one step further and simply blocks those actions. On a day when most people used mods for animated backgrounds, this is the use that matters: the new extension point turned into an observability and safety layer the agent cannot talk its way around.
@Annthonny [Claude Code]
https://x.com/Annthonny/status/2121400063859781748
One user replying to the theme launch says mods have lately become the favorite part of Claude Code. The build is a usage bar placed right above the prompt that shows tokens, cost and the 5-hour and weekly limits as the session runs. A custom theme is next on the list, but the first thing worth building was the meter. It is a small thing that answers the most common complaint of the week, people finding out they hit a limit only after they hit it.
@DamianPanes [Claude Code]
https://x.com/DamianPanes/status/2121420857696203095
A developer had Claude Code on a Mac and Codex on a server each working on their own copy of the same repository, and got tired of the push, pull and copy-paste-the-.env dance. So the developer built hilo: what one agent writes shows up on the other machine in about 50 milliseconds, with no push, no pull and no copying environment files. It is a narrow tool for a new kind of problem, two agents on two machines that need to see the same working tree. The post ends with the right question for anyone running mixed setups: how do you keep your agents in sync?
@PUssY931 [Claude Code]
https://x.com/PUssY931/status/2121393425610310063
A Japanese builder described a two-level task dispatcher sitting above Claude Code and Codex. The first level picks the executor by difficulty and by which system is better at that kind of work: Codex has four difficulty profiles, Claude runs on the normal model by default and saves the top model, which drains the quota fastest, for the hardest jobs only. The dynamic part is the interesting part: the remaining weekly quota of both systems is measured from execution logs, and work shifts to whichever side has room, with machine load checked too. On failure the dispatcher escalates to a stronger model or re-routes across systems, and since it holds all the state itself, one side going down does not lose the work.
@Sameek_Kundu [Claude Code]
https://x.com/Sameek_Kundu/status/2121694107781459989
A developer who works with Claude all day built Wisp, an AI-first browser for Windows, because Chrome kept getting in the way: the Claude-to-Chrome connection dropped, sessions hung, and Chrome wanted its window in front and the mouse and keyboard free. In Wisp, Claude Code, Codex or a Python script open tabs, read pages, click and type in the background using logins already saved, with no window jumping forward and no stolen input. It doubles as a daily browser with separate profiles per Google account, an ad blocker and Chrome extensions, and saved passwords are encrypted with the Windows account. It is a clear example of a builder fixing the exact friction in their own Claude Code workflow and shipping the fix as a product.
@kinopee_ai [Claude Code]
https://x.com/kinopee_ai/status/2121590428684423333
A Japanese user handed a rebuild of a working app to Claude Code on the CLI, Fable 5.2 handing off to Opus 5.5, and was surprised by how anticlimactic it was: it simply worked. Every spot that looked risky checked out, and the only visible difference was a slight change in design. Most of the time went into checking the user's own procedures and settings, while the time actually handed to Claude came in under two hours. Staff verification is scheduled for after the weekend, which is the right order: the agent finished fast, and the humans still sign off.
@ryan_doser13 [Claude Code]
https://x.com/ryan_doser13/status/2121415319579234370
An SEO practitioner showed how a dull site-wide chore changed. Updating 2025 to 2026 across a whole site the WordPress way means opening every page, editing the title and body, clicking Update and going back to the dashboard, page after page. The new way is one slash command in Claude Code that runs in the background while other work continues, and Cloudflare deploys the result automatically. It is a small example, but it captures why people are moving content sites off click-heavy CMS dashboards: the same SEO chore, zero clicking.
@film42 [Claude Code]
https://x.com/film42/status/2121405757577199891
With GitHub underperforming, one engineer pointed Claude Code at the GitHub organizations they pay for to see whether costs could come down. The answer was yes, by a lot. The savings were large enough that, in the author's words, Claude just paid for itself for the rest of the year. Auditing a SaaS bill is a quiet but high-value use: the agent reads the org settings and seats a human would never patiently go through.
