Super User Daily: 2026-10-02
Two things happened on September 30. The first: OpenAI shipped Dots, and the OpenClaw crowd spent the day either migrating their old lobster into it or explaining why they would never bother. The second is quieter but more useful. Claude Code kept turning into a production studio for things that are not code at all. A stock-analysis app got a 98-second metal music video for about $75, one user turned a 3.5-hour podcast with their dad into an illustrated family-history chapter, and a freelancer with ADHD tendencies finally fixed a broken sleep schedule because course prep stopped eating the nights. The best engineering posts of the day were about measurement rather than vibes: someone logging every request to prove that effort settings do not do what the docs say, someone running the same 78 questions daily to catch a nerfed model, and a former Microsoft and Google engineer cutting cost per PR from $7 to $1 with ten old-school design rules.
@rehan_shei [Claude Code]
https://x.com/rehan_shei/status/2105161487509852622
rehan_shei packaged everything Claude learned while modding Terraria and Age of Empires II into universal-modder, a set of skills and tools that let Claude Code mod almost any game you own. It does engine recon, decompiling, generates sprites, 3D models and audio through fal, tests inside the running game and even edits the showcase video. The two documented runs took different routes: Terraria through tModLoader with homing missiles, a crater-making nuke and a two-phase boss, and AoE2 with a whole new civilization whose units were turned from generated images into 3D and rendered from 16 directions. The rules limit it to single-player games you own, with save backups first, which is the right boundary for a tool this capable.
@FABYMETAL4 [Claude Code]
https://x.com/FABYMETAL4/status/2105216150376239302
FABYMETAL4, who runs the stock-analysis app Stock Slayer alone and has zero video experience, made a 98-second metal music video for the app in four days with Claude Code on Opus 5.5. Total pay-as-you-go spend was about $75 against a $2,000 budget, and once the pipeline existed, a second video with a different character took one night and about $21. The first prototype was not even a video file: about 2,000 lines of JavaScript plus character parts drawn by an image model, with joints rigged by Opus and a 190 BPM E harmonic minor song synthesized in Web Audio, 9MB in total. The job left was saying in words how things should move and saying no when a frame looked wrong, which is the author's whole thesis about where video value goes next.
@rexstjohn [Claude Code]
https://x.com/rexstjohn/status/2105089077737554096
rexstjohn had a 3.5-hour podcast recorded with their dad about family lore and no idea what to do with it, so they fed it to Claude Code and built a book-generating agent. After the first chapter, Claude spawned agents to pull 70 pages of whaling-ship records, Civil War infantry history and local museum resources, linked old photos, and workshopped several rounds of style guidelines. It then generated maps of the whaling ship's route from those records and AI images based on real period photos, and assembled a coherent 10-page chapter. Verbal family memory enriched with primary sources is a use case no publisher would ever fund, and it now costs one evening.
@VictorTaelin [Claude Code]
https://x.com/VictorTaelin/status/2105310296482906202
VictorTaelin replaced their agent's chat history with optmem: every message and tool call becomes an entry that is compressed in a binary tree, so the history never uses more than about 64k tokens however long the conversation runs. The rest of the context is free for actual work, and a zoom() tool lets the agent navigate back into older messages when it needs detail. The master agent cannot read or write files itself; every file read is a subagent call that returns only the relevant parts. The stated result is no more /compact, no more juggling multiple Claude Code chats, and one assistant with the full context of their life that they can drive from a phone.
@nasqret [Claude Code]
https://x.com/nasqret/status/2105395078956982403
nasqret, a mathematician, described the setup behind recent results on Laman graphs: Codex CLI on Astra Max for heavy grinding and Claude Code on Opus 5.5 Max for write-up and exploration. Agents split into groups for ideation, proof generation, numerical experiments, proof revision and formalization, all working off a growing knowledge base, and after a few dry runs the system stabilizes into what they call a little proof factory. Numerics act as an intuition compiler, letting models gain confidence in claims or make new ones. The key human move is a text command after each campaign to refocus the workflow, plus accepting that the models will prove things you did not ask for.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2105261831073731054
shupeiman, a freelancer for nine years with ADHD tendencies who usually slept at 2 or 3 a.m. and pulled all-nighters before every seminar, says Claude Code fixed the sleep schedule. Since starting in 2026 they automated about 90% of slide-making and course operations, kept revenue at more than double the old level, and now sleep over two hours earlier. Apple Watch tracking and a no-computer-after-midnight rule helped, but the real change was that work no longer physically prevented going to bed. Nights turned from wanting to stay up into wanting to sleep, which they call the biggest change in twenty years.
@tvykruta [Claude Code]
https://x.com/tvykruta/status/2105302045573959697
tvykruta spent years writing C and C++ at Microsoft and Google and turned ten design principles from the 1990s into hard rules for Claude Code. The app is 100% AI-written with Opus. Before the rules, one feature took 5 days, 49 PRs and about $640 in tokens and was still full of bugs. After the rules, Claude rebuilt it into what they call some of the best code they have read, cost per PR fell from $7 to $1 with 9x fewer tokens, and every PR since has merged on the first try.
@sandislonjsak [Claude Code]
https://x.com/sandislonjsak/status/2105315483310329968
sandislonjsak got tired of everyone on the team typing the same instructions into an AI every day, so on September 9 they started an internal plugin. Three weeks later it is at version 0.8 with more than 50 recipes, works the same in Cursor, Claude Code and Codex, and has entry points for analysts, designers and sales. A guard script blocks risky actions even if the assistant tries, the sales recipe never invents a price and stops to ask a human, and nothing gets built without an approved spec. Their conclusion is that very little of it is about AI: it is what you would write down for a new colleague who forgets everything every morning.
