Super User Daily: 2026-11-10
The phone became the remote control for Claude Code this weekend. One person watched an agent delete 170 dead Alexa devices by tapping through a mirrored iPhone, another directed a full video edit from the back of an Uber while a Mac mini at home rented GPUs by the minute, and a PM keeps five or six agents running for a day at a time and reads a single status page instead of the chat. The second thread is people who finally measured instead of guessing. One team ran Google's new spec-review plugin against plain prompting and got the same test score for 2.2 times the bill, and another person checked ten ChatGPT app ideas against verified revenue and found only two with any money behind them. A lot of the best work came from people who don't write code for a living: a novelist turning a chapter into a trailer, an IT consulting manager who handed off a dozen recurring jobs, and a first-time game builder with 11,700 lines of C# they never typed. The loudest warning in the replies is about Don't Ask Again, which writes approved commands, passwords included, into a settings file that keeps ending up on public GitHub.
@marcos_placona [Claude Code]
https://x.com/marcos_placona/status/2119544264686223516
@marcos_placona moved Hue bulbs to Zigbee2MQTT and was left with about 170 dead Alexa devices, and the new iOS Alexa app has no bulk delete, so each one meant four taps. The fix was to install phone-harness from the Hermes plugin catalog, pair iPhone Mirroring on a Mac mini, grant Accessibility and Screen Recording, and let Claude Code, driven through T3, screenshot the phone, read the text and tap. It was not smooth: unlocking the phone to check progress killed the mirroring, and the first script scrolled up a pull-up sheet that closes when you scroll up. What made it work were boring rules, starting with only touching rows marked unresponsive, which is the real lesson for anyone pointing an agent at a phone UI.
@stas_sorokin_ [Claude Code]
https://x.com/stas_sorokin_/status/2119503454792020344
@stas_sorokin_ measured Google's new Elephant-Goldfish spec check instead of just sharing it: 8 functions from vague one-line specs, Opus 5.5 in Claude Code, hidden tests, 4 runs per arm. Going straight to code passed 24 of 32 for $0.28 in 70 seconds, and running the goldfish review first also passed 24 of 32 but cost $0.63 and took 156 seconds. The goldfish did find real gaps, like a bare 5 passing as a duration, but then answered its own questions and every fix introduced a new break. The takeaway is sharp and cheap to apply: keep the reviewer's questions, but answer them yourself.
@pap3rw8 [Claude Code]
https://x.com/pap3rw8/status/2119250505889206676
@pap3rw8, who calls themself a certified AI hater, admitted Claude Code solved a hard, undocumented problem in under an hour that had been dragging on for nearly a year. The post itself is one sentence, but it drew close to 9,000 likes, mostly from people who recognized the shape of the story. The detail that matters is undocumented: this is the class of problem where search fails and forum threads go cold. It is also a useful reminder that the most persuasive evidence for these tools tends to come from the people least inclined to praise them.
@rileybrown [Claude Code]
https://x.com/rileybrown/status/2119229301392392221
@rileybrown sat in an Uber and had Claude Duet use its own VM to gather assets for a video idea, then had Duet spin up a Claude Code session on a Mac mini at home to edit the video with Blender for a few visuals. To render faster, Claude Code rented compute by the minute on RunPod for the Blender scenes, finished the render on the Mac mini, uploaded it to a video version control app and sent a review link back through Duet. Five rounds of back and forth took about 30 minutes, all of it by chatting with an agent. The part worth noting is that the agent made the buy decision on compute itself, which is exactly where a spend cap would belong.
@tetsuoai [Claude Code]
https://x.com/tetsuoai/status/2119341048367665201
@tetsuoai released a 5:26 music video built around the Attention Is All You Need paper, taking 26 hours 51 minutes from first lyrics to final file. The run used 10.1 billion tokens, 131 agents and 23,354 model calls, with Claude Opus 5.5 and Claude Code alongside Grok, Qwen-Image, Seedance 2.5, MiniMax H3 and a 51M-parameter model trained from scratch just for the film. Blender rendered 11 3D sequences, ffmpeg with NVENC encoded 84 shots and 573 sung words on screen, and the hardware was an RTX PRO 6000 plus an RTX 5090 with 1 TB of RAM. The piano was written by hand, which is the pattern worth keeping: the human piece goes first and the pipeline is built around it.
