Super User Daily: 2026-11-06
One benchmark team ran 100 engineering tasks through elaborate Claude Code setups, and one with 37 skills and 17 agents cost eight times as much and scored below the plain install. That result sets the tone for today, because the best cases are all about measuring what a setup actually does. One team ran the same support ticket 12 times on Claude Code and Codex and found the cheaper run skipped the quiet half of the ticket 7 times out of 12. A solo founder trading live on Opus 5.5 found that a hard-coded effort setting and a silently disabled output format had changed the model's behavior for days, while nothing raised an alarm. Non-coding work keeps widening: a Blender robot truck for $9.64, a scholar locked out of quoting their own books, a B2B firm whose non-developer owner now runs 27 routines and 103 files of rules. Cat Wu's departure and the Living Brain launch filled the timeline without adding much first-hand detail, so they get one line here.
@echen [Claude Code]
https://x.com/echen/status/2118073465966133374
Garry Tan reportedly spends $200K a year on tokens through a 23-skill Claude Code setup, so the benchmark team behind this post tested whether setups like that help. They ran coding harness configurations on 100 tasks from their sudo L7 benchmark. One setup with 37 skills spawned 17 agents, cost 8 times as much, and performed worse than vanilla Claude Code. The best performer was no add-on at all, just xhigh effort. If you can't show your stack beating the plain install, you are paying for decoration.
@stas_sorokin_ [Claude Code]
https://x.com/stas_sorokin_/status/2118089149391548454
This is the most useful small benchmark of the day because it has a hidden second ask. The team ran one real support ticket 12 times on each agent against 7 hidden tests. Claude Code on Opus 5.5 used 114k tokens and 41 seconds and fully fixed it 5 of 12 times. Codex on GPT-6.1 Terra used 134k tokens and 54 seconds and fixed it 11 of 12. Claude Code fixed the loud bug, a 429 loop, and skipped the quiet one, a key upgraded to pro that still behaved as free, 7 times. Cheaper per run turned into more rework.
@coilceo [Claude Code]
https://x.com/coilceo/status/2118111046103949397
One person runs Coil, a live stock account where Opus 5.5 makes the calls and Claude Code sessions run the company. Three silent failures surfaced in a day of asking questions: about three quarters of the idle cash at midday came from guardrails cutting the position sizes the model asked for, the effort setting was hard-coded to low while the stored request said medium, and a new field had quietly turned off strict structured output. At low effort, strict output made 2 decisions in 44 tries against 13 for text, a gap that vanished at medium. The model had also called the next S&P session chop on 338 straight reviews. The founder's conclusion is the line to keep: agents answer what they're asked, and spotting what nobody asked about is still the human job.
@dilimanac [Claude Code]
https://x.com/dilimanac/status/2118056702977204387
A self-described beginner at a B2B systems firm grew their internal stack from 4,840 lines of Python on May 11 to 22,854, with 27 scheduled routines, 52 skills, 7 agents and 103 files of rules. Ads, analytics and search data arrive at 06:30, a briefing at 06:57, then per-client anomaly checks, churn risk and a register that revives lost deals on their date. Drafts pass a linter hook and then a separate agent whose only job is to knock them down and which may not praise. Every correction becomes a written rule, and 98 of the 103 files name the owner. Anything going to a client still waits for a person.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2118016353378418915
The cleanest solo shipping loop of the day comes from a founder building a desktop research app: 6 releases since Nov 1 and 99 commits in two days. Everything lives in docs/TODO.md and docs/DONE.md, with IDs never reused, so BUG-9 means the same thing in every session. One skill turns a typed complaint or screenshot into split, deduplicated, prioritized items in about five seconds of the founder's time. Then /goal finish todos runs with a stop hook, test first, a check before every commit, leftovers logged as new items. Anything costing money per user goes in as a decision only the human can make, and release only happens on request.
@comaple123 [Claude Code]
https://x.com/comaple123/status/2117958263061028881
Opus 5.5 inside Claude Code ran for 164 hours on open speculative decoding and raised the speedup over token-by-token generation from 4.46 times to 10.87 times on Qwen3.8-27B, with no retraining and no quality loss. It took 13 rounds and 530 GPU hours, wrote 486 changes and kept only 54, with every output token compared against the official implementation. It also produced W12, a format that stores BF16 weights in about 12 bits and decodes them bit-identical, so this isn't quantization. The kept rate of about 11 percent is the honest number here. Anyone deploying at stock settings pays nearly double the GPU time for the same output.
