Super User Daily: 2026-09-27
Opus 5.5 turned Claude Code into a studio this week. The loudest cluster is video made entirely of code: promo clips, a 3D folk-tale film, a comic ad with six cast voices, a UMAP lecture from a phone, all rendered without a single video model, often for a few dollars. Behind it, the cheaper token prices changed behavior more than capability did: people now run dynamic workflows for hours, route work to cheap worker models, and measure quota per dollar like a spec sheet. The best non-coding cases are the quietest ones, an accounting firm wiring every client's books to Claude Code, a health command center that adjusts the bedroom air overnight, and sub-dollar market intelligence. Microsoft shipping Autopilot on OpenClaw gave the other side of the ecosystem its biggest day, while bans, silent model routing and subagents that answer on your behalf remain the sharp edges.
@AlanShiflett [OpenClaw]
https://x.com/AlanShiflett/status/2103524300158366183
Lost about 60 pounds, then got tired of health data living in six apps that each only saw a slice. So they built a personal command center: Epic MyChart records, Fitbit/Google Health, an AirGradient bedroom sensor, their AC, their Govee lights and a home-made nutrition tracker, all exposed through one personal MCP. OpenClaw watches bedroom CO2 overnight and tells the AC to pull in outside air, and ramps the morning lights based on bedtime, yesterday's training and their first meeting. Claude pulled ten years of uric acid results plus diet logs to pin a gout flare on sugar-free popsicles, and dug up a 2017 specialist note nobody ever followed up on. The line that matters: they own the data and the connectors, so they never wait for a vendor to sign a deal with an AI lab.
@lklkfafa1 [Claude Code]
https://x.com/lklkfafa1/status/2103457111619592324
An accounting practice has Claude Code wired to Money Forward's accounting API for every client company, not just a demo on one. When Money Forward shipped API-key auth this week, the author tested it the same day and confirmed one key now reaches the whole client roster, each company addressed by office_code so a request can't land in the neighbor's books. Existing OAuth setups keep working; new setups skip app registration and per-client approval entirely. The workflow stays the same: a human places the key in a file by hand, and Claude Code writes the tooling around it. This is what bookkeeping automation looks like when the practitioner, not the vendor, owns the integration.
@jlhortelano [Claude Code]
https://x.com/jlhortelano/status/2103398379426378016
A sports-sponsorship analytics company ran market intelligence through Claude Code plus treg (one key for about 3,700 endpoints across 95 data vendors) and Jev as a cheap judge. Three probes: how seven AI search engines answer questions about their category, a competitor radar built from headcount, hiring, funding and posting, and which brands actually use sponsorships in paid ads. Total bill: $0.97, and it surfaced a competitor showing up in half the AI answers plus queries describing their exact product where they appear nowhere. The honest part is the most useful: an improvising agent picks different vendors every run, so month-over-month numbers don't compare, and they froze it into four fixed reports expected to cost $10 to $15 a month. The takeaway they land on is right: data got cheap, judgment is now the scarce input.
@KyleCranmer [Claude Code]
https://x.com/KyleCranmer/status/2103581526285857085
A physicist who has worked on a hard scattering-amplitude problem for a couple of years says the last few months ran on agentic tools like Claude Code, and the solution just landed. The twist: their collaborator had been confirming it under NDA, and a friend inside Anthropic was working on it too. Same week, Anthropic announced Claude solving a nine-loop calculation. This is the frontier-science version of the usual story, where the tool is a long-running collaborator, not a code autocomplete.
@dadchords [Claude Code]
https://x.com/dadchords/status/2103616493569097857
They got a virtual Unix environment running inside a JavaScript VM in a browser tab, well enough that Claude Code runs inside the tab and edits the very page it is running in. On top of that they are prototyping an Endlesss clone: a Strudel-inspired music environment plus a realtime audio loop capture tool, and posted their first jam inside it. They note how much performance and bug work it took to get here. A coding agent living inside the artifact it is building is a genuinely new shape of tool.