@marcelkargul [Claude Code]
https://x.com/marcelkargul/status/2119406341206724799
@marcelkargul broke down a 31-second icon animation made entirely in code. Six storyboard frames in Figma went through Figma MCP and Claude Code to SVG, a script split the finished icon into 15 layers, and Remotion with React animated each layer so lines draw themselves, labels type in with a caret and the fill resolves out of dots, all on a tilted CSS 3D plane with a light pool that follows the action. Revisions happened in Claude Code with Chrome previews, and the music was generated in Python and NumPy from the film's own timings, rising from 100 to 138 BPM. The render took 4 minutes at 1080p 60fps, which shows how far a design team can now go without opening After Effects.
@noriyang_crypt [Claude Code]
https://x.com/noriyang_crypt/status/2119251126734467247
@noriyang_crypt turned the opening of chapter 8 of a web novel on Kakuyomu into a short trailer, using the text's own lines as subtitles and narration. The chapter went into Claude Code, which split it into four shots and wrote the prompts, and the shots were generated with Seedance 2.5 through Renoise. In the first version the heroine held the quill in the left hand, so the shot was rewritten to a sleeping pose where the hand is hidden. A follow-up post spells out the habit behind it: every fix gets written back into the next prompt, so the same mistake does not happen twice.
@masashi20555444 [Claude Code]
https://x.com/masashi20555444/status/2119539205001105694
@masashi20555444 is a manager at an IT consulting firm with more than 20 years in, and also writes books on the side, and has been moving the I'm-the-only-one-who-can-do-this work into Claude Code one task at a time. The day job now runs slide decks, meeting minutes, reviews of other people's documents, a task dashboard, issue-to-countermeasure write-ups and the weekly report through it. The side business runs Kindle manuscripts, Amazon listing images and copy, category research, Udemy course videos, scheduled posts for X, note and a newsletter, and a morning goal check. The publish button, the approval of each post and any change to strategy stay with the human, which is the line most people skip drawing.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2119575501115162870
@PawelHuryn is a PM who runs five or six Claude Code agents in parallel, some for more than 24 hours, without writing code. The method is to delegate outcomes rather than steps, with the goal, how success gets measured and what agents may not do on their own, and then leave them alone for 12, 18 or 24 hours. The key habit is not reading the chat: each agent keeps a flight-deck artifact of what is done, what is in flight and what is next, decisions get made inside that file, and Telegram only pings when an agent is blocked. The /flight-deck skill is free on GitHub, and it answers the most common complaint about long runs, which is that a 24-hour chat is unreadable.
@_LouiePeters [Claude Code]
https://x.com/_LouiePeters/status/2119389910381535437
@_LouiePeters has stopped starting fresh threads and doing handovers between them, and instead keeps one permanent thread per project that manages hundreds of subagents, some permanent and some thrown away. With context compaction plus custom memory files and scratchpad notes, some of those threads have passed 10 billion tokens including subagents without a major breakdown. The post is framed as a question to other users of Claude Code, Codex and T3 Code about whether context rot still bites them. It cuts against the common advice to reset early and often, and the replies are worth watching for where that advice still holds.
@ds_nakajima [Claude Code]
https://x.com/ds_nakajima/status/2119304507725353124
@ds_nakajima found that managing every session personally turned the human into the bottleneck, so session management moved to an AI. A single AI managing everything got messy too, with messages tangling together, so the setup was rebuilt around specialist bots. Grok Bot is operated from a phone or a MacBook Air, while the actual work runs on a Mac mini, which lowers the load and keeps the connection stable. The result is that Claude Code and Orca are now almost never touched directly, which is a concrete version of the orchestrator-over-agents pattern many people are only theorizing about.
@JustLikeBit [Claude Code]
https://x.com/JustLikeBit/status/2119406082808279144
@JustLikeBit shared a setup that chains Grok Bot, Claude, a free Oracle cloud machine and a home Mac into one system. Grok Bot is the commander with its own cloud computer and bots for tech, office and media work, Claude Code sits inside that cloud computer and only reviews code, the Oracle box with 4 cores and 24 GB runs 7 tools and acts as the relay home, and the Mac does the heavy lifting with Blender, ComfyUI and an Android emulator. One sentence to the bot travels all the way back to the Mac without opening any port, and the result was a full set of 3D renders for an apartment layout. The monthly bill is about $20 for Claude and $30 for Grok Bot, with the Oracle machine free.