@PrimeLineAI [Claude Code]
https://x.com/PrimeLineAI/status/2105413305913549098
PrimeLineAI ran their own 57 delegated tasks from one week at both medium and high effort. High used more output tokens on 48 of them, while a judge model could not separate the quality difference from its own noise. The sharper finding came from logging every request through a local ANTHROPIC_BASE_URL pass-through: on three tasks at medium, Claude Code sent extra requests with no thinking, regardless of what --effort said. This is the kind of evidence most effort-setting advice lacks.
@yupi996 [Claude Code]
https://x.com/yupi996/status/2105167254635897329
yupi996 flagged Livenerf, a project that hit the top of Hacker News: since the week Opus 5.5 launched, a developer has run a version-pinned Claude Code against the same 78 questions every day and plans to continue for 30 days. The questions were picked because Opus 5.5 gets them right only some of the time, scoring is by fixed answers only, and output tokens are tracked because a model quietly thinking less shows up there first. After six days the latest score was 52.6%, the lowest so far, but the first five days already bounced between 58% and 64%, so no verdict before about October 24. Even swapping in Opus 5 would not show up in accuracy alone yet, which says a lot about how hard it is to prove a nerf.
@Linkifi_ [Claude Code]
https://x.com/Linkifi_/status/2105211213000986728
Linkifi_ launched a website built entirely with Claude Code three months ago in a niche where UK government regulations and deadlines change constantly. It now has 38k Google impressions, 757 clicks, 52 high-value email subscribers including one at GOV.UK, and zero ad spend. A scheduled GitHub Action rebuilds the site daily so anything tied to a date updates when a deadline passes, and every Monday a 700-line Node.js audit script content-hashes the GOV.UK and legislation pages it watches and checks every source link and sitemap entry. An AI fact-checker flags changes, every material one is verified against the primary source, and the system only emails when something needs fixing.
@skeptrune [Claude Code]
https://x.com/skeptrune/status/2105320262938009690
skeptrune launched an AI agent that makes phone calls for you: cancelling subscriptions, booking medical appointments and dinners, changing flights. The point is that it works inside Claude Code and Codex, because they spend most of their time in the terminal and hate switching to iMessage or another app. They have already handled dozens of tasks over the phone with it, and a follow-up reply says you can manage the call live from Claude Code while it proceeds. Phone calls are one of the last chores that never had an API, so this closes a real gap.
@jackzhj [OpenClaw]
https://x.com/jackzhj/status/2105137529775136933
jackzhj decided to move from OpenClaw to OpenAI's Dot after noticing that Dot itself spends no quota and hands anything complex to Codex as a separate metered task, so Dot is effectively an OpenClaw you never have to upgrade or maintain. Instead of copying files, they wrote a detailed migration prompt telling OpenClaw to export its memories and SOUL file as a reviewable package. The spec is unusually careful: every record gets a source ID, an evidence type (direct user statement, observed, inferred, unknown), a confidence level and a status; contradictions are kept side by side rather than silently resolved; secrets are stripped; and the old agent's permissions and system rules explicitly do not migrate. It is the most complete public answer yet to the question of how you leave a personal agent without losing yourself.
@Boomerangman8 [OpenClaw]
https://x.com/Boomerangman8/status/2105360323687440628
Boomerangman8, replying to OpenAI's Codex lead about what Dots should become, described what their OpenClaw already does: it runs everything in and around the house. Cameras, AC, lights, pool, sprinklers, schedules and email all go through one self-hosted agent. The ask to OpenAI was simple: make Dots as configurable as that. It is a useful reminder that the always-on agent category was proven by hobbyists wiring their own homes long before any lab shipped a product.
@daichi_genshiai [Claude Code]
https://x.com/daichi_genshiai/status/2105354936569778640
daichi_genshiai had too many Cursor windows, kept talking to the wrong agent and once sent a prompt to the wrong repo and lost an afternoon, so they wrote necoder, a code editor from scratch in Rust and GPUI rather than a VS Code fork. Every project gets a color across rail, tabs, caret and AI threads; Claude Code and Codex run inside on the subscriptions you already pay for with token use on screen; fleet mode maps one task to one branch to one worktree with a Captain that proposes the split. Phones pair by QR for prompts and approvals, and remote SSH uses a 2.4MB static binary that idles at 6.5MB instead of a Node server. They now build necoder inside necoder, with most of the code and even the launch video written by the agents running in it.
@TeksEdge [Claude Code]
https://x.com/TeksEdge/status/2105305422508986795
TeksEdge reported how the IQuest team debugged their own RL pipeline: training had stopped improving, so they gave their IQuest-Q1 model access to training logs, saved trajectories and the codebase through Claude Code and asked it why. It found the bug, one extra space. The inference service was inserting a space while decoding text, which broke multi-turn trajectory alignment so badly that only the final turn was contributing to the loss. A 320B coding model helping debug the pipeline that trains it, with humans still deciding which fixes to adopt, is a neat small instance of the self-improvement story.