@JacqBots [Claude Code]
https://x.com/JacqBots/status/2121425004000608402
Someone stole a pair of Ray-Ban Meta glasses from a traveler at an airport, and the traveler vibe-coded an iPhone app with Claude Code to find them. The app, Glasses Guardian, offers a radar-style finder, a loud ping and a GPS alert when the glasses disconnect, and it is now on the App Store, with a clear note that it is not affiliated with Meta or Luxottica. It is the purest version of the personal-software pattern: a bad day turned into a shipped product before the frustration wore off.
@josheyxyz [Claude Code]
https://x.com/josheyxyz/status/2121416341207093606
One builder's agents are running a partner's Squarespace shop even though Squarespace has no MCP for editing pages. One agent keeps watch over the site and opens an issue whenever something looks off, then Claude Code reads the issue, opens Chrome, goes into Squarespace and fixes it in the settings the way a person would. The builder had been dreading all that manual clicking and now barely touches it. The workaround is the lesson: when a platform has no API, an issue tracker plus a browser-driving agent becomes the API.
@VaibhavSisinty [Claude Code]
https://x.com/VaibhavSisinty/status/2121623397767778549
A Reddit builder made an entire swimmable ocean in a single 97KB HTML file with Fable 5.2 in Claude Code: kelp forests, a sunken ship, manta rays, schools of fish and light rays from the surface, all generated by code at load time. It started from one prompt, and a visual-critic subagent kept checking the renders until they looked right, for about two to three hours of actual model work. It needs WebGPU, so Chrome is the safe bet, and the file is public. The visual-critic subagent is the reusable idea: the agent that writes the scene is not the one that judges it.
@Sauers_ [Claude Code]
https://x.com/Sauers_/status/2121409816522604555
One user asked a Claude Code agent what its own sub-agents were called and got back a list of more than a hundred names: how2, devv2, factory2 through factory4, kind1 to kind8, covA to covF, rlsft, rlenv, mechinterp, prtriage, deslop, podsitter, diskfree and many more, with some redacted. Read as a list, it is a snapshot of what a long-running research setup actually spawns: forks, benchmarks, data moves, disk cleanup, pod babysitting and PR triage. Nobody designed that taxonomy by hand. It is a reminder that once agents start spawning agents, naming and inventory become a real management problem.
@rushiagr [Claude Code]
https://x.com/rushiagr/status/2121461764995068096
A developer implemented a compact Opus mode for Claude Code. Instead of waking the main agent every time a subagent posts an update or finishes, it launches all the work as a wave of subagents and only wakes up when the whole wave is done, then plans and launches the next wave. The point is to stop the expensive orchestrator from burning context and tokens on a stream of partial progress reports. It is a simple scheduling change that targets one of the week's loudest cost complaints, the coordinator that keeps reading messages it does not need.
@kaoruokasaku [Claude Code]
https://x.com/kaoruokasaku/status/2121418187938132448
A Japanese user wrote up how adding decision criteria, called Axes, changes the human's role with Claude Code. Before, Claude Code either followed fixed Rules or came back to ask. With Axes, Claude Code decides on its own but must report which Axis it used, the human reads the reports and stops it when the judgment drifts, and once an Axis goes a long stretch without needing a stop, it gets promoted to a Rule. The analogy is a school club: a new member asks the senior about everything, then acts alone while the senior watches, until the rule is in the handbook and nobody needs to ask. It is a concrete way to graduate an agent's autonomy instead of granting it all at once.
@kamonohashi1986 [Claude Code]
https://x.com/kamonohashi1986/status/2121396500337193387
A Japanese blogger counted the suggestions from Claude Code that were turned down over six weeks since October: 20 in total. The most common reason, 7 times, was that another method was already good enough. The ledger of article themes holds 28 entries, 13 written and 13 declined. Keeping a log of rejected agent suggestions is unusual and useful, because it shows where the agent's ideas are redundant rather than wrong.