@devxops [Claude Code]
https://x.com/devxops/status/2119449578013643244
@devxops turned an idle PC into a cloud environment for Claude Code: a small Debian VM on Proxmox with 4 cores and 8 GB of RAM. One always-on Claude per repo runs through claude remote-control as a systemd template, so each session can be driven from the Claude app or the web. CLAUDE.md, settings, skills, notes, specs and plans all live in one private GitHub repo that a bootstrap script links into ~/.claude on any machine, and an auto-sync script commits, rebases and pushes on session start and stop hooks plus every two minutes, stopping to ask on a conflict. The payoff listed is that the agent cannot damage the Mac, the laptop lid can stay closed, and work continues from a phone.
@Marko_Poly [Claude Code]
https://x.com/Marko_Poly/status/2119465628700405970
@Marko_Poly relayed a case where a user running 60 Grok bots on the $30 plan used up the weekly allowance in about 2.5 days, and every bot stopped for 40 hours. One week of logs, October 13 to 20, went into Claude Code for a breakdown: 93.2% of tokens were the bots re-reading their own chat, 0.64% was what the bots actually said and 0.37% was new prompts. There were 1,540 scheduled wake-ups a week, and 43 of 45 watchdog runs found nothing. Using Claude Code as the auditor for another agent's bill is the part to copy, because you pay for the reading, not the answer.
@0xMfox [Claude Code]
https://x.com/0xMfox/status/2119517866550595939
@0xMfox asked ChatGPT for 10 apps a vibe coder could build, got told every one was huge, and then had Claude Code check them against TrustMRR, which reads revenue straight from founders' payment accounts. Only 2 of the 10 had real money behind them, an astrology app at $46 and an air fryer app at $88. What actually earned in the last 30 days was boring: Comp AI for compliance paperwork at $898K, Cometly for ad attribution at $212K, a resume builder at $208K and software for disability care providers at $200K. Two of every three businesses on TrustMRR sit under $1K, which makes the check worth running before building anything.
@Pluvio9yte [Claude Code]
https://x.com/Pluvio9yte/status/2119258980199682161
@Pluvio9yte benchmarked two computer-use tools, agent-desktop and Cua, for use with Claude Code, which has weak built-in computer control. Each task ran 10 to 20 times, a second tool verified every claimed success, and the person kept using the computer normally to see whether either one stole the mouse. Both went 20 for 20 toggling System Settings, 10 for 10 on a calculator task and 10 for 10 typing mixed Chinese and English in the background, but agent-desktop was faster at about 1.4 seconds per calculator round against 5 for Cua. Neither could read the text inside an Electron app like Cursor, which is a limit worth knowing before planning any desktop automation around them.
@0xKense1 [Claude Code]
https://x.com/0xKense1/status/2119547082549744043
@0xKense1 is building a Zelda-like forest game with a team of AI agents and logged the pace: concept art late on day one, a first playable build on day two, and a second version with a 28-meter titan boss on the night of day two. That adds up to 11,700 lines of C# and shaders, none written by hand. The stack is Unity, Claude Code on Opus 5.5, Tripo for 3D models and Blender. The claim that one person can now build a game is common, but a line count plus a boss size plus a timeline makes this one checkable.
@dogahack [Claude Code]
https://x.com/dogahack/status/2119452635061268786
@dogahack summarized a solo developer at Whale Dev Studios who built an anime-style FPS with a team of AI agents and shipped a free Steam demo in under a month. The first commit was September 29, the demo went live on October 23, and the repo had 4,558 commits by then, with the developer saying almost no code was written by hand. The tools were Godot, Claude Code and ElevenLabs, and the developer published the build process and the instruction files given to the agents in a thread. Commit count is the useful number here, because it shows the agents were working in small, reviewable steps rather than one giant generation.