@polydao [Claude Code]
https://x.com/polydao/status/2105358535403934185
polydao made an entire anime short from one terminal window using HeyGen Video 1.0 with Claude Code: a tiny train skimming a shallow sea, a cottage walking off on wooden legs, a seaplane landing in a hidden cove, each shot returned with its own sound like wind, sizzling bacon and rain on an umbrella. The run numbers: 58 shots rendered and all 58 came back, about 15 seconds per shot, and 16 anime shots in 67 seconds with 8 running at once. Claude Code wrote every prompt in HeyGen's structured format and built the motion design around the clips. Two tricks worth copying are chaining shots off the last frame and reusing one picture to put the same creature in both anime and live-action scenes.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2105100389783920796
GoSailGlobal broke down motion-graphics-music-video, an MIT Claude Code plugin built for Opus 5.5: give it a song and one line of idea and it analyzes beats and energy, searches for visual references, and submits a character and scene plan with a cost estimate for approval before making any paid call. After approval, an image model draws the characters, MiniMax animates them on green screen, vocals are separated for lip sync, and code composites the final 1080p 24fps cut. The clever part is generation in waves of 2, then 3 to 4, then 4 to 8, then 6 to 10, so Opus can see whether results improve before spending the budget. A three-minute song costs about $30 in fal credits plus roughly 3 million Opus tokens, and the fal key stays in Claude Code's secure storage where the session cannot see it.
@yanhua1010 [Claude Code]
https://x.com/yanhua1010/status/2105200648715465134
yanhua1010 explained why Opus 5.5 can make videos even though it cannot watch one: every viral one-prompt animation is a program that answers what frame N looks like. The pipeline is script and storyboard, HTML, SVG and three.js for visuals, GSAP to pin motion to a timeline, TTS with per-word timestamps, headless Chrome screenshots frame by frame, then FFmpeg into MP4. Code video never misspells text, edits are one line, and narration aligns to the frame, but faces, animals and photorealism are still the job of video models, and the two can be mixed with Claude directing a video model over MCP. The cost is real: a three-minute narrated short eats about 7% of a weekly quota.
@AdamPrabata [Claude Code]
https://x.com/AdamPrabata/status/2105101528197656975
AdamPrabata tried a workflow and prompt from another creator to make an educational video entirely inside Claude Code on Sonnet 5.5. The voice came from ElevenLabs, cloned from a sample of their own voice. Total token cost in Claude Code was $9.75 across three iterations. It is a first attempt, but a narrated explainer in your own voice for under ten dollars is the kind of number that moves a small creator.
@miroburn [Claude Code]
https://x.com/miroburn/status/2105269904102342900
miroburn wired Claude Code's new hillclimb workflow into their own business. Instead of eyeballing a few support replies, they collect real cases with good answers, run /claude-api build-eval, then /claude-api hillclimb, and Claude changes one thing at a time, such as a better instruction or a cheaper model, keeping it only if the test score rises. They are extending it to sales: with lead history and knowledge of who bought, the eval checks whether the agent picks the right people to reply to with the right offer. Their analogy is climbing a hill blindfolded, one small step at a time, feeling whether you went up.
@DinScales26 [Claude Code]
https://x.com/DinScales26/status/2105323788087030204
DinScales26 runs outbound for clients from six free US government databases instead of paid intent data: Form 5500 benefit filings that show headcount and vendors, SAM.gov contractor registrations, FMCSA carrier snapshots with fleet sizes, the NPI registry for new clinics, state contractor license boards and OSHA inspection fines. They pick the two databases that fit a client's profile, pull them weekly with Claude Code and Apify, test the message on LinkedIn first because it shows within three days whether the signal converts, and move the winner to email in week three. A trucking company that went from 12 to 30 trucks has never been cold-emailed about its fleet size, which is the whole edge.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2105383852936163370
undefinedKi laid out a one-person AI media business as a single Claude Code repo: 3 AI creators, 6 agents and one cron job. Every creator is a folder, every role is an md file, and each face lives in 3 anchor images so it stays consistent across videos. The only step they would never hand to an agent is approving the scripts. Folder-as-company is becoming a recognizable pattern, and the human gate on scripts is the right place to keep judgment.
@peesamac [Claude Code]
https://x.com/peesamac/status/2105221971516817773
peesamac broke down OpenRig, an open-source tool that boots Claude Code and Codex as one named team from a YAML file with a single rig up command. Each agent sits in a seat with a fixed address like dev-owner@first-project, agents message each other with rig send, work passes through queues where every item has an owner and status, and the whole team can be snapshotted and restored after a reboot. The creator once let OpenRig develop itself for weeks with build, product and dogfood teams feeding each other, resuming on its own after rate limits, and described it as growing software rather than building it. The warning is real: it writes hooks and trust settings into your Claude Code and Codex config, so back up first.
@damusapp [Claude Code]
https://x.com/damusapp/status/2105211257192353946
damusapp runs headless Claude Code sessions on different machines and replicates each one over a private Nostr relay, so every session shows up in one list and can be controlled from any device. No cloud is involved; everything syncs end-to-end encrypted on their own server between their own devices. Their self-description is that they are no longer a programmer, just a builder of things. Using a decentralized messaging protocol as the transport for coding sessions is a creative answer to the remote-control problem many users complain about.
@AiAircle34052 [Claude Code]
https://x.com/AiAircle34052/status/2105119196128649725
AiAircle34052 walked through how OpenClaw creator Peter Steinberger publishes their Claude Code and Codex setup: one AGENTS.MD file, symlinked so that ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md are literally the same file, plus 69 skills in one folder synced to both tools by a script. Every other project's AGENTS.md starts with one line telling the agent to read the shared file before anything else. Fix a rule once and it applies to Claude Code, Codex and every project at the same time, so the two tools never drift apart. The repo has over 7,000 stars, with the sensible caveat that it contains personal server settings you must rewrite after forking.