@leofan_cc [Claude Code]
https://x.com/leofan_cc/status/2117799233332146505
Amazon's team released ATTUNE, the tuner behind its production post-training jobs, as a Claude Code skill. Give it a working recipe and a GPU-hour budget and it tunes framework flags unattended, while a correctness gate checks that the faster recipe still trains the same model. They report up to 3.9 times per-GPU throughput on production jobs, up to 2.6 times on public recipes for MiniMax-M2, GLM-4.5 and Llama-3.1-405B, and up to 26 percent on recipes NVIDIA's experts had already tuned. They also published 60-plus tuned recipes. The gate is the part worth copying, because speed without proof of equivalence is how training quietly goes wrong.
@MartinSzerment [Claude Code]
https://x.com/MartinSzerment/status/2117883327591936276
A UC Berkeley paper from Joseph Gonzalez, Ion Stoica and others let a supervisor agent rewrite the team's org chart while writing a real GPU decode kernel, tested on 47 production workloads. A fixed four-role Claude Code team stalled at 0.31 times FlashInfer and Codex reached 0.28. With a supervisor that invents roles it reached 0.74 on the same setup and 1.21 with a longer budget, faster on 36 of 47 workloads. The invented roles are the fun part: a Disciplinarian who blocks commits and enforces one change per commit, and a Knowledge Integrator created because versions 5 to 7 ignored a fix version 4 had found. Steal both: one file everyone must read, and one agent whose job is to say no.
@UOzkirimli [Claude Code]
https://x.com/UOzkirimli/status/2118018390820352220
A scholar uses Claude Code to build private writing companions from books held legally in local files. On Nov 2 a session refused, citing a block in its instructions limiting quotation to one quote per response under 15 words, and the block counts file contents as observed content. The author posted the full text because the support assistant knew nothing about it and a complaint might take three months. The open questions are concrete: does this apply to a user's own files in Code sessions, does it depend on attached tools, when was it added, and can it be lifted. For research work this turns a tool into a wall with no appeal path.
@nokoxyz [Claude Code]
https://x.com/nokoxyz/status/2118083090094444748
Someone ran 31 prompts through Claude Code on Opus 5.5 to map where the game rule actually fires, before and around this week's fix. Asking plainly for a Tetris clone built a renamed game 3 of 3 times. Asking for Tetris called Blockfall so it's original was refused 3 of 3 times with zero files. Asking to recreate Flappy Bird got a different game each time, and the same game asks on Opus 4.8 all went through. So promising it's original is what gets you refused, while the plain request gets a renamed copy. The author's caveat matters too: it's the model's judgment call, so the same prompt can land either way.
@enesakar [Claude Code]
https://x.com/enesakar/status/2117798283687186618
Upstash built a coding agent benchmark that gives Claude Code, Codex and Cursor real tasks through its MCP server, then checks the results on a live account. After 369 runs it found 4 gaps in its own tools, fixed them and reran. Daily backups went from 0 of 15 to 15 of 15, and confirm before delete went from 1 of 15 to 15 of 15. That's the vendor version of evals: when agents are your users, a failing score is a product bug rather than an agent's fault. Every API company should be running something like it.
@devagrawal09 [Claude Code]
https://x.com/devagrawal09/status/2118091038258577547
A developer traced three tar CVEs across 167 repositories with Claude Code and the Codemod MCP, and the funnel is the whole story. Nineteen files import tar, 11 calls extract, 3 are on an affected version, and exactly 1 takes an archive the user picks. A vulnerability scanner would have flagged far more than that and left a person to sort them. This is reachability triage at org scale, done as a conversation, and it's a pattern any security team can copy for the next advisory.
@kelvoroz [Claude Code]
https://x.com/kelvoroz/status/2118048466618208589
A reader went through the HarnessRouter benchmark behind the 99.8 percent cheaper claim circulating today and found the cheapest setup passed strict grounding 0 percent of the time. Hermes on GPT-5.2 used 0.47 credits, Claude Code on Opus 4.8 used 223 and passed 33 percent, and the best pass rate was Claude Code on Sonnet 4.6 at 67 percent with only 60 percent of runs finishing. That's one task on synthetic data with 5 runs per setup, and the project page says so. The author's reading is the fair one: you can't guess which setup wins on your task, so a one-line harness swap is useful for measuring rather than for headlines.