@whothefukboris [Claude Code]
https://x.com/whothefukboris/status/2119506519855075765
@whothefukboris made an animated story in the style of Roblox shorts in 30 minutes with Claude Opus for $0, after almost a week of trying to get the pipeline right, and gave the whole prompt away. The prompt builds a 1080x1920 60fps short entirely in three.js with no Roblox assets and no AI images or video, rendered frame by frame in headless Chrome. It specifies R6-style blocky avatars with exact part sizes, faces drawn as flat sticker decals for seven expressions plus talking versions, costumes per role and a sunny plastic toy town with specific sky colors. It is a good example of the real cost of these one-prompt videos: the prompt is the product, and it took a week to write.
@mattworkman [Claude Code]
https://x.com/mattworkman/status/2119541684195221655
@mattworkman prototyped a Mimic treasure chest mechanic inspired by Frieren for the iPhone Duo and is considering building a small 3D mobile game around it. It was built in Claude Code with Opus 5.5 on High, using the fal MCP to generate assets. The post is short, but it shows a pattern that keeps coming up: people prototype one mechanic on a new device form factor first, and only then decide whether a game is worth building around it. A dual-screen phone is exactly the kind of hardware where a cheap mechanic prototype tells you more than a design doc.
@froessell [Claude Code]
https://x.com/froessell/status/2119453901686739134
@froessell got hooked on Yet Another Zombie Survivors on the Switch and built a first prototype of a similar game in a personal style, made in Claude Code from a phone. There is no laptop in this story, which is the point: remote control of a session has made the phone a viable place to start a game, not just to check on one. The post is a prototype rather than a finished product, but it is first-hand and comes with a video. It belongs with a growing set of weekend projects where the trigger is simply playing something and wanting your own version.
@BeyondTheBoxCN [Claude Code]
https://x.com/BeyondTheBoxCN/status/2119429774330261513
@BeyondTheBoxCN highlighted Simran, who recreated a perspective-illusion effect inspired by the Japanese creator @garments__ using Claude Code, Three.js and GLSL shaders, and then packaged the whole method as an open-source skill called perspective-illusion. The trick is to use an extreme telephoto lens to flatten a 3D street into what looks like a flat illustration, then move the camera and change focal length so the hidden depth opens up. The skill bundles scene templates such as streets, a supermarket and a rainy crossing, plus camera moves, shaders, video rendering and an image-check step for Claude Code to follow. Packaging a visual technique as a skill is the interesting move, because the next person gets the workflow, not just the clip.
@aifindr [Claude Code]
https://x.com/aifindr/status/2119564775420047743
@aifindr had Claude Opus 5.5 build a storm inside a snow globe in one HTML file with three.js, from a single prompt with no edits, in 25 minutes for $5.65. The prompt specifies a lighthouse on a rock, water that sloshes when the globe shakes, a boat tossed by waves, rain, lightning and a sweeping beam, plus a title, a control panel and an auto-shake every few seconds. The important detail is that it drives everything from window.renderFrame at 30fps with a fixed duration and a ready flag, which makes the scene renderable to video deterministically. The honest caveat is included too: $5.65 is API list price for one run in Claude Code, mostly output and cache, and your number will differ.
@WoodsDolls [Claude Code]
https://x.com/WoodsDolls/status/2119366552952987831
@WoodsDolls deployed about 25 landing pages automatically with Claude Code, with no brief and no complicated strategy, each one generated from a keyword and a title, and they are already starting to rank. The site now gets 500 to 1,000 returning visitors a day, with some users staying up to an hour and a half. The thesis is that SEO has not changed in ten years, business keywords plus content plus backlinks, but retention has never mattered more. The project is expected to pass 10K euros in revenue in its first month, with customers outside France who found it on their own.
@AtownBrown [Claude Code]
https://x.com/AtownBrown/status/2119260905988829220
@AtownBrown loaded the Porkbun and Cloudflare MCP servers into Claude Code and in about 20 minutes bought 15 exact-match .com domains for $177, built 10 separate guide sites each written for its own keyword, and pointed the other 5 domains at the closest site. Every site compares one investing app with five competitors, with sources, says who published it and links back to the app stores, all hosted on Cloudflare with HTTPS. The aim is to rank both on Google and inside answers from ChatGPT, Perplexity and Gemini. The next step is analytics to see whether any of it drives downloads, which is the only number that will settle whether this works.