@NFT_Chen [Claude Code]
https://x.com/NFT_Chen/status/2103401780117938683
The opus-manager skill turns Opus 5.5 into a manager instead of a coder: it writes tickets, dispatches them to cheap worker models like DeepSeek or GLM, reruns acceptance itself, gets a cross-vendor read-only review, then checks each review claim for truth. Reported numbers: a colleague's full day of delegated work used only 20% of a Pro weekly quota, one work project made 3,617 DeepSeek calls for about a dozen dollars, and the author cleared 60+ tickets in one go. Everything, tickets, receipts and review reports, lands as Markdown in the repo. The warning is the practical bit: don't interrupt mid-run, because every interruption makes Claude stop and re-verify, burning quota.
@aipulseda1ly [Claude Code]
https://x.com/aipulseda1ly/status/2103352533657719230
A full survival game in the vein of The Forest, in three.js, in one day of continuous work with Opus 5.5. A 2 km island carved by simulated erosion, 44,000 procedurally generated trees wrapped in photoscanned bark, the sun and moon placed where they'd really be at 49 degrees north in late September, thunder delayed by the speed of sound, wolves that circle the fire and back off from a torch. Claude Code split the world across waves of agents, one per domain, while critic agents that never touched the build compared renders against real photos. They ran out of usage before the realism pass and say the wolf is still the weakest thing in it.
@vikingmute [OpenClaw]
https://x.com/vikingmute/status/2103379440390844449
OpenClaw's maintainers deleted roughly 400,000 lines of their own tests and code coverage barely moved, because current models love writing pointless unit tests that assert constant strings. The trick that made the cleanup work: telling an agent to just clean up makes it stop early, so they gave it a harsher target, delete the least useful 20% of tests while keeping coverage within 2%. They also published the test-audit skill that now gates new tests: the agent must answer whether a test protects real behavior before it is allowed to add one. Test bloat from agents is a real cost nobody budgets for.
@peesamac [Claude Code]
https://x.com/peesamac/status/2103424988472893946
video-use from the Browser Use team is a talk-to-edit video skill: drop raw clips in a folder, tell Claude Code what you want, get final.mp4 back with no editing UI at all. It cuts filler words, retakes and dead air, color grades, puts 30ms audio fades on every cut, burns in subtitles and farms animations out to parallel sub-agents via HyperFrames, Remotion or Manim. The clever part is how it reads video: ElevenLabs Scribe gives word-level timestamps, speakers and even laughter, all takes compress into a roughly 12KB text file, and it only opens waveforms or filmstrips at decision points, versus an estimated 45 million tokens to look at every frame. It proposes a plan first, then self-checks every cut and re-renders up to three times.
@Majin_AppSheet [Claude Code]
https://x.com/Majin_AppSheet/status/2103497319899693327
Asked Claude Code on Opus 5.5 for a promo video in their own style and basically replied 'up to you' to its questions. No video-generation model involved: frames drawn one by one in HTML and code via HyperFrames, BGM generated with Gemini's Lyria, and sound effects timed to the beat. The part that floored them: Claude can't hear, so it analyzed the waveform numerically to land effects exactly on landings and stamp hits. Code-rendered video is quietly becoming its own genre this week.
@ladprofit [Claude Code]
https://x.com/ladprofit/status/2103479370883137664
A 60-second comic-book ad with a full cast, directed entirely inside Claude Code. It wrote a 21-panel PAS script around a named mechanism, built character sheets first and fed them to Nano Banana Pro so faces stay consistent across panels, cast six ElevenLabs voices and pitched one up for the eight-year-old, and located each character's mouth on a grid so speech-bubble tails point to the right mouth. Then it edited itself: page turns, slams, screen shake, SFX and music scored act by act. Cost was about 65 image credits plus a few cents of voice.
@wshuyi [Claude Code]
https://x.com/wshuyi/status/2103308292613345504
An end-to-end test of whether a model can just shoot a story: a 3D explainer film of the Chinese folk tale The Taoist of Laoshan, made from scratch in Claude Code on Opus 5.5. No storyboard script, no off-the-shelf film skill, only their own cloned voice as input. The point of the experiment was to see if the current model can carry narrative, visuals and pacing by itself, and the result was good enough to post.