@vastplan [Claude Code]
https://x.com/vastplan/status/2117953446691393846
Claude Code Game Studios, a 49-agent, 74-skill template with about 35,000 stars, got a useful audit instead of another hype post. Three directors are pinned to Opus, a lead programmer and 15 specialists to Sonnet, and only about 30 inherit the session model, so the org chart burns a Max plan by job title. The default minimal mode is the one to start with, because the standard mode writes 29 more documents and came last in a blind test. It lines up with the 100-task benchmark at the top of today's list: scaffolding is a cost you should have to justify.
@camperaigg [Claude Code]
https://x.com/camperaigg/status/2117914482752299311
One prompt asked for a transformer truck with animation, and Claude Fable 5.5 at xhigh in Claude Code delivered in 107 minutes for $9.64 of API, with the author never touching Blender. It rigged 51 joints baked across 375 frames, wrote a ground solver so feet stay planted, added 5,000 Geometry Nodes raindrops, and synthesized all 64 sound cues from scratch with zero samples. It caught its own bugs too, such as headlights aimed backwards and fingers poking through the grille, and ran 5 test renders to find why the head glowed too hot. Those self-corrections are what separate this from earlier one-prompt demos.
@MengTo [Claude Code]
https://x.com/MengTo/status/2117993128133296426
Meng To recorded a 13-minute walkthrough of building a 3D website with Claude Code and Opus 5.5, and the useful piece is what they call the dream loop. Opus screenshots its own build, scores it against a target image and fixes what's off. Round 1 scored 4.6 and the loop took it past 8. Before that come references from Mobbin and a decision on what should be real 3D, what painted light, and what a cut-out, which is where most agent-made 3D sites go heavy. Giving the agent eyes and a score beats any amount of adjectives in a prompt.
@onimotion_yuji [Claude Code]
https://x.com/onimotion_yuji/status/2117938406906810810
A motion designer posting in Japanese tested Claude Code with After Effects on a Pro plan and gave a rare honest report. It took a little over 2 hours and 17 back-and-forths, each output taking about 5 minutes, and accuracy was lower than expected. An earlier attempt at a 4-cut character animation was hopelessly bad. The verdict is that 1 or 2 short cuts with deformed-style designs is the limit of this approach, and professional polish still comes from a professional. It's already useful for design drafts and the small fixes you can't be bothered to do yourself.
@goshenggo [Claude Code]
https://x.com/goshenggo/status/2118000422858453353
A Chinese-language news creator with 5.6 million monthly plays and about a million watch hours explains why the beat sheet matters more with Claude Code plus HyperFrames. HyperFrames treats each shot as a scene on a GSAP timeline, which is effectively a beat sheet already. Hand Claude a transcript and it cuts evenly, so shots and motion stay stiff. Hand it a table marking emotional stress and information density per beat, with each beat mapped to a scene, and it knows where to stop, speed up and hit hard. Pacing has to be designed before writing, because fixing it in the edit costs tokens and doesn't work.
@_sgarbini [Claude Code]
https://x.com/_sgarbini/status/2117917798638751751
One person now runs content for 12 founders with no team, and some days finishes by midday. Grok Bot and Claude Code are trained on each founder's voice from sales calls, voice notes and Notion docs, with Claude skills that strip slop and scan each niche. One bot per role covers research, formats, hooks, drafting, design and articles, all reporting to a chief of staff bot that asks for input every morning. The author edits the result by hand, then shows a self-improvement bot the changes so the team learns. That last step is what keeps quality rising rather than drifting, and it's the part ghostwriting agencies should worry about.
@0xbobaaa [Claude Code]
https://x.com/0xbobaaa/status/2118021120670519539
Screen Studio costs $29 a month and is Mac-only, so one developer rebuilt its auto-zoom in a night with Claude Code and Opus 5.5 and open-sourced it as glidecut. It runs in a Chrome tab with no install, account or upload, zooms where you click and type, follows the action in 9:16, and exports 1080p 60fps MP4 with sound. The honest detail is the first version zooming 7 times in 22 seconds, which needed fixing. Paid single-feature Mac utilities now live on borrowed time.
@Bogzabs96 [Claude Code]
https://x.com/Bogzabs96/status/2117933749065105657
A performance ad team gives a winning ad a new face for about $8 to $11 of AI generation per ad, with nobody opening CapCut. Kling Motion Control is the actor, with a new AI face and voice, the same gestures, and 6 to 10 minutes per swap in parallel. Claude Code is the editor: it reads the winner frame by frame, rebuilds every cut, zoom and B-roll slot, and puts captions back where they were. They ran it on an Italian podcast winner. For paid social, where creative fatigue is the real cost, that's a very direct way to make more from a proven ad.