@amitmuflis [Claude Code]
https://x.com/amitmuflis/status/2119418410396451137
@amitmuflis built 7 free agents in Claude Code for job hunting, aimed at a specific failure of AI resume tools: they make you sound perfect until the interviewer asks about a skill you never had. The agents decide whether a role is worth applying to, tailor the application, and reject anything the real resume cannot back up. That last rule is the useful one, because it turns the agent into a check on exaggeration rather than a generator of it. It is a small project, but it shows people building agents whose main job is to say no.
@GoldshteinAnton [Claude Code]
https://x.com/GoldshteinAnton/status/2119508844510581079
@GoldshteinAnton built a free, local tool that works like Google Analytics for Claude Code sessions and projects, after losing track of a codebase at 25 sessions a day. It draws a live map of the project's entities, such as users, subscriptions and payments, and which files implement them, and it flags stale parts that keep polluting context, along with how many tokens MCP servers, CLAUDE.md and memory add to every turn. It also ranks bad habits in dollars, such as resuming after the cache expired or using Opus for questions Sonnet would answer, alerts when two sessions edit the same file, and logs every prompt with the files changed, whether tests passed and what it cost. That covers most of the questions people keep asking about where their limits went.
@stretchcloud [Claude Code]
https://x.com/stretchcloud/status/2119340898773262612
@stretchcloud argued that a CLAUDE.md line saying never touch production is a request, not a guardrail, and that the popular seven-layer fix assumes one agent, one harness and one settings file. That breaks once Claude Code runs next to Codex and Aider, each with its own permission model, so Campfire was built as one shared approval layer for Claude Code, Codex, Goose, Aider and OpenHands. Permission requests are decided by majority among whoever is watching, any single deny blocks the action, and a 30-second timer covers the case where nobody is around. Every session replays afterward with each tool call logged by agent and backend, and agent races run the same task in isolated git worktrees so a bad command never reaches the real branch.
@andyramosdesign [Claude Code]
https://x.com/andyramosdesign/status/2119470387977158881
@andyramosdesign added a portfolio case study about a Claude Code plugin built to take Figma designs into Optimizely without rebuilding every test by hand. The write-up covers the workflow and what feedback from the team's test engineer changed in it. A designer building the bridge between a design tool and an experimentation platform is exactly the kind of non-engineer tooling the agent era enables. Including the test engineer's corrections is the detail that makes it credible, because it shows the plugin was used by someone other than its author.
@TheMattWalmsley [Claude Code]
https://x.com/TheMattWalmsley/status/2119230420277747934
@TheMattWalmsley listed the wild mistakes Claude Code made while editing videos: it left a restart in the cut three times, once after the edit had already been approved, killed the music three different ways, and brightened a thumbnail until the creator looked far too pale. Each of those failures now has its own fix built into the pipeline. It is a refreshingly honest counterweight to the flood of one-prompt video claims. The pattern is the same one serious users keep describing: every mistake becomes a permanent rule, and that accumulated list is the actual asset.
@hnshah [Claude Code]
https://x.com/hnshah/status/2119510287401509060
@hnshah described an in-house workflow built on top of Claude Code and Codex, running since February. Each project has a lead agent with persistent memory and a task board, which starts other agents in separate worktrees and tracks their progress. Changes pass automated checks before merging, and anything that needs human judgment lands in one review queue where the actual work can be inspected and feedback sent back to the agent. The team has recently started building dedicated software for it, which suggests the review queue, not the agents, is what needed a product.
@tomreilly_ai [Claude Code]
https://x.com/tomreilly_ai/status/2119511780057612375
@tomreilly_ai runs a consulting business on both Grok Bot and Claude Code, split by the kind of work. Grok Bot handles operations: triaging client emails, light research, queuing prospect follow-ups, scheduling time and tracking priorities. Claude Code handles the deep work: polishing the website and offer, building software projects and creating video content. The deciding question is simply whether the task needs someone to sit down for it, which is a cleaner rule for splitting a personal agent from a coding agent than most feature comparisons offer.
@wy_ones [Claude Code]
https://x.com/wy_ones/status/2119301047600771467
@wy_ones shares one memory and one set of skills between Claude Code on two computers, kept in sync automatically with Mutagen. Two lessons came out of it. Do not sync login credentials, because refreshing the login on one machine kicks the other offline. And do not let one side simply overwrite files that both have edited, because both machines write to memory and an overwrite silently loses notes. It is a small post, but these are exactly the failure modes people will hit as cross-machine agent memory becomes normal.