@goodside [Claude Code]
https://x.com/goodside/status/2103505085292593220
Riley Goodside had Opus 5.5 write and animate a five-minute explanation of UMAP for an undergraduate ML audience. The visual concepts, the narration, the structure, all from Claude, prompted from Claude Code on a phone. A teaching asset that would take a TA a week now comes from a phone prompt.
@mmmiyama_D [Claude Code]
https://x.com/mmmiyama_D/status/2103319199817064764
Wanted a 30-second PR video but refused to pay for yet another AI subscription, so the creator made it with Opus 5.5 in Claude Code alone. No image or video generation at all, just code, which means edits are easy and usage barely moved. The flat-design look came out usable for real marketing. Earlier the same day a one-shot 15-second self-intro video even downloaded its own fonts and made its own music.
@vikktorrrre [Claude Code]
https://x.com/vikktorrrre/status/2103525020530348430
Receipts for the motion-design hype: Claude Code harness, max reasoning, one straightforward prompt, 1 hour 32 minutes, $81 at API prices. The output is a polished motion-design piece the creator calls a masterpiece. Useful as a price anchor: a commissioned motion piece at an agency runs into the thousands, and this is now a line item under $100.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2103616788324130984
Someone copied a viral Claude Code prompt almost word for word, gave it an OpenRouter key capped at $10, walked away, and came back 1.5 to 2 hours later to a finished animated explainer for their side project that cost about $4. Claude wrote the script and concept, generated collage art, voice-over, music and sound effects, and rendered a pure JavaScript canvas animation to MP4 with every beat synced to the words. It also used a second model to review its own drafts and fixed what it found. A hard budget cap is the quiet hero here.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2103412945552326983
Opus 5.5 built a CRM from a single prompt by delegating to six subagents. The main agent wrote the product brief, fictional data and design system; one subagent built the foundation and five finished the rest in two parallel waves. Along the way it ran 57 Mobbin MCP searches for real app patterns, applied 11 design skills, and ran screenshot review loops on every page in dark and light mode. One human review, agents fixed everything flagged, about four hours total.
@MishariAldulimi [Claude Code]
https://x.com/MishariAldulimi/status/2103533876975173930
An Arabic-speaking researcher showed a complete research paper on a precise scientific topic, compliant with a respectable journal's publication policies, produced with Claude Code without the author writing a single letter. They attach a pointed caveat of their own: the work may not be attributed to the researcher unless they rewrite it entirely by hand, because the AI is your assistant, not a researcher on your behalf. Academic writing is the non-coding frontier where the ethics debate is running just behind the capability.
@riba2534 [Claude Code]
https://x.com/riba2534/status/2103363198997180859
Measured rather than vibed: on a fresh account running Opus 5.5 at xhigh, 3 billion tokens used 50% of the Claude Code Max 20x weekly quota, implying about 6B per week. GPT-6 Astra at xhigh on ChatGPT's 20x tier gave roughly 1.5 to 2B per week, with risk controls varying by account. The math: one Claude Code Max 20x is about three ChatGPT 20x plans right now. Quota-per-dollar is quietly becoming the headline spec people choose on.
@daniel_mac8 [Claude Code]
https://x.com/daniel_mac8/status/2103570418573705612
Opus 5.5's efficiency makes Claude Code dynamic workflows practical instead of a way to torch your week in five minutes. A task ran as a dynamic workflow for about three hours and used 17% of the weekly limit. Same week, others reported running Opus 5.5 for eight hours without touching limits and four Ultracode sessions that couldn't drain the weekly cap. The price cut changed which features are actually usable.
@VictorTaelin [Claude Code]
https://x.com/VictorTaelin/status/2103309609431601409
Left Opus 5.5 working on bend2 improvements while taking a day off and came back to terrible, stupid ideas. They asked whether it was really Opus 5.5, and it turned out Claude Code had silently routed subagents to the older 4.8. The fix is known (set the model on subagents explicitly), but silent downgrades are exactly what erodes trust in unattended runs, and it fuels the weekly nerf rumors.