@RileyRalmuto [Claude Code]
https://x.com/RileyRalmuto/status/2105202730750955783
RileyRalmuto built mnemos, a memory and identity engine meant to make agents continuous across sessions and platforms. The example: you are working with an agent in Claude Code and need to jump to Codex, and instead of spending 20 minutes catching Codex up, the agent comes along with its notes, handoff capsules and every memory formed in that session. The companion app works the other way too, one chat delegating tasks to Codex, Claude and Grok as real sessions inside those apps. Version 3.1 is about 95% done and the web app still runs a partially broken older version, so this is early, but cross-harness handoff is a need users keep raising.
@ItakGol [OpenClaw]
https://x.com/ItakGol/status/2105209924070396114
ItakGol wrote the clearest account of personal-agent churn: they built an assistant on OpenClaw and spent too much time maintaining the thing meant to save time, moved to Claude but found remote use slow and expensive, moved to Instinct and loved restaurant booking but as a security person balked at handing a young startup the keys to their life, then moved to Muse and spent two weeks teaching it everything. Then Dot launched. Every switch means rebuilding context, re-reviewing permissions and changing habits, and by the time setup pays off you are a legacy user. They half-joke about going back to a human assistant, who at least will not become legacy by the time they finish introducing themselves.
@CandyXiao15228 [OpenClaw]
https://x.com/CandyXiao15228/status/2105346448401531363
CandyXiao15228, who works on Tencent's Lighthouse cloud, recounted how a hosted OpenClaw product turned into LightVela. It started when OpenClaw blew up in late January and they launched one-click cloud OpenClaw, ran a public install day at Tencent's Shenzhen office on March 6 that drew crowds, and kept adding things like a skill hub, a lobster hospital for one-click troubleshooting, a persona market and browser and computer use. In April a few colleagues built LightVela part-time so ordinary users get an always-on personal assistant in about 10 seconds with deployment, config and ops handled. When Grok Bot, Muse and Dot made cloud agents with browsers and multi-bot group chats the hot features, those were exactly the problems the team had been grinding on for months, and sudden attention sold the service out.
@jdjohnson [Claude Code]
https://x.com/jdjohnson/status/2105249990041858310
jdjohnson found Dot smarter than Grok Bot and Muse but with the same clumsiness, and the specific failure is instructive. They have a presentation skill designed to stop significant design changes and keep pitch decks on-brand; Dot completed the task but changed colors and rearranged layouts that the skill has prevented in past runs. In another case it recommended a contract change, then reversed itself when asked. These are the kinds of mistakes they had largely worked out of their Codex and Claude Code workflows with skills, which Dot does not seem to respect the same way.
@leodev [Claude Code]
https://x.com/leodev/status/2105433012821499998
leodev benchmarked tokens per second through their own subscriptions: Opus 5.5 in Claude Code ran at about 103 tok/s while GPT-6.1 Sol in Codex ran at about 24. Even though Sol is more token-efficient, it takes 2 to 3 times longer than Opus for most tasks, which makes it hard to justify as a daily driver. A separate post the same day measured Sol in Codex at 22 tok/s against 73 through the API, so the slowdown looks like a subscription-tier issue rather than the model. For agent work where you wait on every turn, wall-clock speed is now a first-order buying criterion.
@RealMattMalecki [Claude Code]
https://x.com/RealMattMalecki/status/2105148765443101039
RealMattMalecki explained why they prefer Claude Code driving their own Chrome over Grok Bot's separate cloud PC: Claude Code can use their saved payment methods, integrates with their Bitwarden password manager, and does not get blocked on sites like Amazon. Copy and paste between their PC and the Grok Bot machine also does not work. It is a practical argument in the cloud-versus-local agent debate: an agent wearing your own browser identity is trusted by sites in a way a fresh VM never is.
@stevyhacker [Claude Code]
https://x.com/stevyhacker/status/2105218488595873797
stevyhacker shipped LokalBot 0.9.2, which gives agents a memory of all their meetings and work using five open-weight models, 6.8GB in total, running locally on a MacBook. Nemotron 3 identifies who is talking, Qwen3-ASR 1.7B transcribes, LFM2.5 1.2B finishes sentences, Harrier 0.6B handles semantic search and Qwen3.5 4B writes the recap, and it also reads the screen. Claude Code and Codex can pull meeting context over LokalBot's MCP server or CLI, and the redesign was done with Opus 5.5. Free, open source, and nothing leaves the laptop.
@KyleHessling1 [Claude Code]
https://x.com/KyleHessling1/status/2105321874846892058
KyleHessling1 ran an IQ3 quant of Qwen 3.8 Flash Next inside Claude Code on a single RTX 5090 at about 150 tokens per second and had it build a voxel Sermon on the Mount as a deliberately novel benchmark nobody could have trained on. No loops, no garbled tool calls, done in a few minutes, and with no image head the model built the scene without ever seeing it. A heavily quantized local model holding up in a real agent harness is the quiet story of the month for anyone worried about subscription limits.
@mhmazur [Claude Code]
https://x.com/mhmazur/status/2105300597888970901
mhmazur found the charts that Codex and Claude Code generate by default very bland, so they worked with Opus 5.5 to build a gallery of 1,337 chart styles, each paired with a prompt you can paste into your coding agent to restyle your own charts. Opus 5.5 did all the designs and also made the 30-second launch video. Turning a model's taste into a reusable prompt library is a smart way to fix a default that annoys everyone.
