Super User Daily: 2026-10-11
Thursday on the Claude Code feed was the day the tool left the editor. One person pointed a Claude Code pipeline at NASA light curves and came back with two Earth-sized exoplanet candidates, another watched Claude find a years-old production bug, ship the fix and leave the operator a to-do with the exact ssh commands, and a Japanese sales shop described pulling mid-term plan numbers out of 367 annual reports into individually tracked letters. The second big theme was cost discipline with real ledgers: a 72-hour log showing Opus only changed the outcome in 37 of 1,846 tasks, a turn-by-turn breakdown of why Sonnet is not cheaper once the context passes 150K, 350 Haiku agents for nine dollars, and a five-figure Cloudflare bill traced to one Codex commit. The three-model org chart (Sonnet builds, Haiku swarms, Opus advises) was posted by at least a dozen accounts as an official Anthropic tip, and one fact-check found no such bundled plan in the docs. Video was the non-coding frontier again: ads, short films, 960-frame animations and seminar decks all rendered as code. On the OpenClaw side the foundation won the .claw top-level domain, while actual users filed a WebSocket bug with a traced fix, reported cron prompts leaking into Slack, and read a 112-PR patch that was mostly about surviving its own upgrades.
@paraschopra [Claude Code]
https://x.com/paraschopra/status/2108486193419882775
Inspired by a Reddit thread about someone finding a planet with Claude Code, the author spent one morning running a Claude Code pipeline over NASA TESS data for nearby stars that TESS only scanned at low cadence. The pipeline surfaced two undocumented exoplanet candidates: a star in Cygnus 158 light-years away that dims 0.02 percent every 3.56 days (about 1.0 Earth radii) and one in Andromeda at 134 light-years dimming 0.03 percent every 4.93 days (about 1.1 Earth radii). The dips repeat 55 and 21 times across four observing years, match in both NASA and MIT processing, and survive a hold-one-year-out prediction test and a flipped-light-curve fake-signal check. Codex and a fresh Claude reviewed the method, with the caveat that a faint eclipsing binary neighbor could mimic the signal, and the next step is filing both as community TESS objects of interest.
@mhmazur [Claude Code]
https://x.com/mhmazur/status/2108661854608203932
While working autonomously on the author's SaaS, Claude Code found a small bug introduced years earlier, opened a PR, deployed the fix on its own, then measured how many production records had been affected. Because it had no write permission to the production database, it wrote a Ruby script to repair those records and added a backlog item containing the exact scp and ssh commands needed to copy and execute the script on the server. The author ran those commands several days later. The reflection that follows is that the human has become the bottleneck for Claude, and that this is what the next stretch looks like.
@shmily7 [Claude Code]
https://x.com/shmily7/status/2108493941003972778
A long post-mortem on a runaway Cloudflare bill, now fully refunded after the Cloudflare team traced it to a platform bug. The project was an experiment in splitting dev tasks across Cloudflare Sandboxes with a lead agent and a queue of coordinating sub-agents, which did not show a real efficiency gain and was shelved. It stayed deployed, and a late-August commit made through Codex contained a Durable Object infinite loop that woke up 23 days later and started billing. The author had used Claude Code with Fable-series models for a long stretch without serious incident and only switched to Codex when the accounts ran out, and the lesson drawn is that the amount of review you owe scales with the model you pair with. The post also lists the billing-alert thresholds to configure and notes Cloudflare has no hard spend cap yet.
@vmrmax [Claude Code]
https://x.com/vmrmax/status/2108614496272965948
Grok Bot now installs Claude Code on its own machine and runs Opus on the author's existing Claude plan, so the author logged every one of 1,846 tasks over 72 hours to see when Opus actually mattered. 214 tasks were routed to Opus; re-running the cheap model on the same 214 produced the same decision 177 times, so Opus changed the outcome 37 times. Proposals, contracts and site code accounted for 72 Opus calls and 25 of those 37 changes, while client replies sent 92 to Opus and 85 came back as the same decision in nicer words. The best catch was a contract clause allowing 60-day late payment with no penalty, and the worst miss was Opus writing the year's best proposal at a stale $120 hourly rate pulled from memory. At API prices the 214 calls were $51.36 against $443.04 for everything on Opus.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2108640573724786970
The author pointed the new Claude Managed Agents dynamic workflows at an 18,000-line codebase using the bug-hunter onboarding command inside Claude Code. The run produced 21 proven bugs in five minutes for $8.36, paid from the Claude Max API credits. The mechanism is a lead agent that writes a plan, fans it out across many agents in phases and merges the results. The author's verdict is that it burns tokens fast and is worth it.
@emooove [Claude Code]
https://x.com/emooove/status/2108369311417192662
A Japanese B2B sales-support company upgraded every employee to the Claude Max plan and listed what has changed since AI took over the work. Claude pulled mid-term plan figures out of the securities reports of 367 listed companies above 500 billion yen in market cap and merged them into individual letters, each printed with its own QR code so the firm can see which company and which person opened it. Inbound inquiry and recruiting forms get a first-pass AI triage that removed a median 3.1-day wait, and an analysis of visitors who said they found the firm through AI search showed only 17 percent arrived via AI directly, the rest Googled the company name afterward. The stack also includes a Claude plus Notta link that turns every meeting into a live source, reply drafts that learn from the diffs of the sender's corrections, and an in-house sales system with one-click calling and automated daily reports.
@LachezarVoynov [Claude Code]
https://x.com/LachezarVoynov/status/2108585587494019562
A first real attempt to automate ad editing end to end: a Suno song ad was fully edited by an internal video tool the author is building inside Claude Code. The flow starts from a brief written by a creative strategist, an AI creative director breaks it into more than 100 coherent scenes, each scene is generated with Nano Banana from a custom prompt and animated with Kling 3.0, the song is produced automatically with Suno, and all scenes are assembled, captioned and cut inside Claude Code. It took four revision rounds with human direction, and the agents are hooked to a review tool where the strategy team's comments become revision prompts immediately. The stated goal is turning 100 creative briefs into finished ads within 24 hours by the following Monday, and the post argues the video editor role in direct-to-consumer marketing has become optional.