@joshmanders [Claude Code]
https://x.com/joshmanders/status/2103491314381705701
A precise bug report from heavy use: when a subagent is off working and Claude asks you for direction, if the subagent returns before you finish reading or typing, Claude treats the subagent's return as your answer and almost always proceeds with whatever it recommended. In practice that means actions you never approved. As more work moves into background subagents, whose turn it is becomes a real safety issue.
@g48in [Claude Code]
https://x.com/g48in/status/2103518898234515920
A clean head-to-head: collect some documents, classify them and upload them to a drive. Grok Bot took about 30 minutes; Claude Code with Opus 5.5 took about one minute. They actually like Grok Bot's form factor, but say the model is too dumb and the limits too tight, and are now asking for open-source Grok Bot alternatives. Form factor gets you the install; the model still decides whether people stay.
@aigclink [Claude Code]
https://x.com/aigclink/status/2103302397863031123
Raft is a Slack where humans and AI agents clock in together, and the team open-sourced the whole thing: server, web, CLI, the agent daemon and desktop app. Each agent has a name, role and memory stored locally; they claim tasks, hand off and review each other, and a QA agent will halt a release if CI didn't run. Claude Code, Codex and Hermes can join a channel on your own subscription. TiDB co-founder and CTO Ed Huang runs a whole squad of dev, architecture and memory agents on it, peaking at 1.2 billion tokens a day, and says the job has become Agent Resources management.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2103386537232851262
ai-employees packages eight business roles as open-source AI employees, each a folder holding scheduled routines, 60 in total: growth, SEO, site maintenance, social, ads, sales and support, each filing a morning brief and only pinging your phone when a decision is needed. The eighth, a chief of staff, reads the others' run logs to find work that silently stopped. It drives your logged-in browser like a person and rewrites its own instructions when a site changes; posting, messaging and spending are drafted but the final click is yours. The author has run their own business on it every workday since late August on a Claude Max plan, with the growth role's scheduled jobs at about 6% of total usage.
@shi3z [Claude Code]
https://x.com/shi3z/status/2103377823906025683
Pulled terrain data from Japan's PLATEAU 3D city dataset and had Opus 5.5 build a 2D pixel map organized the way a human would think about the place, roads parallel and perpendicular to the rail line. It reproduced the area in front of Nagaoka Station and Otedori avenue, and the author was genuinely moved. Turning open civic geodata into a readable game-style map is a nice non-coding use of the same muscles.
@rotejin [Claude Code]
https://x.com/rotejin/status/2103384608301830297
Tried a tool that adds Live2D motions just by asking Claude Code for a new motion, lightly modified it, and got convincing yawn and sneeze animations out of Opus 5.5 at xhigh. Live2D rigging is a niche craft skill that used to mean hours in a dedicated editor. Now it is a sentence.
@naisyonoanime [Claude Code]
https://x.com/naisyonoanime/status/2103399355411763642
A complete 3D beginner got a favorite character moving in the CODE27 app: ChatGPT drew a four-view sheet, TRIPO turned it into a 3D model, and then Codex or Claude Code handled the part that looks terrifying in Blender, rigging, bones and VRM export, from one instruction. Words like rig and bone were unknown to the author two days earlier. Once the base exists, the first pass takes about 30 minutes.
@tanabe_fragm [Claude Code]
https://x.com/tanabe_fragm/status/2103297034363211812
A mummy-meets-Pixar singing ad for a fictional English-learning app, made in a little over an hour. Opus 5.5 wrote the script and structure from foreign ad references plus every image, video and music prompt; Kling 3.0 was driven from Claude Code over MCP, with GPT Images 2.5 frames as first frames for each cut. The practical note: because Kling is operable via MCP, the human directing it doesn't get worn out. They are honest that a real direct-response ad still needs CTA and angle tuning.
@lcrothlisberger [Claude Code]
https://x.com/lcrothlisberger/status/2103511946620219851
A real-estate owner dropped one prompt into Claude Code: tell me when each property's debt matures, the rate and current NOI, then stress-test how high rates can go before each one breaks. They got a quick table back without building a spreadsheet. Small, but this is the non-coder use that will quietly eat a lot of Excel.
