Super User Daily: 2026-11-01
This edition covers four days, October 27 to 30. The sharpest thread was context: one builder measured what compaction actually costs and replaced it with retrieval from the old session, and another found the same waste from the inference-engine side. The best workflows were loops with boundaries written down, like a Sentry watcher that fixes, tests and asks before it deploys. The non-coding uses kept getting more ordinary and more useful: a contract rewritten after clients found holes, a family tree pulled from parish registers, credit card points cashed out from a phone, car insurance quotes compared in one pass. On the OpenClaw side, long-time users explained in detail why they moved their daily agents elsewhere.
@kevinma_dev_zh [Claude Code]
https://x.com/kevinma_dev_zh/status/2115782178294620482
This developer runs a Claude Code loop inside a local monorepo that checks Sentry on a schedule, sends subagents to read the code, error logs, Supabase data and server logs, attempts a fix, and hands it to Fable for code review before deploy. The boundaries are written down: product interaction, copy, schema changes, data edits and every deploy need the human's sign-off, and each issue leaves a worklog so later agents can reuse what was learned. Problems that turn out to be design questions become Linear issues instead of live patches, and tests run on throwaway accounts with the agent starting the simulator itself. The visible result is that the old Sentry backlog has been mostly cleared and new alerts keep falling. The author's summary is the useful part: how much an agent can do depends on context, tools, permissions and clearly defined acceptance, not on the prompt.
@Processing_Unit [Claude Code]
https://x.com/Processing_Unit/status/2115808441516322945
This hardware engineer designed a 64-bit RISC-V CPU core with Claude Code and booted Linux on an FPGA board within a month. The core implements RV64GC with L1 and L2 caches, an MMU, a branch target buffer, an FPU and a JTAG debugger. That is a list of parts that normally takes a small team a long time to get through verification. It is one of the clearest signs this week that agentic coding has moved well past web apps into work where a single wrong bit means nothing boots.
@marekl [Claude Code]
https://x.com/marekl/status/2115184085438017905
A Czech user gave Claude Code a special skill and watched it search civil and parish registers on its own, read the scanned records and assemble a family tree without hand-holding. The author says they did not believe it would work and sat staring at the screen when it did. Genealogy is a good test case because the source material is old handwriting, inconsistent spelling and records split across archives. This is the kind of patient, tedious research that people pay specialists for, now done by a coding agent with a browser.
@Alezander9 [Claude Code]
https://x.com/Alezander9/status/2115842962521509968
This builder measured compaction on a real 400k-token Opus 5.5 session: 38 seconds and 1.99 dollars when the prompt cache had expired, 29 seconds and 17 cents when it had not, and the summary still lost the exact error message or file path you needed. Their tool Mainplane does the opposite: it keeps every session as a file, and when context fills it starts a new session told to retrieve what it needs from the old one as it goes. Same session, same model, it got back to work in 15 seconds for 10 cents, with no dependence on a cache hit and nothing paraphrased. The argument is the bitter lesson applied to memory: give the model the full record and good search instead of deciding for it what to remember. Mainplane is MIT licensed.
@shi3z [Claude Code]
https://x.com/shi3z/status/2116272460316500166
This Japanese developer built their own DeepSeek inference server and pointed Claude Code at it, and the wasted work became visible: large redundant prefills and cold prefills, and after each compaction a restart from zero that took tens of minutes even at 700 tokens per second. The conclusion is that token consumption depends heavily on the pairing of harness and inference engine, not just the model. A harness built not to compact, like Mainplane, plus an engine tuned for it is far more efficient. Running a stock harness against a self-hosted backend mostly shows how much time disappears into system-prompt prefill alone.
@Dandiggastech [Claude Code]
https://x.com/Dandiggastech/status/2115139454268539260
After reading the error-analysis chapter of Hamel Husain and Shreya Shankar's evals book, this developer realized the dataset was already on disk: hundreds of Claude Code, Codex and Hermes session traces nobody had read. They added one button per step to their agent dashboard, mark this as the failure, with a short note, and a failures list that exports to CSV. The first failure it found was in the sample itself: 193 of the 200 newest Claude rows were Untitled, because a usage check running every 10 minutes had left 1,893 throwaway sessions that pushed the real ones out. The first real mark was an agent saying a metronome app was done before it had looked at it, which needed a stop hook to force a screenshot. The takeaway is plain: the agent's answer is not the data, the trace is.
@undefinedKi [Claude Code]
https://x.com/undefinedKi/status/2115891841488867723
This post breaks down how Stripe runs agents on its repeated work: the same integration task took 20 days with plain Claude Code and 4 days with their setup, and payment integrations that used to take up to 6 months now take 2 to 6 weeks. They studied 125-plus past integrations and turned each step into one of 100-plus saved prompts, and while one agent builds, a second reads its transcript and notes where it got lost or which tools failed. After every run the agents leave learning and decision logs that another agent turns into prompt fixes. They first used an LLM orchestrator, found it slow, expensive and prone to doing the work itself, and rewrote it as plain code; agents also run test payments in sandboxes before any human reviews the PR. The copyable lesson is to fix the prompt, not the output.