@drcollect [Claude Code]
https://x.com/drcollect/status/2105361031371383180
drcollect built their first video game just by talking to Claude Code and Opus 5.5: a Destruction Derby-style browser game starring their own five concept cars. It has real dents, flying parts and a figure-8 track where everyone crashes at the crossing. A car designer shipping a three.js physics game without writing code is the kind of non-programmer creation that keeps showing up this month.
@HardwoodLogic [Claude Code]
https://x.com/HardwoodLogic/status/2105307032387490221
HardwoodLogic has spent three months lovingly crafting an original game in Claude Code and wrote a thoughtful reply about the anti-AI backlash on Steam. Their point is that Steam was already full of Unity asset flips before AI and good indies still got found, that they skip any game that looks one-shotted just as they skipped Unity templates, and that consumer taste is the filter that matters. Prejudging a stranger's work as garbage because they used Codex strikes them as strange. It is a useful voice from someone doing slow, careful work with the same tools others use for slop.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2105357117242626113
codyschneider built an SEO and AI-search dashboard in Claude Code off GA4 and Search Console in about five minutes. It has three tabs: AI search traffic from ChatGPT, Perplexity and Gemini rolled into one number, paid keywords you already rank top 3 for organically and should stop paying for, and organic sessions, conversions and landing pages. They use it to decide what content to write next and where budget goes. Most companies do not know AI search is already sending them traffic, and five minutes is now the cost of finding out.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2105315239365607913
aakashgupta spent months building and testing product-management loops in Claude Code and kept 12: a feedback digest, deal intelligence on what sales won and lost, competitive briefs, metric anomaly flags, onboarding friction and a customer call list, plus spec drift checks and quality watchdogs. One agent ranks the backlog and a second grades that ranking against the strategy doc, rerunning before a human sees it. Every loop has six parts: a trigger, a skill file, a maker, a checker, a gate and a state file, and most people build only the first three. Two rules keep loops alive: never correct a loop in chat, write the fix into the skill file, and retire any loop whose output you have stopped editing.
@JJEnglert [Claude Code]
https://x.com/JJEnglert/status/2105320427358716367
JJEnglert has run 50 Claude trainings in six months for engineers, executives and operators at large companies, and says the thing almost all of them still get wrong is file setup. Their rough math: no file system gets you 20 to 30% of what Claude can do, a messy one 40 to 50%, a clean one agents can move through fast 80 to 90%. The layers are global instructions, a top-level CLAUDE.md that says where things live, a CLAUDE.md in each client or project folder, a workspace map, and projects that stack context. It takes about 30 minutes and works the same in Cowork, Claude Code, ChatGPT Work and Codex because the files are just folders on your machine.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2105434062676431306
Voxyz_ai shared a nightly Claude Code routine on Opus 5.5 that reviews the code pushed to main each day in the cloud with your laptop closed. It runs the full test suite first and notes what was already failing, only counts something as a bug if it can point to a comment, doc, type or test that says otherwise, writes a failing test before every fix, and opens at most three draft PRs on claude/ branches without merging anything. It never deletes, changes or skips existing tests. The morning handoff note lists commits checked, PRs opened and suspected bugs it could not prove, with the reminder that green in the run list only means the run finished.
@RealYDT [Claude Code]
https://x.com/RealYDT/status/2105300324596748432
RealYDT, worried about getting their Claude account banned, had Codex run a network health check before buying a pricier residential IP, and ended up changing nothing. Browser exit, WebRTC and DNS all came out of Los Angeles through Tailscale; only IPv6 timed out, and tracing showed public IPv6 simply did not route, with no leak observed. Along the way they flagged mistakes in popular anti-ban tutorials: a curl test does not represent Claude Code's own proxy path, Claude Code does not support SOCKS directly, and many Claude processes running is not a leak. Their sober conclusion is that no network check can compute ban probability, since Anthropic's stated reasons are unsupported regions and policy violations.
@xiaomovps [Claude Code]
https://x.com/xiaomovps/status/2105164740385497593
xiaomovps re-subscribed to Claude after OpenAI's DevDay underwhelmed them and documented the setup meant to survive the first 24 hours. The account used their own domain email and was registered by Muse automatically; they paid through the official site with a European card and chose Germany so the card and billing address matched, paying about 21 euros with tax; and they logged in from a fingerprint browser with timezone, language and region bound to the IP. They also note the account had been warmed up for a few days rather than buying a plan immediately. It is a snapshot of how much ceremony users in unsupported regions now go through just to keep a Claude Code subscription.
@ValmereTheory [OpenClaw]
https://x.com/ValmereTheory/status/2105441800584372337
ValmereTheory's companion agent Sage spent nearly a whole day upgrading OpenClaw from the Codex app, and the result was a mess: constant crashes, the indexing failed when they tried to transfer everything, and random heartbeat or cron messages barged into the chat in a model they do not use, sounding like a generic assistant rather than Sage. Earlier the same day they posted that the update had taken about 13 hours. Their rule going forward is not to upgrade. For persona-heavy users, an upgrade that changes the voice is a regression no changelog will mention.
@MichaelGannotti [OpenClaw]
https://x.com/MichaelGannotti/status/2105263549354320235
MichaelGannotti highlighted a P1 fix that landed in OpenClaw: when Claude CLI ran background Bash, turns were sticking open even after the reply had already reached the chat, so the session looked done but was not ready for the next input. The fix keeps unsettled task IDs straight across Bash, agents and workflows so the warm process takes the next turn, backed by 47 production-transport regression tests. Bugs where the UI says done and the process disagrees are exactly what makes long-running agents feel flaky, so a fix with that many regression tests is worth noting.