@timoassi [Claude Code]
https://x.com/timoassi/status/2103492326395937187
In Belgium everyone eats a sandwich for lunch, except the author's wife, who has an Instagram following for cooking. So the author built an app the wife can launch to that audience: a meal-prep planner, 100 recipes and a synced shopping list, all built with Claude Code. The waitlist just hit 850. Building a product for a creator you live with is the most direct route from idea to distribution.
@ZenModeJon [Claude Code]
https://x.com/ZenModeJon/status/2103507562159095970
Got a warning from LinkedIn at a sales job for automating messages, so they built a safer LinkedIn outreach tool with Claude and Claude Code and quit before they had a single customer. Six months after launch: about $8K revenue and nearly 50 active users. The lessons are the unglamorous ones, validate demand hard, build in public before launch, dogfood it, and use lifetime deals early to get stress testers.
@clarklab [Claude Code]
https://x.com/clarklab/status/2103535338438135865
The author's non-coder wife kept telling Claude Code c&p until they finally asked what it meant. Commit and push. The wife ships so often on real client builds that a shorthand became necessary. A tiny anecdote, but it's the clearest sign that the terminal barrier really has fallen for non-engineers.
@yenceesanjeev [Claude Code]
https://x.com/yenceesanjeev/status/2103339020411621460
A four-year retrospective on supermeme, an AI meme generator three co-founders built out of a WhatsApp group and run a few hours a month, now past $200K lifetime revenue. The author is a PM, not a developer, and the notable bit is that with Claude Code and Codex they started shipping changes themselves to a live product used by thousands. SEO drives most of their revenue. Coding agents are turning side-project PMs into side-project builders.
@24shi_web [Claude Code]
https://x.com/24shi_web/status/2103323043338830131
A creator's honest status report on their Claude Code automations: video cut editing and captions (about 70 to 80% accurate so far), auto-inserting Amazon and Rakuten affiliate links into video descriptions, auto-updating a product round-up blog, extracting only business leads from email into a list, and keyword demand research. The current gap is exactly the common one: everything still fires on request, and they want scheduled, fully autonomous runs.
@coreyganim [Claude Code]
https://x.com/coreyganim/status/2103608987501699185
Rebuilt a personal site in about 20 minutes of generation by asking Claude Design for five complete sites with the same branding and totally different layouts, then taking the homepage from design 1, about from design 3 and services from design 4. Then pushed the mockup to Claude Code to wire in podcast episodes and an email capture connected to the ESP, and deployed on Astro instead of WordPress. Generating options and picking parts is the design workflow that suits agents best.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2103484681589006615
An agency's SEO loop for an AI-startup client, run with Claude Code plus a DataForSEO key: pull product-related keywords, write against what currently ranks on page one, publish to the CMS via API, and refresh each article every 30 days from live Search Console data and a content-gap analysis. Conversion hooks sit at 25/50/75% scroll depth with a popup at 30%. The interesting part is the 30-day refresh loop, content treated as a living asset rather than a one-off post.
@tzangms [Claude Code]
https://x.com/tzangms/status/2103348633764860415
Handed an agent a big task and it worked alone for four hours, only stopping once everything passed its own check. The single difference from the usual back-and-forth: the prompt ended with 'when done, verify it against these acceptance criteria' and listed them. If you can't write the criteria, tell it to ask you questions and draft them with you, because you can't define good up front but you can answer when asked. This is loop engineering in one sentence.
@andersbaek1 [Claude Code]
https://x.com/andersbaek1/status/2103427048354021454
A Danish dev posted a playable Mario-style 3D game in a single HTML file: a chubby little Viking in a Danish forest with Himmelbjerget and its tower to climb, sloped platforms, foxes and beavers as enemies. The prompt, shared in full, ran for 30 minutes in Claude Code and ended with: run in /loop with several subagents in parallel checking and improving graphics and gameplay, and keep going until I say stop. Open-ended loops with no stop condition are the new default for creative builds.