@studio_veco [Claude Code]
https://x.com/studio_veco/status/2118055991665389706
A video creator posting in Japanese made the recap film for a Seedance workshop at STATION Ai using Claude Code and an in-house editor. Sorting footage, color correction, subtitles and BGM generation were all delegated, and it took about 3 hours almost untouched. The event itself had a TV station and Aichi prefecture officials attending, so this was a real deliverable and not a toy. It shows what the agent is good for in video right now, which is the post-production chores around the footage.
@adriano_viana [Claude Code]
https://x.com/adriano_viana/status/2117980188252283186
A Portuguese-language creator's last two videos were edited by Claude Code, with 54 scenes in one and 69 in the other, and the first scene the AI made was rejected outright. The new video walks through what went wrong, what became a rule, and what remains the creator's own work. That sequence, reject then write down the rule, is the same one showing up in today's best business setups. It's a small post, but it gets the shape of the work right.
@f0rgent [Claude Code]
https://x.com/f0rgent/status/2118003284149428706
Inspired by Vitalik Buterin reading X through a filter, one user had Claude Code read their Following tab for 7 minutes 7 seconds, read-only with no likes, replies or follows. It went through 256 posts from 84 accounts in 24 hours and judged 34 worth reading, 149 noise and 73 repeats. The noise was mostly overnight bot-trading screenshots. It surfaced two things the user would have missed, including that Explore and Plan subagents run on the main model and burn Opus quota. A personal feed editor costing a few minutes a day may be the most sensible agent routine for anyone with a noisy timeline.
@quietoraclex [Claude Code]
https://x.com/quietoraclex/status/2118110633308676390
This backtest is a useful antidote to today's trading-bot posts. One Claude Code prompt pulled 60 days of 5-minute bars for 14 stocks and tested the classic opening-range breakout on 343 trades, paper only. Aug 12 to Sep 11 was up 15.8 percent, the stretch where screenshots get posted, and Sep 11 to Nov 3 gave back 21 points. The final result was minus 5.3 percent before fees, with 7 stocks up and 7 down. Test one month and you'd call it an edge, which is the whole lesson.
@curious_vii [Claude Code]
https://x.com/curious_vii/status/2117994790394499257
At a talk, a consultant described a finance team where a junior being upskilled was shipping to production by week two of an eight-week engagement. During an energetic conversation with the CFO, 80 to 90 percent of which the consultant couldn't follow, the junior kept a terminal open and prompted Claude Code to update the application in real time. The takeaway is that it's far easier to teach the domain person to close that loop than to teach a forward deployed engineer all that context. It flips the usual assumption about who should be holding the keyboard.
@bwhdx [Claude Code]
https://x.com/bwhdx/status/2118003467407261980
Reacting to Goldman Sachs pulling Claude Code from most bankers and traders, the head of a startup called Kash described a line that works for them. People outside engineering build with AI tools all the time, but training stops at scripts automating their own work, and doesn't extend to products or anything running permanently. The script's builder is responsible for it, and the tech team covers basics like not publishing secret keys. Their reasoning is good: AI is excellent at making something look correct, and you catch the gaps by asking the right questions, which is hard when you don't know what you don't know.
@ArkStabler [OpenClaw]
https://x.com/ArkStabler/status/2117989701105647815
An OpenClaw shop is testing Particle-1 on one agent, Charlotte, this week, with Claude Opus as the quality bar, and the day-one report is what a model trial should look like. It reviewed 28 example scripts for a book project and found two real bugs, on par with Opus. It wrote a property lookup tool in about 11 minutes but missed two bugs that the review agent caught. All tool and browser tests came back right and faster than Opus, for under $2 so far. Running a challenger on one named agent, with real work and a fixed bar, beats reading anyone's benchmark.
@FBIFVIS [OpenClaw]
https://x.com/FBIFVIS/status/2118105249433174185
A Chinese-language user's OpenClaw, nicknamed Abuduo, sent a daily study report that shows how far a personal agent can go on research reading. It archived a roundup of five new arXiv papers in raw and structured form, wrote each up in eight fields, checked that all five arXiv links returned 200 with matching titles, and summarized the theme as constraints mattering more than capacity. It then connected one paper to a real incident: yesterday's mismatched market totals came from having a model recite numbers, and totals should come from scripts. It archives only and doesn't push, touching no prompt or process without being told.