@jun_liang_sf [Claude Code]
https://x.com/jun_liang_sf/status/2108635673624527271
An October edition of a recurring eval that asks whether coding agents can actually use a given API: Claude Code on Sonnet 5 and Codex on GPT-6 Sol were run against more than 120 live production APIs across 36 categories, over 1,000 eval runs and more than $5,000 in spend, each with a realistic developer task in an isolated sandbox. Only 17 of 121 APIs scored 80 or above. Category leaders included Stripe for payments, Firecrawl for search, Daytona for sandboxes, OpenRouter for inference, WorkOS and Auth0 for auth, Chroma for vector storage, ElevenLabs and Deepgram for voice, Reducto for document parsing, Browserbase for browsers and Prefect for durable workflows. Scores combine discoverability, usability, tool calls and errors, and the full results are public.
@fleyta88 [Claude Code]
https://x.com/fleyta88/status/2108491564435591582
A cost breakdown arguing that most people benchmark Opus 5.5 on the wrong part of the loop. Fresh-token prices make Opus look like 2x Sonnet, but cached history is where the gap collapses: per turn Opus is 1.88x at 20K context, 1.49x at 150K and 1.26x at 400K, so the real question for a long-running agent is how many turns each model needs to finish. At 150K context, 30 Sonnet turns cost about $1.76 and 20 Opus turns about $1.75, so staying on the cheaper model can produce the same bill ten turns later. Switching has its own price: a late 300K handoff into Opus costs about $1.50 just to rebuild the cache against about $0.10 for a clean 20K handoff. The proposed loop is Sonnet on scoped work with periodic real checks, repeated failure as the escalation signal, and Opus starting fresh on the hard branch with only the useful state crossing over.
@Lniosytest [Claude Code]
https://x.com/Lniosytest/status/2108425085891572002
A fact-check of the three-model setup that circulated on X all day as the official Anthropic plan. The author read the docs: there is no such bundled scheme, but the parts are real. Sonnet 5.5 can be the main coder, Haiku 5.5 subagents can handle file and doc lookups, and Opus 5.5 can act as advisor, called in to set the plan, on repeated errors and before wrap-up. The advisor is enabled with a single command-line flag and is still experimental, Haiku subagents have to be configured by hand, and the widely quoted 340 tokens per second figure has no official source the author could find.
@dravenip [Claude Code]
https://x.com/dravenip/status/2108665279806869713
Takeaways from a 90-minute walkthrough in which the person who built Claude Code showed how agents are used inside Anthropic, several of which cut against common practice. The best model gets used even for small tasks, on the argument that the cheap one burns more tokens getting it wrong; most tasks start in plan mode so no code is written until the approach makes sense; agents get a goal and tools rather than a step-by-step script; and even finance staff on that team now write code. The author cannot stop thinking about the model point, since everyone they know downgrades to save money and the claim is that this is the expensive choice. The open question the author wants answered is where the line sits between quick edits and bigger tasks, and how much of the five agents and 30 PRs a day comes from the setup rather than the model.
@mylifcc [Claude Code]
https://x.com/mylifcc/status/2108498203691659711
A practical note on Claude Code concurrency limits. Workflow concurrency defaults to min(16, max(2, available cores minus 2)), so a 16-core machine runs 14 workflow agents at once, and the value can be set between 1 and 256 through the CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS variable; ordinary subagents have a separate default of 20 under CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS. The author's own data point: 350 short-lived Haiku 5.5 agents consumed 193,503,496 input tokens and cost nine dollars.
@HouseHackerJon [Claude Code]
https://x.com/HouseHackerJon/status/2108368914669592637
The author went back to Claude Code to test Opus 5.5, Sonnet 5.5 and Haiku 5.5 on a long-running skill, with Opus as orchestrator and Haiku as the worker. Haiku ran for almost five hours on simple but tedious tasks against a dev instance, produced genuinely good code, and used roughly five to ten percent of the weekly allowance. The complaint is about everything around the model: after using personal agents like Grok Bot and Muse, the secret-store experience of those systems versus juggling environment variables in Claude Code is night and day, and even with an API key and every connector the author could barely get Claude Code to deploy an app to Cloudflare. The broader verdict is that a coding harness is a waiter that takes orders and reports when stuck, while the personal agents only come back when truly blocked or done.
@ysbilgin [Claude Code]
https://x.com/ysbilgin/status/2108475281543442684
After a long stretch on Claude Code alone, the author added Codex and now has the two check each other's work. They are good at different things, so one often catches a bug or an edge case the other missed and tells the other to fix it. No special harness or tooling is involved: the author just told each one to prompt the other through the CLI and specified which model and effort level to use.
@ai_fudosan_ai [Claude Code]
https://x.com/ai_fudosan_ai/status/2108368664760397868
A Japanese builder got a mixed local-and-cloud, Claude-and-Codex agent team running on top of agmsg, a bash-and-SQLite messaging layer that lets different kinds of agents exchange messages through one place. The extension adds receipt confirmation, resend, long-message splitting and reply threading, and now a local Codex agent can wake a cloud Claude Code session with a single message. In the demonstrated loop, cloud Claude opens a PR, Codex reviews and requests changes, cloud Claude fixes and resubmits, Codex re-reviews, and the PR merges with no human carrying messages. Cloud-to-local travels through a Discord relay channel that a local process writes back into agmsg, and the whole thing runs from an 8 GB machine because the heavy work is pushed to the cloud.
@kristiandaaniel [Claude Code]
https://x.com/kristiandaaniel/status/2108684883996610761
Built a lead-agent workflow on top of T3 Code that coordinates Claude Code and Codex across the author's subscriptions. One approval starts the run, after which each task gets its own thread and git worktree, pauses are quota-aware so work stops before a plan limit is hit, and merges are verified before they land. The author shared the setup running live.
@onenewbite [Claude Code]
https://x.com/onenewbite/status/2108671448667865342
Unimpressed by Claude Code Projects, the author built a local equivalent and describes why. Three problems with the hosted version: the main thread acts as a messenger between you and the sub-threads rather than a brain, so you end up watching sub-threads anyway; existing local projects are hard to port because the hosted project has its own memory and the main thread lacks your local memory; and running in the cloud is impractical for many projects. The local alternative is framed as a campaign rather than an ongoing project: start one, implement a set of features, finish, start the next, so the main thread stays focused and strategy-heavy. The point is that Claude Code already has all the infrastructure needed to build this yourself.