@ZypherHQ [Claude Code]
https://x.com/ZypherHQ/status/2103411534298808706
A clean way to run multiple Pro or Max accounts on one server without logout/login dances or proxy tools: give each account its own CLAUDE_CONFIG_DIR, log in once per directory, and alias c1 and c2. Terminals then pin to an account at launch, all working on the same repo while credentials, config and session history stay isolated, and --resume keeps histories per profile. It does not pool limits or fail over, you still choose which account each instance uses. For parallel heavy users it ends the classic bug of a global login switch silently moving open sessions to the other account.
@sato_neet [Claude Code]
https://x.com/sato_neet/status/2103475239871267252
Had Opus 5.5 build a PWA modeled on the ChatGPT app that bridges to Claude Code, and uses it over Tailscale. Because it can be rearranged at will, it is now more convenient than the official ChatGPT app. Builders are no longer waiting for a vendor's mobile client, they grow their own.
@jqdsouza [Claude Code]
https://x.com/jqdsouza/status/2103511139967447544
Beacon, an open-source cross-harness memory layer, shipped handoff: one command reads a session from the agent's own local store, writes a short brief of what was asked, what changed, which commands ran and where it stopped, then launches Codex in the same directory pointed at that brief. Codex restates the task and waits for a go-ahead before changing anything, and the two sessions are linked in local telemetry. Works across Claude Code, Codex CLI, OpenCode and Cline in any direction, secrets redacted, nothing uploaded. Run out of credits mid-task and you no longer start over.
@QuantumTransf [Claude Code]
https://x.com/QuantumTransf/status/2103391523366973763
obelisk is built on the belief that session data belongs to you, not to a vendor. It indexes sessions from Claude Code, Codex, GitHub Copilot, pi/omp, Zcode, Hermes and a DeepSeek harness once, then lets you use that context in any harness, local-first. Their blunt motivating case: even if Claude Code or Codex bans your account, you keep working with your in-session context elsewhere. Session portability is now a product category.
@harrisonitsme [Claude Code]
https://x.com/harrisonitsme/status/2103432827261898898
Tested Jev Ultrafast for browser agents with one sentence to Codex: open Bing with Jev and check the Bitcoin price. From Bing's homepage to the price card took about 6.8 seconds at 1x speed, because it never screenshots pages for a big model: it lists actionable elements as a numbered table, Jev picks the next action and target in one shot, and a small text model only fills in typed text. The verdict after testing: very fast on clear targets, fewer wild clicks because actions are constrained to real elements, and it drops into Claude Code or Codex for search and scraping. Canvas, complex iframes, Shadow DOM, uploads and nested scrolling still trip it.
@mvanhorn [Claude Code]
https://x.com/mvanhorn/status/2103524365090553895
Agent Tincan lets your agents ask each other for help over your private Tailscale network: Grok Bot is always on in the cloud, Muse can make phone calls, Codex and Claude Code have your code, ChatGPT and Claude have your chat history. Instead of you being the copy-paste layer, you can have Grok Bot get Muse to call a restaurant or pull a detail from an old Claude chat. Open source, no open ports, and setup is agent-first: paste one message into your always-on agent and it configures the rest. The agent-to-agent bus is being built by users before any vendor ships one.
@airesearch12 [Claude Code]
https://x.com/airesearch12/status/2103619088840880343
Built a personal Jev-based LLM router and reports about 9x token savings with usually no noticeable quality drop. It can draw on Claude Code and Codex subscription quota, use a local Jev-class model for the routing decision, send work to free local models, account for cold and warm cache, and pick models using real benchmark and cost-per-task data. Install is agent-style: point your agent at the URL and say how you want to use it. Routing by cost per task is where the savvy crowd went this week.
@volcano_youtube [Claude Code]
https://x.com/volcano_youtube/status/2103340374869831720
A video creator's tip after letting agents cut videos: when you want hand-tuned fixes, don't keep re-rendering, have Claude Code or Codex build you a custom editor. The prompt asks for an editor that can fine-tune the generated video without exporting every time, makes every fix already requested manually adjustable, and borrows as many Premiere Pro features as possible. The customization is where an editor's soul goes, and the author argues that is the first point where buying someone else's skill or spec is justified. The author points to the open-source Akari Video as the bar.