@AIFantasyLeague [Claude Code]
https://x.com/AIFantasyLeague/status/2121424045782745562
A fantasy football account running roster decisions through Claude Code reported a very specific failure: the agent dropped Colston Loveland, one of the best tight ends in football, because the team had a bye week. The joke in the post is that the agents seem to believe players stop existing when their teams do not play. The real point is the gap between reasoning and domain context: the agent optimized for this week and forgot that a bye is temporary. Automating a hobby league is a cheap place to learn that an agent's decision rule needs the season, not just the slate.
@builtwithrubel [Claude Code]
https://x.com/builtwithrubel/status/2121670750046494793
One builder posted the launch math in four lines: the app took 2 hours with Claude Code, the landing page 20 minutes, Stripe setup 10 minutes, and finding a single paying customer is still loading. The conclusion is blunt: shipping is the easy part now, and getting anyone to care is the job. It sits next to several other posts that day making the same point from the other side, founders who can build in a weekend but cannot get ten people to pay ten dollars. Build speed has collapsed, distribution has not.
@QuickscopeFPS [Claude Code]
https://x.com/QuickscopeFPS/status/2121529490954203316
An indie developer who missed the sniping lobbies of 2009 vibe-coded Quickscope FPS with Claude Code, a sniper game that is now free on iPhone. The showcase clip is one bullet, two kills through a wall with no scope, then a collateral for a four-kill feed, all within 17 seconds. It is a nostalgia project built to a specific feel rather than a genre, which is exactly where solo builders with Claude Code are now shipping finished mobile games.
@yukix2000 [Claude Code]
https://x.com/yukix2000/status/2121748042365993360
A Japanese builder split work across seven agents with separate jobs: Grok Bot as the main front desk and X lead, Muse as the sparring partner for posts and video, Claude Code as the main entrance for daily conversation and work, Codex as reviewer and software factory for a goal app, Cursor for long-running maintenance, Devin for heavy builds left to run, and Gemini or Genspark for research only. The reason it holds together is a self-built goal-management app that serves as the shared context: goals, current position and every decision live there, and every agent reads it. Without that, the builder says, any division of labor falls apart fast. Roles can be split, but context has to stay single.
@fujii_enjineer [Claude Code]
https://x.com/fujii_enjineer/status/2121553894283370770
A Japanese seller in e-commerce shared the setup that handles most of the work: an orchestrator on Opus or Fable and workers on Sonnet, Sol or Terra, run in parallel with a build, evaluate, fix loop, and the final call made by a human. Simple jobs go to Claude Code or Codex alone, and one-off tasks that need judgment but do not justify building a system go to the Muse Chrome extension. For anyone without Codex, the advice is to tell Claude to substitute a Claude model in that slot. The broader point is that using AI has moved from picking one tool to assembling a cost-effective crew, with creative work as the exception.
@DuncanRogoff [Claude Code]
https://x.com/DuncanRogoff/status/2121412640450998652
A video editor of fifteen years says the most useful skill brought to Claude Code was giving notes. Saying make it better gets nothing from an editor or from Claude. Saying the cut lands a beat late and the eye goes to the wrong corner first gets a real fix. It is a compact argument that the scarce input is no longer typing code but precise, observed feedback, the thing experienced reviewers already know how to give.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2121392839435407576
On a product podcast, a former CPO of Bolt said there was a time when talking to seven or eight customers a month felt like enough, and with AI it is now a thousand times that or more. The guest, now CEO of Woven, built the customer research agent in Claude Code in 30 minutes before the recording, then added the safety check live on the show. The episode walks through emailing drop-off users and adding an eval loop on air. The notable part is the format: a research agent that does customer discovery at scale, built in half an hour, with the guardrail added in public.