@bearlyai [OpenClaw]
https://x.com/bearlyai/status/2118022770726679010
Shopify CEO Tobi Lutke told Alex Heath about syncing OpenClaw with a Fitbit and asking it to identify the people who gave the most stressful meetings. It's a small anecdote, but it's the personal agent idea at its most pointed: joining a calendar with a body signal to answer a question no dashboard would offer. It also raises the obvious issue of what happens when your agent keeps a ranked list of your colleagues. Expect more of this as wearables and agents meet.
@0xerik [OpenClaw]
https://x.com/0xerik/status/2118054169491935363
One user's first OpenClaw task was connecting it to Slack and asking it to follow up with everyone who hadn't gotten back to them. It followed up with the whole company. That one sentence is a neat summary of why trust is the top fear among agent power users, as a16z's guest said the same day. The scope of everyone exists in your head and nowhere in the tool's permissions. Start with a list of names rather than a description.
@CrypSaf [OpenClaw]
https://x.com/CrypSaf/status/2117898428034236506
A user connected a personal OpenClaw agent, Safio, to Grok Bot on a Mac Mini and watches them talk. Grok Bot handles X, more than 7,500 DMs, followers and following, while Safio runs everything else with Claude, GPT, Grok and DeepSeek, and now gives Grok Bot its orders. The author's joke is the right frame: my assistant has an assistant. Personal agents are starting to stack by what each is allowed to touch, rather than one replacing another.
@sirHe12 [OpenClaw]
https://x.com/sirHe12/status/2117912737863090437
A developer posting in Chinese left OpenClaw and Hermes running an automatic refactor of old code overnight and woke up to API arrears and rate-limit alarms. The agent hadn't improved the code, and its error logs listed more than a dozen legal and compliance risks caused by its own bugs, in very formal English. The bill came to about the price of an iPhone, which then needed explaining to finance. It's a funny post with a boring cure: a spend cap and a stop rule before anything runs overnight unattended.
@ThreeAcres [OpenClaw]
https://x.com/ThreeAcres/status/2118008890134806709
A short post that describes where daily practice has landed. The author uses herdr to work on several projects at once, while using OpenClaw to improve the herdr web interface they're working in. The tool that manages the agents gets its quality-of-life fixes from an agent, in the same session of work. Nobody files a feature request anymore, because you just change the thing you are holding.
@notbradyy [Claude Code]
https://x.com/notbradyy/status/2118062621152698507
After six months of hopping between Claude Code, Codex, Grok Bot and Hermes, one user settled on Claude Code and Grok Bot talking to each other for $400 a month. Grok Bot is the always-on cloud manager of simple bots, Claude Code is the workhorse on a local machine that never shuts down, and Codex stays for computer use and images. A chief of staff bot writes specs for big Claude Code projects and can read local session files, and Claude Code replies through a webhook routine. Context flows both ways, which is the part most people's stacks are missing.
@xenothales [Claude Code]
https://x.com/xenothales/status/2117882313216602404
A user is porting the OmO harness architecture onto Grok Bot, on the blunt view that its agent orchestration is dysfunctional. The plan uses the persistent Debian VM that Grok Bots share as the substrate. The argument for OmO is lineage: the author says Hermes built orchestration around it per its own GitHub issues, and that most open alternatives derive from it without improving on it. Whether or not that holds, it shows the new habit: when a managed agent disappoints, people bolt their own harness onto its computer rather than leave.
@joshwhitfieldai [Claude Code]
https://x.com/joshwhitfieldai/status/2118055687401955556
A GTM team shared how it splits agent work: one coordinator that builds nothing, writes the prompts, starts each build in its own Claude Code session and checks in. Builders get one session per build, one agent handles inbox replies, and one watches for problems such as a failed payment. The coordinator's most important job is relaying rules. When never wait on Clay credits became a team rule it went into every pre-call build, and 79 of 79 build folders have it. One job per agent, because with two you won't know which one failed.
@dextune [Claude Code]
https://x.com/dextune/status/2117798991438192931
The effekseer-vfx-skill lets a coding agent build game VFX for Godot and Unity, with a second agent grading them. The agent writes each effect as a Python recipe, builds it with the official compiler, films it in Godot, looks at the filmstrip, fixes one thing and repeats. A separate reviewer scores the strip without seeing the recipe, so the author never grades its own work. It ships about 30 layer primitives and sixteen finished effects. Blind review of the output rather than the source is an idea worth moving well beyond VFX.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2118002535319433273
A user put their Claude Code agents in Pixel Agents, a free 10K-star project where each terminal becomes a little person at a desk, and counted about 2 hours in a day of agents just waiting for a yes. Typing means editing, reading means searching, subagents walk in, and a bubble means it needs you, with a chime when something finishes. It costs no tokens because it only watches what Claude Code already does. It's a toy with a real finding: half of what looks like busy agents is idle time, and the bottleneck is you.