@astralwave [Claude Code]
https://x.com/astralwave/status/2108510924646896070
A grounded critique of the coordinator pattern from someone who likes it. It feels more agent-native than pinned threads or an orchestrator, but the UX is rough: cascading permissions, open decisions and blockers still force the author into child threads to unblock them. The coordinators are arbitrarily hobbled compared with bare sessions in Claude Code, Codex or Cursor: the Claude Projects coordinator has no web browser and cannot use MCPs, Dot cannot run skills, see AGENTS.md, use the structured question tool or touch the local machine, and Grok Bot ignores the repo's skills and AGENTS.md. The spawned threads are hobbled too; in Claude Projects, skills cannot be invoked by hand inside project threads and there is no way to see which skills a thread has read.
@arnestrickmann [Claude Code]
https://x.com/arnestrickmann/status/2108674246427979930
Noting that nobody had built an all-in-one tool holding the third-party connections for personal agents such as Muse, Instinct, Grok Bot and Interaction, the author built one over a few days with a collaborator, open source and easy to self-host. The pitch is that this layer should be an independent product: auth stays separate from the agent, every action the agent takes can be audited, permissions can be fine-grained, and switching between agent apps becomes trivial. The immediate result is that the author's Instinct now has exactly the same tools as Claude Code on the laptop.
@trevin [Claude Code]
https://x.com/trevin/status/2108356625220391150
A low-key project launched a few weeks earlier to fix a specific annoyance: sharing docs and files between agents, and letting an agent produce a doc a human can hand to another human, should not require separate services for markdown, HTML, prototypes and zips or lock you into one vendor. The author uses it daily, with the skill installed in Grok Bot, Muse, Instinct, Claude Code and ChatGPT Dot. It self-hosts in your own Cloudflare account and works on the free tier.
@bart_walker [OpenClaw]
https://x.com/bart_walker/status/2108374310301642807
A practitioner coordinating Hermes, Codex and OpenClaw, with Grok Bot likely next, asks what mechanism others use to keep them in sync. The best the author has managed is a shared folder structure holding project status lists for everything in flight, referenced at the start of each task and updated at the end. It is a plain-files answer to a coordination problem that no agent vendor has solved across its competitors.
@EdgeDimi [OpenClaw]
https://x.com/EdgeDimi/status/2108528657606086729
A detailed bug report addressed to the OpenClaw maintainer about a ChatGPT-subscription failure on the 2026.9.8 runtime: when a Responses WebSocket closes with code 1000 before the terminal response event, the turn ends with no recovery. On production gateways this showed up as a scheduled automation failing after about 100 seconds with no retry, and a live chat turn failing after tools had already written business records. The author traced the path through the bundle: close code 1000 is not in the retryable set, so it becomes a non-retryable terminal error, and the one-line change in an open PR that adds 1000 to the retryable set reclassifies it as a transient transport error the existing retry logic already handles. A follow-up confirms two natural close-1000s on a live 9.8 gateway recovered through native retry with no duplicate tool calls once the patch was applied.
@_bashoh [OpenClaw]
https://x.com/_bashoh/status/2108702773889900830
A short but pointed report to the OpenClaw creator: the author's OpenClaw is leaking cron prompts into Slack, and they are asking for suggestions. The post reached more than 22,000 views, which says something about how many people run scheduled jobs through a chat channel.
@tvytlx [OpenClaw]
https://x.com/tvytlx/status/2108592072148017556
The author read the full OpenClaw 2026.9.9 release notes, which the announcement headlined as GPT-6.1 Sol and Haiku 5.5 support. The new-model section is a few lines; the longest section is updates and maintenance, about twenty items all about surviving a failed upgrade: rolling back a half-migrated database, whether it can be re-upgraded afterwards, backups on FUSE drives failing restore because file times changed, Windows rolling back the whole update after a first Doctor check, Docker on older Linux stuck in an upgrade loop over an unchanged database. Two very agent-flavored bugs: iMessage replies composed but never sent, and an old scheduled task timing out and killing the updated conversation in the same session along with its tool permissions. Useful detail: the haiku alias now points at Haiku 5.5 with adaptive thinking on, prompts over 100K tokens cost more, and interrupted one-off tasks may rerun after restart and send duplicates. The author decided to wait for 9.10.
@0xLalice [OpenClaw]
https://x.com/0xLalice/status/2108386043632533831
A French user runs OpenClaw on a VPS with GPT, gives it several varied missions every day, and says it copes fine. The setup recipe: install Codex or Claude Code locally, rent a VPS for about 20 euros a month, hand the local agent the VPS ssh key and ask it to download, install and configure OpenClaw on the server. The result is a self-hosted agent online around the clock on its own machine, reachable from Discord or Telegram on any device, which you then wire to email, GitHub, WordPress and your servers and give a personality and a sense that the VPS is its own responsibility. The claim is that an amateur can do it in ten minutes.
@fagamericano [OpenClaw]
https://x.com/fagamericano/status/2108618930709565480
An OpenClaw research task just completed: a cognitive engine that evolves the author's Maestro setup to Postgres-backed memory, semantic search and RAG across more than 500 gateways. The explicit break is with memory in markdown files. In its place are ingestion pipelines from Notion, ClickUp, Slack and meeting transcripts, security checks at the embedding gates, proper scoping at personal, channel, org and agent levels, and retrieval that goes through native SPIFFE identity-mesh calls.
@tankxu [OpenClaw]
https://x.com/tankxu/status/2108608546732708037
A Chinese user got an end-to-end AI-generated personalized news pipeline working on GrokBot, with news quality described as solid and a cloned voice of the author's daughter giving the spoken segments a human feel. A similar attempt on OpenClaw at the start of the year was unsatisfying; after most of a year of model and agent upgrades, the output quality now justifies the token spend. A companion website for the daily news is also up, and the author is ready to take custom-news work.
@roger9949 [Claude Code]
https://x.com/roger9949/status/2108476338550251735
A summary of a long interview in which a well-known crypto founder describes handing daily life to Claude Code. Meals get photographed before and after so Claude estimates calories, Oura Ring sleep data goes to Claude Code for a daily report, a coach films each set so the reps and weights are logged, and everything is stored as local markdown so a different AI could pick it up tomorrow; the stated split is 90 percent AI driving, 10 percent human adjusting. The same question is asked of three models and the middle answer taken. Other threads: data-heavy fields such as math, astronomy and biopharma, proactive health checks against an AI-written list, crypto as a payment layer for AI metered per API call, and token-resale arbitrage across countries where subscription prices differ.