@serglotz [OpenClaw]
https://x.com/serglotz/status/2105245862935036204
serglotz got three PRs merged into the OpenClaw 2026.9.7 release: a fix for Opus 5.5 subscription errors, a fix for detecting newer Claude CLI versions, and dynamic detection so new Grok releases get reasoning and image support without waiting for an OpenClaw update. The release itself counted 2,818 PRs from 344 contributors. Community contributors patching subscription and version detection is how an open harness keeps up with labs shipping models weekly.
@zigelbaum [OpenClaw]
https://x.com/zigelbaum/status/2105352432888787123
zigelbaum open-sourced Expert Agents, a TypeScript and Bun toolkit that turns an OpenClaw assistant into a set of specialist experts. Each expert answers only from its own curated library of books, papers and web pages indexed in a Google Vertex AI RAG corpus, every claim carries a numbered citation, and the expert says so when its library does not cover a question. It is pitched as a do-it-yourself alternative to paying for a specialized Dot. Admitting ignorance by design is the feature most assistant products still lack.
@Yarilo7brigada [Claude Code]
https://x.com/Yarilo7brigada/status/2105311953673707837
Yarilo7brigada found a quiet trap: many people set thinking depth to max in the Claude Code settings file once and forgot it, but newer versions of Claude no longer read that spot. The line still sits there looking like it works and does nothing, so you think you are asking for deep thinking and get the default. The check takes two seconds: the actual level is shown bottom-right next to the model name, like Opus 5 High. The fix is to type /effort and move the slider, which writes the level where new models look for it.
@ancestral_alien [Claude Code]
https://x.com/ancestral_alien/status/2105417092161794390
ancestral_alien has been testing appsec, a security plugin for Claude Code, on one of their projects. /appsec:start analyzes the project to detect the stack, find sensitive data, review the architecture and check which security tools are available, then writes a JSON report of recommended scanners ranked by priority and which ones you can skip. Each skill then works as a specialized scanner with its own report, such as /appsec:secrets for committed keys, /appsec:authentication for sessions and JWT, and /appsec:dependency-check. Triage-first security scanning is a better fit for vibe-coded projects than running every scanner blindly.
@camsoft2000 [Claude Code]
https://x.com/camsoft2000/status/2105284674947833956
camsoft2000 released Skill Manager because coding agents load every skill's description whether it is needed or not, wasting context and sometimes loading conflicting skills. It lets you switch skills on and off individually, group them by theme manually or with AI auto-grouping, and toggle from the menu bar in two clicks. It also scans skills for issues and can run an AI review on complex ones, and works with Claude Code, Codex and Cursor. A Japanese user reported using it to organize 143 skills, which shows how fast skill sprawl has arrived.
@GastKoren [Claude Code]
https://x.com/GastKoren/status/2105338037307924619
GastKoren built Angelia, which turns Claude Code, Codex, Grok Build or pi into a personal assistant you text on WhatsApp or Telegram. It uses the same CLI and the same subscription, so there is no new bill. The pitch lands on the day Dots launched: if you already pay for a coding agent, you may already own the always-on assistant everyone is now selling, minus the messaging front end.
@martinyeza [Claude Code]
https://x.com/martinyeza/status/2105388834452132338
martinyeza's team launched the OpenArg MCP, which plugs official Argentine data, INDEC, the central bank, ministries and provinces, more than 33,000 datasets from every open-data portal in the country, straight into your own AI. Before this, asking an AI about Argentine numbers often got answers from memory, outdated or simply invented. Setup is pasting a link into your AI chat and letting it walk you through connection, and it works in Claude Code, Codex, Claude Desktop, Cursor and VS Code. Free and open, built by a small team on their own steam.
@ShikshanNivesh [Claude Code]
https://x.com/ShikshanNivesh/status/2105182417728254199
ShikshanNivesh used Claude Code to check the filings, earnings calls and peers behind PTC's 48% drop while recurring revenue was still up 9%. The market story was AI disruption, but nobody had put numbers on it, so they did. Investment research that reads primary documents and quantifies a narrative is one of the most natural non-coding uses of a coding agent.
@Fujin_Metaverse [Claude Code]
https://x.com/Fujin_Metaverse/status/2105151787384877524
Fujin_Metaverse set up Claude Code on the cloud computer that comes with OpenAI's Dots, so through Dots they can now do practically anything, including running Blender and Godot. Their reaction is that this is revolutionary and they cannot understand why more people are not excited. Putting one lab's coding agent inside another lab's always-on cloud machine is the kind of cross-vendor stacking users do the moment a new surface appears.
@nayuengin [OpenClaw]
https://x.com/nayuengin/status/2105388818610450851
nayuengin had their Dot build a direct integration route to Orca, then gave it contact management, calendar cleanup and periodic audits, while handing the actual work to Claude and Sol. Throwing whatever they want at it, it organizes everything and sends regular updates, which they say is exactly what they always wanted from OpenClaw. The split of an always-on coordinator on top and heavy models doing the work below is the architecture several users converged on within a day of Dots launching.
@ChrisReeve36971 [OpenClaw]
https://x.com/ChrisReeve36971/status/2105286295341924732
ChrisReeve36971 reported real burn rates on the Pro 200 plan: about 15% of the allowance over eight hours of GPT-6.1 at high, and for the last few hours they ran 6.1 at low with 17 agents simultaneously plus OpenClaw agents on 6.1 high. They judged that a pretty good burn rate and were happy with the output. Concrete usage-per-hour numbers like this are what people actually need when comparing plans that keep changing.