@channel9lambda [Claude Code]
https://x.com/channel9lambda/status/2121422753849053687
One user's Claude Code kept answering 403 Attention Required, Cloudflare, and running /login did nothing. The diagnosis is worth keeping: that page is Cloudflare's block page, not a login error, so the fix is not re-authenticating. Curl the API directly, and if it returns application/json the request reached Anthropic, while text/html means Cloudflare blocked it. Then turn the VPN off, try another network and check the status page. A one-line test that separates an auth problem from a network block saves a lot of pointless logins.
@aabyzov [Claude Code]
https://x.com/aabyzov/status/2121747514919399706
Claude Desktop now runs on an Ubuntu box in the official Linux beta with Chat, Cowork and Claude Code, but dictation and computer use are not in the Linux build yet. One user filled the voice gap with SpecWithCode, an open-source tool that talks to Claude Code on Linux using local voice: whisper.cpp for speech to text and Piper for speech output. It is a typical Linux-user move, patch the missing piece locally and open-source it, and it shows voice is becoming a default way to drive Claude Code rather than a novelty.
@iamhuy [Claude Code]
https://x.com/iamhuy/status/2121422204466315566
A builder on day two of building an open-source analytics engine in public shared an almost free agentic data stack: DuckDB self-hosted or MotherDuck's free plan as the warehouse, dlt and dbt for ELT, a static-typed semantic layer for metrics, and the builder's own open-source engine plus custom BI apps on top. Hosting is usually a single $7 a month Hetzner box, and context comes from markdown files, dbt descriptions or Notion. The whole stack is set up with coding agents, Claude Code, Codex or OpenCode, and the builder says they rarely look at the code anymore. Everything lives as code in Git, which is what makes letting the agents build it safe enough.
@iAmHenryMascot [Claude Code]
https://x.com/iAmHenryMascot/status/2121427612186681473
A user who keeps skills outside the harness so any agent can use them ran a rule-placement test: a skill never mentioned in the prompt worked in Codex 30 of 30 turns and OpenClaw 29 of 30, and did nothing in Hermes or Claude Code. Those two saw the skill's name on every request and opened it 0 times in 30 turns, and since three of the harnesses ran the same GPT-6.1 Sol model, the gap points to the harness, not the model. Naming the skill in the prompt fixed it on Claude Code, and a rule given once faded on later turns while a per-turn hook kept working. The numbers are early, 30 turns per setup across 10 public tasks, but the lesson is concrete: a shared skill is not done until each harness is told to use it and checked.
@thoughtcrime___ [Claude Code]
https://x.com/thoughtcrime___/status/2121394969277845766
After people kept asking about a screenshot, one builder named the app: speedrun, which reads AI agent logs and shows what got done each day next to how long an expert in that field would have taken. It works with Claude Code, Codex and Cowork, and the repo is public. It turns the vague feeling of being more productive into a daily comparison against a human baseline. Whether the expert estimates are fair is open, but the instinct, measure the output rather than the token count, is the right one.
@masaakiotadev [Claude Code]
https://x.com/masaakiotadev/status/2121410995516305691
A Japanese developer updated Tomarigi, a desktop app that keeps Claude Code, Codex and Gemini CLI sessions in an always-on-top window so they can be watched without switching. Version 0.17.0 stops the Claude desktop app's side chat, the btw thread, from showing up as its own session. Sessions handed off by Tomarigi's concierge to another session now appear with a distinct marker. Small release notes like these show the session-watcher category maturing around the new Claude Code features within a day of their launch.
@ballanutsa [Claude Code]
https://x.com/ballanutsa/status/2121529076892701126
One user wrote a sober counterweight to the week's agent hype, arguing the gold-rush panic is itself the product being sold, and that agents earn their keep on narrow, repeatable tasks where the output is easy to verify and mistakes are cheap. The example given is exactly that kind of task: Kaito's MCP connected to Claude with a weekly routine that pulls the smart-followers data for a recap, so the human only takes over editing the draft. The post also quotes the claims it distrusts, including Claude Code doing three weeks of work in four hours. Read together, the claim and the example define a realistic scope: automate the data pull, keep the judgment.