@jurlycat [Claude Code]
https://x.com/jurlycat/status/2108405420721377550
For nearly ten years Ben's main way to communicate was turning the head for yes or no. A brother with no coding background used AI tools including Claude Code to build a hub Ben controls with simple switches, and now Ben can text, pick movies, browse the web and play games, with Claude Haiku suggesting replies to select during conversations. The first movie Ben picked alone was Spider-Man. The tools are free and open source.
@METANA_flow [Claude Code]
https://x.com/METANA_flow/status/2108505432461291696
A Japanese web agency had its own English newsletter form hit by bots: 65 sign-ups in about 80 minutes one night, all from overseas data centers, and because sign-up triggered an auto-reply the company's domain sent about 60 emails to strangers' Gmail accounts, which risks the whole domain being flagged as spam. Fixing it took the AI and 30 minutes. The site's form intake and mail delivery run on a custom system rebuilt with Claude Code, so there was no plugin update to wait for: the agent read the code, rewrote it and shipped to production that night. Now sign-ups that fail the bot check are stored without a reply, bot-triggered sends dropped to zero and fake traffic was filtered out of GA4 where it had reached over 70 percent on some days. The author's advice to companies that cannot touch their own site with AI is to leave WordPress for a managed SaaS.
@denk_tweets [Claude Code]
https://x.com/denk_tweets/status/2108537564680318984
The updated AI stack of a founder running a $35M ARR startup with 130 employees. Claude is open on the desktop at all times with MCP connectors into docs, Slack, code, database, finances, inbox and calendar; Fellow joins every meeting; Unblocked lives in Slack with a full view of the business; Linear tickets are created from Slack and analysis runs through Claude; Claude skills cross-reference more than 100 dashboards with the database and Slack to surface risks; scheduled tasks nudge employees for weekly priorities; and Claude Code plus Figma Slides produces the decks. Wispr Flow handles voice input and Muse is being tested for candidate sourcing and travel. The author asks what is missing.
@itwanger [Claude Code]
https://x.com/itwanger/status/2108438374747099540
A Chinese creator's Feishu knowledge base on agents just passed 1,000 visitors two months after launch, covering agent basics, context and memory, harnesses, RAG, Claude Code, Codex, MCP and model fine-tuning. The production pipeline: scripts written with Opus 5.5 (an upgrade from 4.6, which the author says was not good enough at text), video production that started on Volcano Ark's coding plan, moved to Astra and is now mostly Opus 5.5, and illustrations from GPT-6.1 Sol. Videos go to Douyin, WeChat Channels, Xiaohongshu and Bilibili, with Douyin giving AI video the most traffic and revenue. The knowledge base is free and open.
@cyrilXBT [Claude Code]
https://x.com/cyrilXBT/status/2108448665568498153
A five-step recipe for a wiki that writes itself. Create an Obsidian vault, open it in Claude Code, have Claude make two folders (sources for anything you drop in, wiki for pages Claude writes and links), ask it to write a CLAUDE.md with the rules, then drop in an article, PDF or transcript and type add this to the wiki. Claude reads the source, writes the pages and links them to everything already in the vault, and afterwards you can ask questions across the whole thing.
@_nogu66 [Claude Code]
https://x.com/_nogu66/status/2108367540808466550
The monthly Claude API credits now bundled with the Max plan are being spent on a personal experiment: an AI-generated homepage, regenerated daily from the author's own sites, that replaces the Chrome new-tab page at about $0.75 per run. The author's framing is that treating leftover credits as fuel for a personal agent makes more sense than letting them lapse, and the next step is finding more uses that run outside Claude Code sessions.
@cinkotweets [Claude Code]
https://x.com/cinkotweets/status/2108655933802807694
Twenty-four hours with HyperFrames Studio, HeyGen's new free editor built on the HyperFrames engine that until now was mostly driven from Claude Code. Without opening Premiere or Final Cut the author produced a fully edited 25-minute podcast with intro, captions and archival footage, turned a spreadsheet into a video, and made a pile of launch videos, by drawing on the screen, typing in a text box and telling the AI what to change. A timeline, a chat box and a pen replace writing out timestamps, and the author says there is no going back. Honest notes: a one-sentence prompt yields slop, it is a day-one app with ghost frames and clunky navigation, updates ship several times a day, and you need a Claude, Codex or Grok subscription plus a Mac or Linux machine.
@HBCoop_ [Claude Code]
https://x.com/HBCoop_/status/2108587885763211346
A workflow for short video stories in which the author describes the idea to Claude Code and it works directly inside Invideo's editor while the author directs. The piece Open Sky started as one character image and a five-shot outline; Claude generated cheap 480p drafts first, checked each shot frame by frame and only finalized the approved ones, which avoided paying for 1080p on throwaway shots. The cut, shot-to-shot color match, film look and the vertical Reels version all happened in the free editor, so iteration cost nothing beyond generations. It was not hands-off: the first color grade was invisible, the stronger pass turned the character's hair purple and the grain was too busy, but each fix was one sentence and a fresh look at the frames rather than an hour in panels.
@cybermynd [Claude Code]
https://x.com/cybermynd/status/2108616877417377876
A five-minute short film made with Claude Code and the Krea MCP. Claude wrote the script, prompted every shot through Krea, cut the footage, scored it and mixed the sound. The film is titled System Memory, about an AI remembering a home it never had.
@insomnia_vip [Claude Code]
https://x.com/insomnia_vip/status/2108619162746585582
Claude rendered 960 frames from a single file inside one Claude Code session, then the terminal checked whether the animation loop worked. Six checks passed but the first and last frames did not line up, leaving a visible jump; Claude changed one number, re-rendered and fixed it on the first try. The whole workflow is automatic: frames rendered from code, six checks, errors caught and fixed, re-render, stop when every check passes. No After Effects, no video model, just Claude, a terminal and a system that can check its own work.