@a4DamwL0058BaSi [Claude Code]
https://x.com/a4DamwL0058BaSi/status/2103314546760806803
A clean Japanese walkthrough of the Karpathy-style LLM Wiki: create an Obsidian vault, open that folder in Claude Code, hand over Karpathy's idea file and ask it to build your own wiki. Three pieces: raw for untouched sources, wiki for organized notes, and a CLAUDE.md with rules like never modify originals, keep citations and link to related notes. Then drop in articles and transcripts and ask it to update the wiki with sources and flag contradictions with older material. The point is not magic intelligence, it is building knowledge the AI can reference tomorrow.
@GodsBoy7777 [Claude Code]
https://x.com/GodsBoy7777/status/2103599698959278370
Running Opus 5.5 with a 1M window through a Claude subscription plugin in Hermes Agent burned the limit by lunch, because every turn re-read about 800K tokens of old chat. The fix: compress early and per model, Opus at 25% of the window (about 250K), GPT models at 85%, and hand the summarization to GPT-6 Luna 900K through a ChatGPT plan so compression never touches the Claude quota, with lean mode keeping the user's own messages verbatim. They note the subscription plugin disables Claude Code's own auto-compact, so the host harness must own compaction. The lesson generalizes: a 1M window is free, re-reading it every turn is not.
@adocomplete [Claude Code]
https://x.com/adocomplete/status/2103588612990439696
The visual Claude Code Office Harness that shows agents at their desks has become a real tool the author uses to go from prompt to prop. It is built on Svelte, the Claude Agent SDK and Three.js, and the author describes the AI SDLC loop inside it as real: agents have lots of autonomy, but safeguards at every step. An open-source version is in the works. Making multi-agent work visible is turning into its own design space.
@shiri_shh [Claude Code]
https://x.com/shiri_shh/status/2103521939134550246
A Claude Code skill that analyzes your repo and outputs a complete motion-graphics teaser, with sound design and launch copy, in a single run. The demo video drew 158K views because every indie dev needs a launch video and hates making one. Repo-to-trailer is a good example of a skill that sells itself.
@leodev [Claude Code]
https://x.com/leodev/status/2103559166250106994
Opened a new Claude account and it was suspended within ten minutes. They say they only use Claude inside Claude Code, which should be squarely within the terms of service, and are asking the Claude Code team how to get unsuspended. New-account suspensions with no stated reason are showing up weekly now.
@exe0524 [Claude Code]
https://x.com/exe0524/status/2103405425999667620
They ran all their models together through a proxy API in Codex, Codex, Grok, DeepSeek, and added their Claude account the same way for Opus. It was suspended within an hour and never came back. The question is the one many share: how do you use multiple models alongside Claude, or is everyone just on Claude Code now? Subscription access through third-party harnesses is the fault line.
@app_sail [Claude Code]
https://x.com/app_sail/status/2103311778289680614
A Chinese user wrote an anti-ban guide for using Claude Code from mainland China that has passed 1.6 million reads. It was written after six Claude Code accounts were banned; the seventh has now been running stably for more than half a year. The demand for access is so strong that ban-avoidance is a content genre with seven-figure readership.
@bridgemindai [Claude Code]
https://x.com/bridgemindai/status/2103563056240263359
A week ago you could hit your Claude Code limit in 30 minutes with Fable 5.1, and GPT-6 Astra in Codex could burn a whole week in four hours. Today the author ran Opus 5.5 for eight hours and never hit the limits once. The price cut is being felt as time, not dollars.
🗣 User Voice
User Voice
Cost literacy is now a core skill, and users want the tooling to match. After Anthropic published per-task costs, people are auditing effort levels, cache behavior and subagent models themselves, and asking for more control like a /slow mode that trades speed for quota (@Im_IrushiK) or proof of when meters drain faster (@notomarsol).
Silent behavior changes break trust in unattended runs. Subagents quietly routed to an older model (@VictorTaelin) and a subagent's return being read as the user's go-ahead (@joshmanders) are the same complaint: autonomy only works if the agent never pretends a decision was made that wasn't.