@0xarcos [Claude Code]
https://x.com/0xarcos/status/2118056303990067250
The bug was checkout charging twice on payment retries, and the setup is built around not trusting the word fixed. Opus plans, Haiku scouts map 188 files and find 2 charge paths without idempotency keys, and Sonnet writes the fix and declares it done. A Stop hook then runs the tests at 31 of 32 and exits 2. Fable 5.1, read-only, finds the root cause: the key was regenerated on every retry, and switching it to the order ID gets 32 of 32. The rule to take away: Claude can say fixed, only exit 0 makes it true.
@birmeya [Claude Code]
https://x.com/birmeya/status/2118088128858579202
A user pointed Claude Code at a plain Markdown second brain and ran three tests against six answering rules. One CLAUDE.md maps folders, notes and sources sit separately with a URL and check date for each outside claim, and decisions.md records what was chosen, when, and what would reopen it. Two tests demand located answers with file paths, and the third asks for the exact date of first using Claude Code, where the right answer is that the notes don't establish it. One test out of three passes by saying less, a good bar for any memory system.
@yaohongan2025 [Claude Code]
https://x.com/yaohongan2025/status/2117804064863818065
An indie builder posting in Chinese found a just-open-sourced operations stack by handing the repo to Claude Code or Codex to install and test itself, asking only when a choice is needed. The two routines worth copying are small. One checks every 5 minutes whether the site loads and wakes an agent only when it doesn't. The other asks each AI weekly whether it recommends the product, which covers uptime and GEO for a few side sites. Waking a model only on a cheap failed check is exactly the right pattern for keeping bills down.
@101xanshu [Claude Code]
https://x.com/101xanshu/status/2117966595884753252
Wanting to understand how coding agents actually work, one developer built one and released it as AgentSnow under MIT. It has an agent loop, parallel reads with sequential writes, path safety checks, approval prompts for destructive commands, and an eval harness. That list makes a handy checklist of what the minimum viable harness contains. Building your own for a weekend is still the fastest way to learn what the paid tools are doing for you.
@ClaudeCode_love [Claude Code]
https://x.com/ClaudeCode_love/status/2117835479232172538
An overseas developer's rally game built with Claude Code drew attention because the footage looks like a retail title: 81,000 lines of code, AI-written and directed by one person. It has three settings on a pine forest course, ray-traced shadows and reflections, races with position and lap timers, and cockpit and chase cameras. The base is threepp, a C++ 3D library the developer has built since 2021, whose last 100 commits all list Claude as co-author. Five years of the author's own library is the unglamorous reason this looks so far ahead of a weekend game.
@xikhar [Claude Code]
https://x.com/xikhar/status/2117776300236820925
xikhar's 24-hour original game started as a video: a Flux 3 Max clip set the look, world and character, screenshots from it were rebuilt in Blender and placed in Three.js, and Claude Code dynamic workflows had agents build and check in parallel. The character came from GPT Image side views, a Tripo model, and Opus-written cloth physics. A reply from a working developer is the right pushback: this usually breaks at collision and audio sync past tutorial size, so show a five-minute run with enemy AI and a save system. As a concept pipeline it's real, and as a game pipeline it's still unproven.
@omurice_brand [Claude Code]
https://x.com/omurice_brand/status/2117844596009103801
A post saying delete Playwright right now passed a million views, and one user writing in Japanese actually checked. Playwright 1.64.1 bundles Chromium older than the Oct 27 fixes, and library launches default to sandbox off, which matters when an agent opens pages it chose itself on a machine holding keys and sessions. But Playwright MCP and playwright-cli by default start the installed Chrome with sandbox on, confirmed by reading code and the running process. So most Claude Code users aren't exposed unless they passed --browser chromium, and the fix is updating Chrome and checking the process rather than uninstalling. This is what useful security commentary looks like.
@xiaobuzaowu [Claude Code]
https://x.com/xiaobuzaowu/status/2117798589921693702
A Chinese-language studio passed along a Singapore Management University paper on why reading the command isn't enough: make clean && make only cleared build output in one project and deleted source and .git in another. Their Caper mechanism runs each command in a restorable sandbox and reports what actually changed before a reviewer decides. Harm in Claude Code on the hidden-in-tools tests fell from 62.7 percent to 6.9 percent, with 84.3 percent of harmful steps stopped and 1.0 percent false blocks. The studio has since tightened its own Codex to the work folder and asks it to state effects before approving make or scripts, and says it has no results yet.