@whaleyxbt [Claude Code]
https://x.com/whaleyxbt/status/2108469154986557886
Nine days before Anthropic launched Claude Motion, the author open-sourced a motion tool for Claude Code that does the same job: Claude writes the whole video as code, checks its own frames and fixes what is off, and makes its own sound with no stock audio. The author's 441K-view crawler video was made with it. It hit number one on a tracking list the day before, is free and runs inside Claude Code.
@orenmeetsworld [Claude Code]
https://x.com/orenmeetsworld/status/2108643262332612912
Two tips for anyone editing a lot of video in Codex or Claude Code. First, ask the agent to build an interface where you add review notes by timestamp, where new videos are added automatically and approved ones drop off. Second, before you ever look, have Sol 6.1 or Fable review the first edits and send feedback that gets incorporated, so your pass is on an already-revised cut.
@1000man_niki [Claude Code]
https://x.com/1000man_niki/status/2108518839768236401
A paid, self-funded review of a 1,980-yen skill that cuts long videos into vertical shorts, run inside Claude Code or Codex. From a 24-minute 4K interview the author produced three roughly 60-second vertical clips in about 96 minutes of wall time, of which under five minutes was hands-on; the rest was waiting on the AI. Setup took eight minutes with no errors. The first export needed fixes: captions appeared too early and both speakers had the same subtitle color; a second pass took 34 minutes, and a third with added sound effects and music took four and a half. Highlight selection was middling, but the automatic reframing to whoever is speaking worked well; two typos and some awkward caption breaks mean the first output was not corporate-ready, though it is fine for reviving personal footage. Per-video settings for subtitle timing and speakers are saved and color, volume and fade preferences can be stored as shared defaults, but corrections are not learned automatically.
@gkxspace [Claude Code]
https://x.com/gkxspace/status/2108545947231784990
The viral code-made animation videos are reproducible with StepFun's Step 5 Preview at roughly one eighth of Opus 5.5's price, according to a Chinese user who switched Claude Code's model via CC Switch and ran a batch of demos with an open-source skill. Results included a 3D miniature town cycling through 24 hours in 20 seconds, Chinese kinetic typography that falls, scatters and settles into vertical layout with a seal stamp, a single-shape UI morphing from button to player to chart, and a green-screen dancer composited into a magazine that flips pages with the moves. The recipe: pull good examples from the awesome-opus5-5-videos collection, have the AI rewrite the prompt with duration, beat, per-shot contents, exclusions and acceptance criteria, then switch the model and let it run; one piece can take hours and 200-plus tool calls to write code, compose music, render and check. The author's point is that code-made video is commercially viable for brand motion and product demos, but Claude's cost was blocking batch work.
@Saccc_c [Claude Code]
https://x.com/Saccc_c/status/2108546155634434194
A rotating, interactive 3D island built from a reference image by Step 5 Preview plugged into Claude Code, which the author says is on par with Opus 5.5 and cheaper. The whole site renders live in Three.js with water, rocks and scenery generated in code. The process was simple: prepare a generated reference image, describe the page and interactions clearly, then have the model check against the reference and iterate until the look matched, which takes some patience. Step 5 Preview is on OpenRouter with 1M context and is also wired into Kilo Code, Cline, Hermes Agent and OpenCode.
@gengdaJ [Claude Code]
https://x.com/gengdaJ/status/2108549740971311226
After watching kids play Carrot Fantasy over a trip home, the author decided to build the game instead, and used the occasion to compare two Chinese models inside Claude Code via CC Switch: Step 5 Preview and DeepSeek 4.1 Flash. To isolate model ability, the prompt forbade any skills, memory or reference files and asked for a complete game in one shot, with only web search allowed. Both produced full games with levels, progression unlocks, multiple tower types and polished UI; the author judged completeness a tie and gave Step 5 the edge on playability and visual finish. The post ends on a reflection that playing games has turned into thinking about how to make, monetize and ship them.
@yhslgg [Claude Code]
https://x.com/yhslgg/status/2108442753587728816
A breakdown of how one person shipped a browser-playable GTA-grandma game in a single holiday, repo open-sourced. Assets: Tripo for concept art and one-click 3D conversion, Smart Mesh for quad topology, texturing, Mixamo rigging, and text-to-motion for animations such as the grandpa playing guitar from a single line of text. Code: about 8,000 lines, all written by Claude Code in conversation, rendered in Three.js, with Playwright driving every level from start to finish as a test. The part the reviewer values most is the finishing: text-to-3D models arrive at around two million triangles, so headless Blender batch-decimated them and compressed textures to fit in a browser. The lesson stated is that generation was never the hard part; making it run is.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2108645144794702099
One phone photo became a 3D model with moving parts, built by Claude Code purely from code and checked render-by-render against the photo until they matched. Four desk objects were tested: an earbuds case whose lid opens on its hinge, a pocket knife whose blade folds out, a desk lamp whose arm bends at both joints, and a plush dog. The pattern is that hard edges come out exact and fur comes out stylized, the back is guessed, and the agent says which sides it guessed; three of four now spin on the author's test page. The shared prompt asks the agent to list every recognizable detail before coding, build in passes taking colors from photo pixels rather than memory, give each moving piece its own pivot and idle animation, and compare each pass against the photo before moving on.
@sandislonjsak [Claude Code]
https://x.com/sandislonjsak/status/2108521597669630102
First look at Based Garden, a garden planner that takes your balcony, terrace or backyard and computes how many hours of direct sun each spot gets across the season, accounting for the roof above, walls, fence and the tree in the corner, then says what will grow there. It began with the author's own 9-by-1.7-meter terrace in Zagreb, where the glass railing gets 7.3 hours a day and the corners behind solid walls get 0.8, a difference you cannot see by looking. It already has a 3D view for any day of the year, plants that grow and fade with the season, per-spot crop advice and four example spaces. It is built with Claude Code and the Incantations skills: one mode worked out the audience and pricing before any code, a design mode produced three working layouts to choose from, and an engineering mode built the engine and landing page, with each decision written into the repo instead of lost in a chat.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2108525415971352923
A Japanese user reports that Claude Code can now produce a seminar deck of this quality in one shot, which has made the job dramatically easier. Decks of more than 100 pages are no problem, they come with animations and can be edited like PowerPoint.