@ruyi3699 [Claude Code]
https://x.com/ruyi3699/status/2119225387133669770
@ruyi3699 tried graphify, which compiles code, SQL and documents into a queryable graph, on a half-finished personal project and found the agent reached for raw files much less often. Function calls and database tables live on the same graph, and it plugs into Claude Code, Cursor and Gemini CLI, though PDFs, images and video go through a model and burn tokens. A repo of a little over 200 files built faster than expected, with the caveat that this was one machine's experience. The verdict is a useful filter: not worth it for a small frontend repo, but real savings on data pipelines and multi-service codebases where agents otherwise grep endlessly.
@0xKevin00 [Claude Code]
https://x.com/0xKevin00/status/2119315838042050696
@0xKevin00 distilled a full workflow for making an AI short film with Claude, Codex and Seedance at the lowest cost and the fewest rerolls. Claude acts as director, writing the story, storyboard and every prompt and doing post-production in ffmpeg and Python, Codex draws character sheets and first and last frames, Seedance animates between them, and the human picks direction, approves spend and judges motion and sound. The core rule is to catch errors before they cost money, since a mistake in text costs a few words, in an image one redraw, and in video a per-second charge. A separate review session that has not seen the conversation checks each stage, and it caught a cup that changed size by 13% between two images, a hat brim that appeared and disappeared, and a left-right swap.
@dango12_dango [Claude Code]
https://x.com/dango12_dango/status/2119285368596095226
@dango12_dango kept forgetting what had been fixed while developing with AI, so together with AI built a place where each project gets one page and one line of progress per day. Claude Code can write to it directly when asked to log something to tororo. It is in beta until November 21. It is a tiny tool, but it targets a real gap: the agent remembers the session, while nobody remembers the project.
@ClaudeCode_love [Claude Code]
https://x.com/ClaudeCode_love/status/2119378395830350316
@ClaudeCode_love pointed to the creator Meng To, who had Opus 5.5 recreate a personal room in three.js code so convincingly that it looks like live footage. Inside that 3D room, the video shows Claude Code rebuilding an iPhone app that originally took 20 days to make in 2022. Working from the old demo video, it wrote 1,635 lines of SwiftUI and reached the first build in 20 minutes. Twenty days down to twenty minutes is the number that makes this more than a pretty render.
@Bajanpunkw [Claude Code]
https://x.com/Bajanpunkw/status/2119573610029293745
@Bajanpunkw, 22 and with no video editing experience, built a Claude Code skill in two days that clones a viral video style entirely in code. In one week it cloned the styles of four YouTube channels ranging from 55K to 23M subscribers, with 6.5 billion views combined. The skill takes a video apart frame by frame, covering palette, fonts, characters, camera, pacing and sound, and then Claude rebuilds the style and makes a new video in it, with no AI video tools and no editor. The whole skill, every stage and script, is on GitHub, which makes the claim checkable rather than just a screenshot.
🗣 User Voice
User Voice
Don't Ask Again is leaking secrets. @stretchcloud lays out Apiiro's numbers, 1,122 public settings files with live credentials and 2,539 secrets once related files were counted, plus bug reports that were auto-closed, while @yoppy0123 and @ShaneBOly turn it into a house rule: approve the pattern, never the password, and keep .claude/settings.local.json out of git.
People want a plan between $20 and the top tier, and limits they can predict. @Da7_Tech asks for a $50 Claude Code plan because the jump from $20 to $100 is too steep, @loverofapps says even the 20x plan runs out for real work, and @ekinciio warns that unattended background sessions burn through limits about five times faster.
The harness keeps changing under people's feet. @opanterasi relays a long-time user who counted 244 releases this year, most of them touching the system prompt, @who_rsehgal measured about 17.7k tokens spent just to answer hey, 71% of it tool schemas, and @0xLagosaur found agent files that silently never load when the name has a colon or the YAML does not parse.
Approving from a phone needs a way back. @rodri_shema says remote approval is only half the safety without a snapshot before the action and a clean rollback after it, and @embw_l0x reports that Duet lost the ability to see inside Claude Code sessions it had been checking on twice a day.