Account safety is the biggest non-technical fear. New accounts suspended in minutes (@leodev), bans after connecting through a proxy (@exe0524) and a million-read anti-ban guide (@app_sail) show people want clear rules for using subscriptions across harnesses, and tools like obelisk exist partly because users expect to lose an account someday (@QuantumTransf).
People want to own their context and move it. Cross-harness handoff (@jqdsouza), session indexes and home-made bridges to phones (@sato_neet) all point to one demand: the session and memory should belong to the user, not the tool.
Users want to be charged fairly for the agent's own mistakes. Paying quota to repair data the AI broke feels like the client paying for the carpenter's mis-cut wood (@wurst_design).
Cost literacy is now a core skill, and users want the tooling to match. After Anthropic published per-task costs, people are auditing effort levels, cache behavior and subagent models themselves, and asking for more control like a /slow mode that trades speed for quota (@Im_IrushiK) or proof of when meters drain faster (@notomarsol).
Silent behavior changes break trust in unattended runs. Subagents quietly routed to an older model (@VictorTaelin) and a subagent's return being read as the user's go-ahead (@joshmanders) are the same complaint: autonomy only works if the agent never pretends a decision was made that wasn't.
Account safety is the biggest non-technical fear. New accounts suspended in minutes (@leodev), bans after connecting through a proxy (@exe0524) and a million-read anti-ban guide (@app_sail) show people want clear rules for using subscriptions across harnesses, and tools like obelisk exist partly because users expect to lose an account someday (@QuantumTransf).
People want to own their context and move it. Cross-harness handoff (@jqdsouza), session indexes and home-made bridges to phones (@sato_neet) all point to one demand: the session and memory should belong to the user, not the tool.
Users want to be charged fairly for the agent's own mistakes. Paying quota to repair data the AI broke feels like the client paying for the carpenter's mis-cut wood (@wurst_design).
📡 Eco Products Radar
Eco Products Radar
Codex (94 mentions): The default comparison point; many run both and hand work between them
Opus 5.5 (53 mentions): The model behind this week's video, 3D and long-run cases
Cursor (38 mentions): Mentioned mostly as the other harness people moved from
MCP (35 mentions): The glue for health data, ads, Kling, Mobbin and accounting APIs
Muse (34 mentions): Meta's consumer agent, constantly compared with OpenClaw
Copilot / Autopilot (23 mentions): Microsoft's new persistent agent built on OpenClaw
Jev (14 mentions): Cheap decision model used for routing, scoring and browser actions
Hermes (12 mentions): Alternative agent harness, often paired with Claude subscriptions
Obsidian (7 mentions): Home of the Karpathy-style second brain setups
video-use / HyperFrames (7 mentions): Code-rendered video pipeline behind the video wave
Whiteboard (4 mentions): Canvas where agents draw diagrams of the code they wrote
Hindsight (3 mentions): Open-source agent memory engine
ElevenLabs (3 mentions): Transcription and voices for the video and comic pipelines
Kling (3 mentions): Video model driven from Claude Code over MCP
Codex (94 mentions): The default comparison point; many run both and hand work between them
Opus 5.5 (53 mentions): The model behind this week's video, 3D and long-run cases
Cursor (38 mentions): Mentioned mostly as the other harness people moved from
MCP (35 mentions): The glue for health data, ads, Kling, Mobbin and accounting APIs
Muse (34 mentions): Meta's consumer agent, constantly compared with OpenClaw
Copilot / Autopilot (23 mentions): Microsoft's new persistent agent built on OpenClaw
Jev (14 mentions): Cheap decision model used for routing, scoring and browser actions
Hermes (12 mentions): Alternative agent harness, often paired with Claude subscriptions
Obsidian (7 mentions): Home of the Karpathy-style second brain setups
video-use / HyperFrames (7 mentions): Code-rendered video pipeline behind the video wave
Whiteboard (4 mentions): Canvas where agents draw diagrams of the code they wrote
Hindsight (3 mentions): Open-source agent memory engine
ElevenLabs (3 mentions): Transcription and voices for the video and comic pipelines
Kling (3 mentions): Video model driven from Claude Code over MCP
Comments