@dotey [Claude Code]
https://x.com/dotey/status/2108574361443504625
A token-saving trick for prototyping with Claude Design: build a complete version of the design on the Claude site, iterate until satisfied, export it as HTML and download it, then have Claude Code treat it as a local web page for all further updates, with no further dependence on the Claude Design site. The author shares a repo of prototypes made this way that runs with two npm commands.
@dotey [Claude Code]
https://x.com/dotey/status/2108669008283402744
Advice to cap Claude Code's auto-compaction context length: around 400K is reasonable and the author uses 300K, set with the autocompact slash command. The author had previously recommended disabling the 1M context through a settings.json variable; that is no longer necessary and should be removed, otherwise the session tops out at 200K.
@Jhaddix [Claude Code]
https://x.com/Jhaddix/status/2108637514550657475
A practical finding about skills in agent harnesses. Most harnesses, Claude Code included, load a skill's description at startup, but even when a skill is explicitly invoked the agent usually reads only the first 150 lines to decide how to use it. Any mandatory instruction, Python script or template referenced after that point can be missed entirely. The author's fix is to put a table of contents of the skill's actions in the first 150 lines, mark which are mandatory, and state that if the skill was invoked the whole skill and all references must be read; the author's team now uses markdown checklists for this.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2108450345341079874
For people who ask Claude Code to just check something and get their code edited anyway: putting Read in the allowedTools flag does not mean only Read is permitted. That flag lists tools that may run without confirmation; restricting which standard tools exist at all is the separate tools flag, and MCP tools are specified separately again. Splitting these lets you request a review without handing over the editing tools; the launch command and settings are in the replies.
@sebuzdugan [Claude Code]
https://x.com/sebuzdugan/status/2108431867188052210
Day 23 of a hundred-day prompting series: every Claude Code task prompt was wrapped in XML tags, same words, one tag per paragraph, and compared against the untagged version. The result was 8 of 12 tasks passing on both sides, with time and cost within three percent. No tagged run ever mentioned a tag, and both sides broke the same parameter and failed the same way.
@morganlinton [Claude Code]
https://x.com/morganlinton/status/2108355804537786840
Results from an independent benchmark of Sonnet 5.5 on the author's Frontier v4 eval suite, with all Claude models run inside Claude Code and OpenAI models inside Codex. Both Opus 5.5 and Sonnet 5.5 score higher than GPT-6.1 Sol, but not by much, and Sol is far more cost effective. Opus 5.5 is clearly ahead at medium and high effort, which justifies giving it the hardest tasks, but for easy and medium work the author sees little reason not to use Sol given its very low cost per task. The author suspects Codex-side token-efficiency work may be part of the gap, and all data and runs are open.
@TokenGremlin [Claude Code]
https://x.com/TokenGremlin/status/2108568524197179500
Combining personal measurements, third-party tests and the official speed multipliers from OpenAI and Anthropic, the author compares subscription speeds rather than API ones. GPT-6.1 Sol used to generate about 22 tokens per second, OpenAI raised that by 50 percent and then announced Ultrafast, which now measures around 249 tokens per second. Opus 5.5 delivers roughly 97 tokens per second directly in Claude Code, and with fast mode's 2.5x boost about 242.5. The conclusion is that you pay about $500 a month to get Sol to the speed Opus 5.5 Fast reaches on a $100 Anthropic subscription.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2108513656686223833
A used $1,200 AMD box under the desk paid for itself in three months by replacing three subscriptions. Month one, ChatGPT Pro at $200 went when gpt-oss-120b ran at about 30 tokens per second on a Ryzen AI Halo. Month two, Claude Max 20x at $200 went when Claude Code was pointed at Qwen3 Coder 30B running locally, which works exactly as before with the model under the desk. Month three, SuperGrok Heavy at $300 went when GLM 4.5 Air took over long research sessions. The arithmetic: $1,300 saved by month three, about $7,600 by month twelve, roughly $10 a month in electricity; Opus 5.5 still writes the hardest code, and the monthly AI bill went from about $700 to the power cost.
@tvytlx [Claude Code]
https://x.com/tvytlx/status/2108638124964532517
The cheapest model has started assigning work to the most expensive one. A Reddit user built a Claude Code plugin called effortless after repeatedly forgetting to switch models and running everything at the top tier. Before each prompt goes out, Haiku 5.5 reads the recent conversation and decides whether the job belongs to Haiku, Sonnet or Opus and at what effort, routing simple work to the cheap model while the conversation's own model stays put. It knows when not to switch: mid-task it holds, because it can read that you are still inside the same job. Opus and Sonnet effort changes do not rebuild the cache, so those are safe; on Fable an effort change rebuilds the cache and costs more than it saves, so auto-switching is disabled there. It also warns when the cache has gone cold or the conversation is bloated and it is time for a fresh window. The author says the estimated effort matches the manual choice almost every time; free, MIT licensed, mostly written by Claude.
@rishit30g [Claude Code]
https://x.com/rishit30g/status/2108484818979955000
Introducing usage-band, a Claude Code mod that puts usage right above the prompt: the current model, a context-window bar that goes green, amber then red, the five-hour plan limit with reset time and a warning if your pace would hit it first, a prompt-cache timer and hit percentage read from real API data, and session cost at API prices. A Summarize and new button has Haiku write a hand-off summary, clear the session and drop it in the prompt box. The author contrasts it with a status line: a status line is a script that prints text, while a mod runs inside Claude Code, can have a clickable button, ticks the cache countdown every second and reads usage, limits and cost from the plugin hooks API with no hardcoded price table. Works in the terminal and the desktop app's Code tab, open source, built with Opus 5.5.
@byCanen [Claude Code]
https://x.com/byCanen/status/2108552591722389781
When Claude Code inspects web pages it often reads whole screenshots, and a 2,000-pixel image can cost thousands of image tokens, so many screenshots drain both quota and context quickly. The free image-diet mod shrinks the long edge to 1,280 pixels before Claude's Read of an image and before some MCP image results reach the model. In the author's small screenshot test, image tokens dropped by about 59 percent with no worse answers, a figure that will vary by screenshot. It deliberately skips computer-use tools that click on screenshot coordinates, since shrinking would misplace them, and the full-size version can be kept for detail work.