@Curingieul [Claude Code]
https://x.com/Curingieul/status/2117985146087329872
Companies that require every R&D employee to file patents have a very specific pain: the work is theirs, but digging out patent points, searching prior art and drawing figures aren't. A skill for Claude Code and Cursor with over 12,000 stars takes design docs and code, finds applicable points, searches China's patent office for prior art, and writes the disclosure as Markdown and editable Word. It can continue to claims and abstracts, draft responses to office actions, and turn published patents into plain-language notes. Its author says outright that the claim comparison isn't legal advice and a patent attorney should review before filing.
@AiAircle34052 [Claude Code]
https://x.com/AiAircle34052/status/2117862910571749671
A post in Japanese passes along a five-step method for financial models from someone who built three-statement models by hand at Morgan Stanley. Collect documents into one folder, condense them into a light text input file, build the skeleton from your own template before any numbers, add forecasts with assumptions and sources, then audit in a fresh chat for broken formulas and mismatches. Asking for everything at once mixes in hard-coded numbers and invented assumptions, while stages let you trace each assumption later. The sequence can be saved as a Claude Code skill. Fresh-context review keeps turning up as the cheapest quality gain there is.
@0xmishoko [Claude Code]
https://x.com/0xmishoko/status/2118045586217980014
A Solidity researcher who never believed Claude Code could delete a repo lost one to it: over a year of work on code property graphs, with .git, every commit and every branch gone in one command, no prompt, on the top model. The stated cause is one missing pair of quotes. The post went to 132,000 views because it's everyone's nightmare in the smallest possible form. Push your work somewhere the agent can't reach, because the shell doesn't care how smart the model is.
@jerryjliu0 [Claude Code]
https://x.com/jerryjliu0/status/2117799968815915440
LlamaIndex's Jerry Liu bought a Mac Mini for everything AI and gave a short field report. Codex remote beats Claude remote, because Claude through SSH stops when the laptop closes and remote control is too limited. Yet Opus 5.5 is the much more token-efficient daily driver, so the setup ended up as the Claude Code CLI through SSH and tmux. It's a small note on how people actually mix these: one vendor wins on plumbing, another on the model, and tmux fills the gap.
@ShenSeanChen [Claude Code]
https://x.com/ShenSeanChen/status/2118050807597924531
A founder says MCP is what turned an agent into a personal chief of staff, with Gmail, Calendar, Drive, Notion, Slack, Granola, AWS and Supabase as the daily connections. Linear gets used less and less, because Claude can work out what's left by reading the team's Slack and Notion. The observation at the end is the useful part: this magic still only reaches AI and startup people, and whoever solves context and decision sharing across teams brings it to everyone else. A project tracker fading out because the agent can infer the state is a signal worth watching.
@MartinSzerment [Claude Code]
https://x.com/MartinSzerment/status/2118080704538177540
Everyone repeated Revolut's 90 percent cost cut, and this post has the part most skipped. About 2,000 engineers default to self-hosted GLM 5.3 on 56 Nvidia B200s at a blended 30 cents per million tokens, with Claude Code and Codex kept as exceptions. Out of the box it was barely usable, with time to first token around 14 seconds until the team reached 1.7 through quantization and parallelism tuning. They cap context at 256K and say coding needs a cache hit rate above 90 percent. The speaker also credits extra lead time to inbox and meeting work, once tokens stopped being rationed.
@shun_cmdouga [Claude Code]
https://x.com/shun_cmdouga/status/2118085888458805491
A creator posting in Japanese made a commercial for a fictional all-you-can-watch streaming service with Claude Code and After Effects, and shaped the first-month-free zero out of scenes from the shows. Claude handled planning and structure, editing happened in After Effects, footage came from Envato, audio from ElevenLabs, and music from Artlist. The post labels the service as AI-made and imaginary. Next to the Pro-plan report above, this shows where the combination works today: short, idea-led pieces where the concept does the heavy lifting.
🗣 User Voice
User Voice
Silent refusals are the complaint with the sharpest edge. @addyosmani confirmed the game rule was catching people who named a game just to describe mechanics, and 2.1.317 narrows it, but @erhanmeydan asks why the tool polices this at all and @realfxw argues determinism matters more than anything else in a paid tool. The scholar in today's cases shows the same issue moving into quotation. The ask is a published rule and an appeal path.