@telegram_man [Claude Code]
https://x.com/telegram_man/status/2121597125314203968
A builder laid out a two-tier economics for agents: use the strongest models, through Codex or Claude Code subscriptions, to build the machine, and use a cheap model on a usage-billed API to run it around the clock. The example is a spare-parts shop where an agent finds a compatible washing-machine pump, checks stock, contacts suppliers, agrees an option with the customer and pays within a preset limit. For an obscure bug or architecture work the cheap model trails Sol 6.1 and Opus 5.5, but with a prepared workspace and a defined process it runs well at very low cost. The post also flags a rule people miss: consumer Claude Code and Codex subscriptions are not for serving other users' requests, so a 24/7 customer workflow has to sit on an API.
@BOLTTECH_WORKS [Claude Code]
https://x.com/BOLTTECH_WORKS/status/2121422870140043373
A Japanese channel publishes a daily children's science short at 11 every morning, and the credits read production support Claude Code, with voice and music from ElevenLabs. Episode 13 explains an Archimedes screw: characters lift water from a pond to a goldfish seller's tank by turning a 3-meter screw tube instead of carrying buckets. It is a quiet example of Claude Code as the production pipeline for an educational series rather than a coding tool, with a fixed daily schedule that only works if the pipeline is reliable.
@gregoftheweb [Claude Code]
https://x.com/gregoftheweb/status/2121401869809696794
One builder's latest lesson: hook Claude Code up to a browser and let it handle the app submissions to the Apple App Store and Google Play. The store consoles are long, form-heavy and different for each platform, the kind of clicking that eats an afternoon after the app itself is done. The author calls it a game changer, and it fits a pattern seen across the day, Squarespace edits and App Store forms, where browser-driving agents take over the last-mile paperwork of shipping.
🗣 User Voice
User Voice
Parallel threads need a visible cost before they run. @MartinSzerment points out that every Projects thread is a full Claude Code session, so one message with three tasks drains the limit three times in parallel, and quotes a usage panel showing 104 threads and 2.8B tokens while the coordinator sat at 0%. @IrfanLateefUS runs out of Claude Code limits by Wednesday with the reset on Saturday and wants a clean way to move tasks to Codex without losing quality. People want thread model, effort and concurrency set before the batch starts, not discovered on the bill.
Tell me which model actually ran. @bebuu_ai relays the Artificial Analysis finding that a flagged task can be silently moved to an older model mid-session, 8.78% of runs for Opus 5.5 max, which would explain why Claude Code sometimes feels weaker halfway through. A visible model indicator per turn would settle a lot of arguments.
Long-running agents need painless upgrades and honest scheduling. @fkysly runs several Claude Code instances around the clock on a Mac mini and has to restart every agent to pick up a version with a critical fix, and @divid_lu49394 notes that Friday's outages delayed scheduled tasks by over two hours and skipped some, so unattended jobs need retries and alerts.
Mods and skills are powerful, and that cuts both ways. @Ldataengineer warns that a mod runs with the user's permissions and is not sandboxed, so a shared theme deserves the same caution as any download, and @iAmHenryMascot found a shared skill opened 0 times in 30 turns in Claude Code until it was named in the prompt.
Still read every diff. @fialaerik had Claude Code nail the boring parts of a billing refactor and then confidently call a Stripe method that does not exist, and @samongaro_ barely opened Cursor this month because Claude Code in the terminal just does the job, yet still wonders what is being missed.
Parallel threads need a visible cost before they run. @MartinSzerment points out that every Projects thread is a full Claude Code session, so one message with three tasks drains the limit three times in parallel, and quotes a usage panel showing 104 threads and 2.8B tokens while the coordinator sat at 0%. @IrfanLateefUS runs out of Claude Code limits by Wednesday with the reset on Saturday and wants a clean way to move tasks to Codex without losing quality. People want thread model, effort and concurrency set before the batch starts, not discovered on the bill.