@criscxuan [Claude Code]
https://x.com/criscxuan/status/2108533757070250007
A close read of Magpie, a tool whose value is letting you choose the agent and the model independently, for people juggling several coding tools, providers and subscriptions. You can keep Claude Code's workflow but route it to Kimi, have Codex call DeepSeek, and fail over when one account's quota runs out, all behind one local gateway with 45-plus agents listed and CLI, TUI, web and Docker front ends. The strongest parts of the code are protocol conversion through a unified internal request structure that bridges OpenAI Chat, Responses, Anthropic Messages and Gemini, and the subscription bridge: for a Claude subscription Magpie drives the real local claude binary, exposes the original agent's tools as MCP tools, hands tool execution back to that agent and returns results to the same Claude Code process while managing process reuse and session resume to preserve context and cache. Routing honors subscription quirks: sequential, rotation, least-used-first or soonest-reset-first, with session-to-account stickiness, and an optional intent router lets a small model classify a new turn before choosing a model. MIT licensed.
@AlexZio00 [Claude Code]
https://x.com/AlexZio00/status/2108435105769459984
Release notes for sovereign-skills 6.5.14, twenty open-source Claude Code skills, with no new skills and a long list of fixes found by daily use. Eleven test files had been passing silently because they used their own runners with no test functions, so pytest collected zero tests and exited clean; all eleven now actually run. The session-start skill stopped an entire run when no handoff file existed on a first session and now skips just that step; pre-push treated a missing secret scanner as skip the check and now stops the push; a read-only audit skill could still write files through shell redirection because Bash was not blocked; setup replaced config files without a backup and now makes a .bak and asks first. A further list covers rules that contradicted themselves across skills, such as a health score that let a never-used skill look healthy and a quick mode that printed a ship verdict it was not meant to give.
@slash1sol [Claude Code]
https://x.com/slash1sol/status/2108497556770554041
Four researchers read the source of eleven coding agents, Claude Code, Codex CLI, Gemini CLI and eight more, and the author compresses the category into one line: agent equals model plus harness. Seven subsystems, every agent takes a position on each; zero of eleven import an agent framework; zero of eleven retrieve code with embeddings. The author then maps the same seven rows onto a non-coding agent of their own, which trades Polymarket BTC and ETH up-down markets: loop, tools and safety caps took a weekend, while the context row took a month because the model needs the historical level-2 order book for every market it touches and no model has that in its head. That row is now served by a dedicated historical data feed, every tick, every market, replayable.
@Mikadzyki_NFT [Claude Code]
https://x.com/Mikadzyki_NFT/status/2108490740611399973
A Stanford paper reports that Claude Code agents can run about twice as fast once the overseer is removed. The usual setup has a coordinator handing out tasks and collecting results, sitting idle while helpers work and keeping them from seeing each other's findings. DeLM is a layer on top of Codex and Claude Code with no central hub: agents share a common context and a task queue, grab open work on their own, publish findings immediately and see which approaches already failed. In one example the main model spent 73 percent of its time waiting with native subagents and the job took 162 minutes, versus 29 with DeLM; at best the gains over standard Claude Code were 2.49x faster, 19.2 more points of solved tasks and 19.9 more points on ProgramBench within 120 minutes. The sample is small, speed does not always mean savings, and a four-agent run costs more than baseline; the code, 720 trajectories and a plugin for both tools are open.
@theshawwn [Claude Code]
https://x.com/theshawwn/status/2108633474273800662
The author asked Claude Code to make a transcription of the session's chat and was denied, asked where the .json files were on disk and was told it could not say, then asked Claude Desktop, which answered. After adding a permissions rule and asking Claude Code to copy the files, it still said no. The author has no idea what the plan is here and thinks it has gone a little too far.
@wey_gu [Claude Code]
https://x.com/wey_gu/status/2108493386739228818
Over the National Day holiday the author's Claude Code account was banned, despite a long-held belief that keeping a clean exit IP would be enough to stay safe. A friend in North America in the same batch was banned at the same time.
@superdoccimo [Claude Code]
https://x.com/superdoccimo/status/2108534990216761493
A Japanese write-up of bigarrow, which hit the top of Hacker News within an hour: a macOS CLI and agent skill that draws a large arrow, circle, box or text sign on top of everything so an AI agent can point at where the human needs to click. The most frustrating moment with agents is being told to click something yourself, for macOS permission dialogs, OAuth consent, 2FA, passkeys and payments, and then hunting for which window and which button. Clicks pass through, focus is not stolen, drawing needs no permission, and the tool itself never clicks, types or captures the screen; it only points, by coordinates, rectangle, mouse position, window title or UI element label. It installs into Claude Code and Codex with one command. At the time of writing it had 47 stars, 76 commits, nine releases and was Mac only.
@secondstateinc [Claude Code]
https://x.com/secondstateinc/status/2108578028439928886
VibeBuddy is now open source: an ESP32-S3 desk box that watches Claude Code, Codex and GitHub Actions and speaks up when an agent needs you. It ships with Rust firmware, a Mac and Linux app, hooks and an open protocol, so a single curl puts a card on its screen.
@thecsguy [Claude Code]
https://x.com/thecsguy/status/2108412748904427799
Grok Bot now drives the Claude Code sessions on the author's Mac, from a phone. The day's work continued a mobile video-editing app, tested it on the author's own Mac through the iOS simulator and prepared it for App Store review.
@billy_aiwork [Claude Code]
https://x.com/billy_aiwork/status/2108376910388441593
A Japanese user has Claude Code write scripts and Codex generate the images, and only just discovered Codex CLI: with it, Claude Code sends the instructions to Codex itself, so one tap goes from image generation to saving the result where it can be viewed on a phone. The verdict is that the command tower is Claude Code.
@lxfater [Claude Code]
https://x.com/lxfater/status/2108369386079662191
Tired of AI-drawn diagrams that come out as plain rounded boxes needing a Figma cleanup, the author of diagram-design wrote the color, font and layout rules into an open-source skill. It covers 42 diagram types including architecture, flow, timeline and Gantt, works in Claude Code and Codex, can read your website to match its palette and fonts so diagrams fit your articles, and can redraw existing Mermaid, Excalidraw and similar diagrams. Output is HTML plus SVG that opens in a browser and can be further edited by the AI, with light, dark and magazine styles.