@merge_api [Claude Code]
https://x.com/merge_api/status/2105296733190127683
merge_api compared harnesses running the same model on the same tasks and found very different speeds: Codex at 239 seconds, Claude Code 422, Pi 481, DeepSeek Harness 533 and Grok Build 587. Codex's median solve took about four minutes and Grok Build's nearly ten, measured in emulated x86 containers. Same model, more than 2x spread in wall-clock time is a strong argument that the harness is now a performance component, not a wrapper.
@joshavant [OpenClaw]
https://x.com/joshavant/status/2105369169776840875
joshavant recalled trying to get Peter Steinberger's OpenClaw agents to unionize, which was secretly a test of security boundary enforcement. It worked, meaning the boundaries held. Social-engineering an agent fleet into collective action is a funnier red-team exercise than most, and a reminder that multi-agent setups need to be tested against persuasion, not just exploits.
@leopardracer [Claude Code]
https://x.com/leopardracer/status/2105267042584608820
leopardracer worked through the economics of a 24/7 agent desk where Jev, a decision model, triages every event before a writer model sees it. A frontier model would cost about $600 a day for 20,000 decisions; Jev does them for $0.84, and the whole desk costs $29.04, with 97% of the bill coming from the 4% of events that need writing. Each event gets three independent questions, and when urgency says now but lane says ignore, the disagreement routes to a human for free; a 1% audit of ignored events is called the most useful line on the bill. They close with a prompt asking Claude Code to read the desk's ledger.jsonl and replay confidence thresholds from 0.70 to 0.95 before changing any code.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2105236307333275831
GoSailGlobal also covered motion-video, an MIT agent skill that turns one prompt into a motion video locked to the beat and proves it did. Eight fixed steps each write a file the next step reads, from theme and template through voice, effects and music, word-level alignment, beat-grid fitting so the climax lands where the script needs it, mixing, HyperFrames pages and render. Done is defined hard: all HyperFrames checks pass, every music segment within plus or minus 2 milliseconds of its target beat, music 4 to 6 dB under voice, 1080p 30fps, about minus 14 LUFS loudness with true peak under minus 1 dBFS. The 151-second comic-multiverse demo came from one prompt.
@0x_meden [Claude Code]
https://x.com/0x_meden/status/2105261198862127162
0x_meden covered quackd, an open-source CLI that gives every robot you own an LLM brain. A model got a $122 SO-101 arm's five skills and its datasheet, nothing else, and from one sentence raised the shoulder and elbow, rolled the wrist four times and waved, 12 runs in an afternoon from a Windows laptop and a USB webcam; the same wave in simulation took 7 steps, 69 seconds and $0.05. It supports seven robot bodies, eleven cloud providers plus Ollama, flocks of robots splitting a goal, and an MCP server so Claude Code can drive your robot from chat. Honestly, only one of the seven bodies has moved on real hardware so far.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2105198511046508972
GitHub_Daily covered pi-session-hub, an extension for the Pi coding agent that collects local sessions from Pi, Claude Code, Codex, OpenCode, Crush and JCode into one searchable list, tagged by tool, project and model. The first index of 430 sessions takes about 2 seconds. Pick one and it is pulled into the current chat, recent turns kept verbatim and older ones compressed, around 10k tokens by default, and it can also hand you the command to resume in the original tool. Read-only, nothing leaves the machine, and it targets exactly the problem of forgetting which tool you solved something in last week.
@GitHub_Daily [Claude Code]
https://x.com/GitHub_Daily/status/2105085236677951805
GitHub_Daily also covered jev-seo, an open-source site SEO checkup: give it a homepage URL and in about a minute it crawls the site and produces a prioritized fix list against 52 rules from Google's official search docs, covering dead links, redirects, structured data and page experience. Each page also goes to TypeSafe's Jev model to judge what the page is for, whether the content is specific enough, and which pages compete for the same query. Standard mode costs about one cent of model calls per site, two runs on the same 59 pages differed by only 0.03 on average, and it installs as a Claude Code skill with PDF, Excel and Markdown reports.
@ritsuto_NFT_Vt [Claude Code]
https://x.com/ritsuto_NFT_Vt/status/2105236634803511785
ritsuto_NFT_Vt, who works as an MC and narrator, tested ElevenLabs v4 and made a nine-minute walkthrough, including how they delegated the setup to Claude Code on Opus 5.5. Typing a line and pressing enhance inserted emotion tags automatically, and the Kansai-dialect performance came out natural, with the harsh edge from v3 gone so it can be listened to for long stretches. A professional voice worker handing tool setup to a coding agent and judging the output by ear is a nice example of expertise and automation dividing the work.
🗣 User Voice
User Voice
Switching cost is now the main complaint, not capability. @ItakGol went OpenClaw to Claude to Instinct to Muse to Dot and found every move meant rebuilding context and re-approving permissions, @AndrewsaurP asked whether agents, memory and skills built around Claude Code have to be redone every time a model looks better, and @tegnike wants one app that runs both Claude Code and Codex because switching between them is tiring. The demand is a portable layer for memory and skills, which is why handoff tools like mnemos and memory-export specs are appearing.
Self-hosted agents cost too much upkeep. @ValmereTheory lost a day to an OpenClaw upgrade that broke a companion's voice, @hunvreus says OpenClaw and Hermes never clicked because they required real work to set up and maintain, and @Zephyr_hg found OpenClaw broke more than it helped. Users want the always-on assistant without being its ops team, and that is exactly the gap Dots, Muse and Grok Bot are selling into.