Tell me which model actually ran. @bebuu_ai relays the Artificial Analysis finding that a flagged task can be silently moved to an older model mid-session, 8.78% of runs for Opus 5.5 max, which would explain why Claude Code sometimes feels weaker halfway through. A visible model indicator per turn would settle a lot of arguments.
Long-running agents need painless upgrades and honest scheduling. @fkysly runs several Claude Code instances around the clock on a Mac mini and has to restart every agent to pick up a version with a critical fix, and @divid_lu49394 notes that Friday's outages delayed scheduled tasks by over two hours and skipped some, so unattended jobs need retries and alerts.
Mods and skills are powerful, and that cuts both ways. @Ldataengineer warns that a mod runs with the user's permissions and is not sandboxed, so a shared theme deserves the same caution as any download, and @iAmHenryMascot found a shared skill opened 0 times in 30 turns in Claude Code until it was named in the prompt.
Still read every diff. @fialaerik had Claude Code nail the boring parts of a billing refactor and then confidently call a Stripe method that does not exist, and @samongaro_ barely opened Cursor this month because Claude Code in the terminal just does the job, yet still wonders what is being missed.
📡 Eco Products Radar
Eco Products Radar
Codex: the constant second seat, as builder lane, reviewer and fallback when Claude Code limits run out.
Cursor: compared head to head with Claude Code, and in the news for the tampered Claude Code extension in its marketplace cache.
Grok Bot: positioned as the manager or front desk that hands coding work to Claude Code.
Antigravity: Google's agent environment, mostly mentioned next to Gemini 4 Carbon and Hydrogen.
Muse: Meta's agent, used as a virtual machine host and Chrome extension next to Claude Code.
Blender: the creative workhorse, with skills packs and MCP servers letting Claude Code model, animate and render.
Hermes: the open agent harness people compare against Claude Code for skills and routing.
Orca, Herdr and OpenMausBot: tools for watching and running fleets of coding agents from one place.
Remotion: the default for turning code into video in agent-driven editing setups.
Obsidian: the shared brain that agents write their progress into.
claude-mem: persistent memory across Claude Code sessions.
Modern Web Guidance: Google Chrome's skill pack for modern HTML and CSS, installable in Claude Code.
Tailscale: used to reach local Claude Code previews and sessions from a phone.
AgentRun, Plannotator, Elephant-Goldfish and Alibaba open-code-review: the week's trending add-ons for memory, plan review, design review and code review.
Codex: the constant second seat, as builder lane, reviewer and fallback when Claude Code limits run out.
Cursor: compared head to head with Claude Code, and in the news for the tampered Claude Code extension in its marketplace cache.
Grok Bot: positioned as the manager or front desk that hands coding work to Claude Code.
Antigravity: Google's agent environment, mostly mentioned next to Gemini 4 Carbon and Hydrogen.
Muse: Meta's agent, used as a virtual machine host and Chrome extension next to Claude Code.
Blender: the creative workhorse, with skills packs and MCP servers letting Claude Code model, animate and render.
Hermes: the open agent harness people compare against Claude Code for skills and routing.
Orca, Herdr and OpenMausBot: tools for watching and running fleets of coding agents from one place.
Remotion: the default for turning code into video in agent-driven editing setups.
Obsidian: the shared brain that agents write their progress into.
claude-mem: persistent memory across Claude Code sessions.
Modern Web Guidance: Google Chrome's skill pack for modern HTML and CSS, installable in Claude Code.
Tailscale: used to reach local Claude Code previews and sessions from a phone.
AgentRun, Plannotator, Elephant-Goldfish and Alibaba open-code-review: the week's trending add-ons for memory, plan review, design review and code review.
Comments