Other agents are starting to fence Claude Code out. @RoyOsherove and @TkunTosn both report Grok Bot refusing to hand coding work to a local Claude Code and pushing them to Cursor cloud agents instead, and @tvytlx argues that teams now collaborate worse because every person's context sits locked in a private session.
Don't Ask Again is leaking secrets. @stretchcloud lays out Apiiro's numbers, 1,122 public settings files with live credentials and 2,539 secrets once related files were counted, plus bug reports that were auto-closed, while @yoppy0123 and @ShaneBOly turn it into a house rule: approve the pattern, never the password, and keep .claude/settings.local.json out of git.
People want a plan between $20 and the top tier, and limits they can predict. @Da7_Tech asks for a $50 Claude Code plan because the jump from $20 to $100 is too steep, @loverofapps says even the 20x plan runs out for real work, and @ekinciio warns that unattended background sessions burn through limits about five times faster.
The harness keeps changing under people's feet. @opanterasi relays a long-time user who counted 244 releases this year, most of them touching the system prompt, @who_rsehgal measured about 17.7k tokens spent just to answer hey, 71% of it tool schemas, and @0xLagosaur found agent files that silently never load when the name has a colon or the YAML does not parse.
Approving from a phone needs a way back. @rodri_shema says remote approval is only half the safety without a snapshot before the action and a clean rollback after it, and @embw_l0x reports that Duet lost the ability to see inside Claude Code sessions it had been checking on twice a day.
Other agents are starting to fence Claude Code out. @RoyOsherove and @TkunTosn both report Grok Bot refusing to hand coding work to a local Claude Code and pushing them to Cursor cloud agents instead, and @tvytlx argues that teams now collaborate worse because every person's context sits locked in a private session.
📡 Eco Products Radar
Eco Products Radar
Codex: still the second agent in almost every stack and the default comparison, 162 mentions in the posts read
Cursor: the third name in nearly every comparison, 76
Gemini: the research and image model people pair with Claude, plus Gemini CLI, 46
OpenCode: the open harness, now with a browser that uses your real signed-in tabs, 32
Grok Bot: the personal agent that orchestrates Claude Code in several setups, and now also blocks it in some, 31
OpenClaw: used alongside Claude Code and benchmarked against it on cost, 26
Hermes: the self-hosted orchestrator people put above Claude Code and Codex, 24
Blender: driven by Claude Code for video scenes, games and apartment renders, 24
Kimi: the cheap model in head-to-head cost comparisons, 19
ffmpeg: the last step of nearly every video made in code, 17
Muse: the free personal agent people compare against, 15
Obsidian, Seedance and Antigravity: the memory vault, the video model and the extra worker, 13 each
Remotion and three.js: programmatic video and code-drawn 3D, 11 each
Claude Duet: now the phone remote for Claude Code sessions, 9
Orca, Conductor, DeepSeek Harness, Xcode and ElevenLabs: the orchestrators, the rival harness, the new Apple plugin and the voice layer, 8 each
Codex: still the second agent in almost every stack and the default comparison, 162 mentions in the posts read
Cursor: the third name in nearly every comparison, 76
Gemini: the research and image model people pair with Claude, plus Gemini CLI, 46
OpenCode: the open harness, now with a browser that uses your real signed-in tabs, 32
Grok Bot: the personal agent that orchestrates Claude Code in several setups, and now also blocks it in some, 31
OpenClaw: used alongside Claude Code and benchmarked against it on cost, 26
Hermes: the self-hosted orchestrator people put above Claude Code and Codex, 24
Blender: driven by Claude Code for video scenes, games and apartment renders, 24
Kimi: the cheap model in head-to-head cost comparisons, 19
ffmpeg: the last step of nearly every video made in code, 17
Muse: the free personal agent people compare against, 15
Obsidian, Seedance and Antigravity: the memory vault, the video model and the extra worker, 13 each
Remotion and three.js: programmatic video and code-drawn 3D, 11 each
Claude Duet: now the phone remote for Claude Code sessions, 9
Orca, Conductor, DeepSeek Harness, Xcode and ElevenLabs: the orchestrators, the rival harness, the new Apple plugin and the voice layer, 8 each
Comments