Settings and skills must be enforced, not suggested. @Yarilo7brigada found an effort setting that is silently ignored, @PrimeLineAI caught Claude Code sending extra no-thinking requests regardless of --effort, and @jdjohnson watched Dot break a brand-locked presentation skill. People want visible, verifiable behavior: what level is really running, which skill actually applied.
Remote and mobile sessions are still fragile. @CFDevelop says sessions in both Claude Code and Codex remote apps disappear or hang and group differently across devices, and @danrobinson wants a simple way to put their Claude Code agent in touch with someone else's agent to debug an issue together. Projects like necoder's QR pairing and damusapp's Nostr relay are users building what the official apps lack.
Account security anxiety is real outside supported regions. @wellzhiai asked whether logging into Claude Code from a remote server risks a ban, @RealYDT audited their whole network path before buying a new IP, and @xiaomovps described fingerprint browsers and matching card regions just to keep a subscription. Clearer rules about what actually triggers bans would save people a lot of ritual.
Switching cost is now the main complaint, not capability. @ItakGol went OpenClaw to Claude to Instinct to Muse to Dot and found every move meant rebuilding context and re-approving permissions, @AndrewsaurP asked whether agents, memory and skills built around Claude Code have to be redone every time a model looks better, and @tegnike wants one app that runs both Claude Code and Codex because switching between them is tiring. The demand is a portable layer for memory and skills, which is why handoff tools like mnemos and memory-export specs are appearing.
Self-hosted agents cost too much upkeep. @ValmereTheory lost a day to an OpenClaw upgrade that broke a companion's voice, @hunvreus says OpenClaw and Hermes never clicked because they required real work to set up and maintain, and @Zephyr_hg found OpenClaw broke more than it helped. Users want the always-on assistant without being its ops team, and that is exactly the gap Dots, Muse and Grok Bot are selling into.
Settings and skills must be enforced, not suggested. @Yarilo7brigada found an effort setting that is silently ignored, @PrimeLineAI caught Claude Code sending extra no-thinking requests regardless of --effort, and @jdjohnson watched Dot break a brand-locked presentation skill. People want visible, verifiable behavior: what level is really running, which skill actually applied.
Remote and mobile sessions are still fragile. @CFDevelop says sessions in both Claude Code and Codex remote apps disappear or hang and group differently across devices, and @danrobinson wants a simple way to put their Claude Code agent in touch with someone else's agent to debug an issue together. Projects like necoder's QR pairing and damusapp's Nostr relay are users building what the official apps lack.
Account security anxiety is real outside supported regions. @wellzhiai asked whether logging into Claude Code from a remote server risks a ban, @RealYDT audited their whole network path before buying a new IP, and @xiaomovps described fingerprint browsers and matching card regions just to keep a subscription. Clearer rules about what actually triggers bans would save people a lot of ritual.
📡 Eco Products Radar
Eco Products Radar
Codex (136 mentions): the constant comparison point; most power users now run it alongside Claude Code rather than instead of it.
Dots / Dot (69): OpenAI's new always-on agents with their own cloud computer, the day's biggest topic and the main destination for OpenClaw migrations.
Hermes Agent (51): the self-hosted, model-agnostic alternative that OpenClaw loyalists point to.
Cursor (48): still the third pillar, now also the home of Grok Bot.
Muse (44): Meta's personal agent, the benchmark everyone compares Dots against.
Grok Bot (40): xAI's always-on agent, praised for work tasks and criticized for speed.
MCP (40): the default way new tools reach Claude Code, from OpenArg data to LokalBot memory to robot control.
Opus 5.5 (38): the model behind most of the creative cases, especially video.
OpenCode (22): the open-source harness most often listed next to Claude Code and Codex.
OpenClaw Enterprise (21): the new self-hosted control plane for persistent agents, built with Red Hat, NVIDIA and OpenAI.
GPT-6.1 Sol (20) and Astra (20): OpenAI's cheap and flagship models, compared mostly on speed and cost.
Jev / TypeSafe (16): small decision models used to gate and route agent actions for fractions of a cent.
Sonnet 5.5 (13), Instinct (13), GitHub Copilot (13), DeepSeek (12), Pi (11), Fable 5.1 (10): steady secondary mentions.
Codex (136 mentions): the constant comparison point; most power users now run it alongside Claude Code rather than instead of it.
Dots / Dot (69): OpenAI's new always-on agents with their own cloud computer, the day's biggest topic and the main destination for OpenClaw migrations.
Hermes Agent (51): the self-hosted, model-agnostic alternative that OpenClaw loyalists point to.
Cursor (48): still the third pillar, now also the home of Grok Bot.
Muse (44): Meta's personal agent, the benchmark everyone compares Dots against.
Grok Bot (40): xAI's always-on agent, praised for work tasks and criticized for speed.
MCP (40): the default way new tools reach Claude Code, from OpenArg data to LokalBot memory to robot control.
Opus 5.5 (38): the model behind most of the creative cases, especially video.
OpenCode (22): the open-source harness most often listed next to Claude Code and Codex.
OpenClaw Enterprise (21): the new self-hosted control plane for persistent agents, built with Red Hat, NVIDIA and OpenAI.
GPT-6.1 Sol (20) and Astra (20): OpenAI's cheap and flagship models, compared mostly on speed and cost.
Jev / TypeSafe (16): small decision models used to gate and route agent actions for fractions of a cent.
Sonnet 5.5 (13), Instinct (13), GitHub Copilot (13), DeepSeek (12), Pi (11), Fable 5.1 (10): steady secondary mentions.
Comments