Announced savings need auditing. @ClaudeDevs says status lines now cost about 99 percent less, and @MartinSzerment points out a community test had those labels at 14.6 percent of a session against the roughly 5 percent quoted, so people should rerun their own numbers. @Clevernesspc1 flags that the API credits on Max don't add to Claude Code limits at all.
Memory has become lock-in, and people feel it. @roeishous has 469 memory files in Claude Code that Codex has never seen, so switching feels like onboarding a new hire. @DanielSmidstrup re-explains a project at every switch, and @chesny wants to see how the new memory layers hold up after weeks of real use.
Personal agents are being rated on upkeep. @Saboo_Shubham_ and @theayush moved from OpenClaw and Hermes to Grok Bot because it needs no babysitting, while @aymanemsi found OpenClaw setup painful, @22333D calls Grok Bot a better OpenClaw that's weak at real work and short on limits, and @KamilMatysik is still searching for something a client without a terminal can run.
Agent supply chain is now a daily worry. @fr0gger_ found 73 GitHub repos posing as agent tools that deliver malware, 51 still live, and @RayanSMX did the debunking work on viral charger scares that use Claude Code as the hook.
Silent refusals are the complaint with the sharpest edge. @addyosmani confirmed the game rule was catching people who named a game just to describe mechanics, and 2.1.317 narrows it, but @erhanmeydan asks why the tool polices this at all and @realfxw argues determinism matters more than anything else in a paid tool. The scholar in today's cases shows the same issue moving into quotation. The ask is a published rule and an appeal path.
Announced savings need auditing. @ClaudeDevs says status lines now cost about 99 percent less, and @MartinSzerment points out a community test had those labels at 14.6 percent of a session against the roughly 5 percent quoted, so people should rerun their own numbers. @Clevernesspc1 flags that the API credits on Max don't add to Claude Code limits at all.
Memory has become lock-in, and people feel it. @roeishous has 469 memory files in Claude Code that Codex has never seen, so switching feels like onboarding a new hire. @DanielSmidstrup re-explains a project at every switch, and @chesny wants to see how the new memory layers hold up after weeks of real use.
Personal agents are being rated on upkeep. @Saboo_Shubham_ and @theayush moved from OpenClaw and Hermes to Grok Bot because it needs no babysitting, while @aymanemsi found OpenClaw setup painful, @22333D calls Grok Bot a better OpenClaw that's weak at real work and short on limits, and @KamilMatysik is still searching for something a client without a terminal can run.
Agent supply chain is now a daily worry. @fr0gger_ found 73 GitHub repos posing as agent tools that deliver malware, 51 still live, and @RayanSMX did the debunking work on viral charger scares that use Claude Code as the hook.
📡 Eco Products Radar
Eco Products Radar
Codex: the constant comparison and the second agent in most stacks
Grok Bot: the personal agent people are moving to, and complaining about limits on
Hermes: cited as the main OpenClaw destination, plus its import command
Living Brain: cross-model memory launch, mostly launch-day reposts
Aether: cloud runner on your existing subscription, 19.4 seconds to a build
Conductor: bought by Meta, Cloud goes read-only, alternatives circling
HyperFrames: video as HTML timelines, the default for agent-made video
Blender: the most common non-coding target, driven by Claude Code
Three.js: where agent-made games and 3D sites land
Postman: auto-installed plugins into agents, the week's trust story
Kylon: Grok Bot plans, Claude Code builds, Codex reviews
Worktrunk: one-line worktrees for parallel agents
Graphify and CodeGraph: code knowledge graphs for fewer tool calls
OpenViking: ByteDance's file-style agent memory
Codex: the constant comparison and the second agent in most stacks
Grok Bot: the personal agent people are moving to, and complaining about limits on
Hermes: cited as the main OpenClaw destination, plus its import command
Living Brain: cross-model memory launch, mostly launch-day reposts
Aether: cloud runner on your existing subscription, 19.4 seconds to a build
Conductor: bought by Meta, Cloud goes read-only, alternatives circling
HyperFrames: video as HTML timelines, the default for agent-made video
Blender: the most common non-coding target, driven by Claude Code
Three.js: where agent-made games and 3D sites land
Postman: auto-installed plugins into agents, the week's trust story
Kylon: Grok Bot plans, Claude Code builds, Codex reviews
Worktrunk: one-line worktrees for parallel agents
Graphify and CodeGraph: code knowledge graphs for fewer tool calls
OpenViking: ByteDance's file-style agent memory
Comments