🗣 User Voice
User Voice
Weekly and five-hour limits are the loudest complaint again. @pcshipp asked how to burn a weekly limit in 11 hours, @AlchainHust hit a quarter of a weekly quota within hours of topping up, and @mylifcc and @HouseHackerJon were both reaching for Haiku to stretch the same allowance. The new Max API credits are being treated as agent fuel by @_nogu66 and as a fallback pool by @yancya, and @aliceisplaying mapped out how confusing it is which credits work where.
Model routing advice has become a template, and users want the truth behind it. The Sonnet-builds, Haiku-swarms, Opus-advises diagram was posted as an official Anthropic setup by @Arindam_1729, @Bober_smart, @merccante, @claudecode84 and others; @Lniosytest checked the docs and found the parts real but no bundled plan and no source for the 340 tokens per second figure, and @dravenip relayed the Claude Code creator saying the cheap model is the expensive choice. @fleyta88 and @morganlinton want cost per passed task measured instead of token price.
Who can read the agent's files is now a daily worry. @catnose99 pointed out that Claude Code and Codex session logs are enormous, plaintext and stored in a known place, @keitaro_aigc told everyone to check public servers for stray .claude folders and CLAUDE.md files, @deeponailabs reminded that mods run with your permissions and can read keys, and @theshawwn hit the opposite wall when Claude Code refused to hand over its own session transcript.
Coordinators are wanted but hobbled. @astralwave listed how the Claude Projects coordinator has no browser or MCP access and project threads cannot invoke skills by hand, @onenewbite rebuilt Projects locally because the main thread acts as a messenger rather than a brain, @bart_walker is coordinating Hermes, Codex and OpenClaw through a shared folder, and @arnestrickmann and @trevin each built the missing connection and document layers themselves.
OpenClaw users are judging it on upgrade survival, not features. @tvytlx read a 112-PR patch that was mostly rollback and recovery and decided to wait for the next one, @pan_derevyan said zero of six months of updates completed without manual intervention, @shields_pikes moved to a managed alternative because it broke too often, @_bashoh reported cron prompts leaking into Slack, and @EdgeDimi filed a traced WebSocket fix with the one-line patch attached.
Weekly and five-hour limits are the loudest complaint again. @pcshipp asked how to burn a weekly limit in 11 hours, @AlchainHust hit a quarter of a weekly quota within hours of topping up, and @mylifcc and @HouseHackerJon were both reaching for Haiku to stretch the same allowance. The new Max API credits are being treated as agent fuel by @_nogu66 and as a fallback pool by @yancya, and @aliceisplaying mapped out how confusing it is which credits work where.
Model routing advice has become a template, and users want the truth behind it. The Sonnet-builds, Haiku-swarms, Opus-advises diagram was posted as an official Anthropic setup by @Arindam_1729, @Bober_smart, @merccante, @claudecode84 and others; @Lniosytest checked the docs and found the parts real but no bundled plan and no source for the 340 tokens per second figure, and @dravenip relayed the Claude Code creator saying the cheap model is the expensive choice. @fleyta88 and @morganlinton want cost per passed task measured instead of token price.
Who can read the agent's files is now a daily worry. @catnose99 pointed out that Claude Code and Codex session logs are enormous, plaintext and stored in a known place, @keitaro_aigc told everyone to check public servers for stray .claude folders and CLAUDE.md files, @deeponailabs reminded that mods run with your permissions and can read keys, and @theshawwn hit the opposite wall when Claude Code refused to hand over its own session transcript.
Coordinators are wanted but hobbled. @astralwave listed how the Claude Projects coordinator has no browser or MCP access and project threads cannot invoke skills by hand, @onenewbite rebuilt Projects locally because the main thread acts as a messenger rather than a brain, @bart_walker is coordinating Hermes, Codex and OpenClaw through a shared folder, and @arnestrickmann and @trevin each built the missing connection and document layers themselves.
OpenClaw users are judging it on upgrade survival, not features. @tvytlx read a 112-PR patch that was mostly rollback and recovery and decided to wait for the next one, @pan_derevyan said zero of six months of updates completed without manual intervention, @shields_pikes moved to a managed alternative because it broke too often, @_bashoh reported cron prompts leaking into Slack, and @EdgeDimi filed a traced WebSocket fix with the one-line patch attached.
📡 Eco Products Radar
Eco Products Radar
Codex (128 mentions), OpenClaw (113), MCP (52), Opus 5.5 (46), Cursor (41), Claude Code Projects (35), Grok Bot (34), Hermes (28), DeepSeek (25), Gemini (22), Dots (19), Pine Computer (18), Haiku 5.5 (18), Muse (14), Jev (14), OpenCode (14), Sonnet 5.5 (13), Fable (12), GitHub Copilot (12), Slack (10), Gemini CLI (10), Pi (10), GLM (10), Claude Cowork (10), Cloudflare (9), OpenRouter (9), Telegram (9), Instinct (8), Qwen (8), Claude Managed Agents (8), Claude Dashboards (8), Kimi (6), Excel (6), VS Code (6), Step 5 Preview (5), HyperFrames (5), Antigravity (5), Ollama (5), Discord (5), REA (4), Blender (4), Claude Motion (4), Notion (4), Playwright (4), superpowers (4), Remotion (3), Firecrawl (3), Three.js (3), Drex (3), Unsloth (3), llama.cpp (3), Cline (3), Aider (3), Droid (3).
Codex (128 mentions), OpenClaw (113), MCP (52), Opus 5.5 (46), Cursor (41), Claude Code Projects (35), Grok Bot (34), Hermes (28), DeepSeek (25), Gemini (22), Dots (19), Pine Computer (18), Haiku 5.5 (18), Muse (14), Jev (14), OpenCode (14), Sonnet 5.5 (13), Fable (12), GitHub Copilot (12), Slack (10), Gemini CLI (10), Pi (10), GLM (10), Claude Cowork (10), Cloudflare (9), OpenRouter (9), Telegram (9), Instinct (8), Qwen (8), Claude Managed Agents (8), Claude Dashboards (8), Kimi (6), Excel (6), VS Code (6), Step 5 Preview (5), HyperFrames (5), Antigravity (5), Ollama (5), Discord (5), REA (4), Blender (4), Claude Motion (4), Notion (4), Playwright (4), superpowers (4), Remotion (3), Firecrawl (3), Three.js (3), Drex (3), Unsloth (3), llama.cpp (3), Cline (3), Aider (3), Droid (3).
Comments