@dotey [Claude Code]
https://x.com/dotey/status/2115282464935473225
A Chinese builder who has been making video-editing skills and a tool called BaoCut says that after Opus 5.5 most of those skills became unnecessary: given only speech-recognition and text-to-speech tools, the model edited the same video better with no skill than with one. The post also distills two practitioners: an ad team lead whose senior editor now ships five times the output at the same quality, and the VEED co-founder, who expects basic edits to fall from dollars to cents within a year. The concrete method is the valuable part: small per-style skills, seeded from 4 or 5 reference pieces, then trained by having the skill re-cut the senior editor's last 25 ads while that editor critiques each one, plus screen recordings of top editors explaining every cut. Most pieces now pass on the first or second version, and BaoCut has demoted its timeline editor to a simple preview.
@FinansowyUmysl [Claude Code]
https://x.com/FinansowyUmysl/status/2115763496742694996
A Polish business owner had paid a law firm 5,000 zloty for a client contract, then watched clients write in with objections that were all fair. Out of frustration they dropped the contract into Claude Code and found more gaps and risks than they could count. They rewrote the whole contract with Claude and are now looking for a lawyer only to verify it. The striking line is about trust: the author now trusts the AI more than the lawyers on this, and plans to run anything a lawyer says past the AI first.
@levelsio [Claude Code]
https://x.com/levelsio/status/2115417533457006656
This indie founder runs everything on VPS servers, with Termius SSH and Claude Code on the server as the only interface. All the businesses and the health data live there too, calories, sleep, and more, and the output is either a chat answer or a website Claude Code builds. The example question shows why: did magnesium last night improve my sleep, and Claude pulls sleep from WHOOP, nutrition from the food tracker, and bedroom temperature and air quality to check for confounders. It is a personal data warehouse with a coding agent as the query language.
@turingou [Claude Code]
https://x.com/turingou/status/2115388647805722809
While working on another product, this developer told Claude Code to take over their connected everyday phone and convert every unused domestic credit card points balance into airline miles, Marriott points, or rice, oil and small appliances shipped to family. It was a task that had sat on the to-do list for a long time. Small as it is, it shows a category that keeps growing: chores people write down and never do because each one is fiddly, not hard. An agent that can drive a phone clears that backlog in an afternoon.
@dontbesilent [Claude Code]
https://x.com/dontbesilent/status/2116071996048191920
When car insurance came up for renewal, several insurers' sales reps added this user on WeChat to send quotes, and the user simply forwarded every quote to Claude Code and got the best option back. The author's reaction is from the seller's side: with AI, undifferentiated products become almost impossible to sell, because the customer can compare everything in one pass. The only answers left are efficiency and price wars, or betting the customer does not use AI. It is a small personal use with a large implication for any business that relies on opaque pricing.
@sugoi_service [Claude Code]
https://x.com/sugoi_service/status/2115416282833301558
This post points at a Japanese air-conditioning installer who built a professional-grade SEO site with Claude Code and Codex, written entirely from first-hand field knowledge and experience. It reads as a technical site by an AC contractor for other AC contractors, and it is monetized through affiliate links for tools and business-startup kits aimed at the same trade. The author argues the model transfers directly to house cleaning, renovation and hairdressing. The real moat here is the craftsman's own know-how; the agent just removed the cost of turning it into a website.
@yaohongan2025 [Claude Code]
https://x.com/yaohongan2025/status/2115268326783906189
This Chinese indie maker went from buying a domain to the first paid orders in two days: two 5.99-dollar orders from Colombia via Google organic search. The method: collect X posts from people actually running money-making sites, hand them to Claude Code and ask it to learn their keyword-finding logic and build a keyword radar, but write the plan first with no code. Candidate keywords are checked by hand against TikTok and Google Trends before buying a domain, and the stack is an open-source SaaS template plus Cloudflare. The most useful line is about pricing: earlier sites never sold because generous free tiers attracted people looking for free tools, and cutting the free trial way down is what started the payments.
@binhara [Claude Code]
https://x.com/binhara/status/2115821078039716048
A Brazilian mobile developer with seven years of experience always believed their side apps failed because of code: a missing feature, a refactor, three more screens. Now Claude Code writes most of the code, and the gain was not build speed but free time to work on what actually sells an app: paywall, price tests, subscription model and marketing. After shifting effort there, the app DivineTalk hit 7,159 dollars in a single month, and the author sold it shortly after. The point is aimed at developers: if AI does the grunt coding in minutes, spending 90 percent of your time coding is the mistake.
@bcherny [Claude Code]
https://x.com/bcherny/status/2115491259381072269
Boris Cherny shared results from the state of Utah, which gave Claude Code to 34 developers across 10 state agencies in a four-week pilot. 80 percent had a working environment within 30 minutes, and 71 percent reported saving 20 or more hours by the end of the pilot. Government IT is usually where new developer tools arrive last and set up slowest. A half-hour setup and a majority saving a full workweek each is a concrete data point for public-sector adoption.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2114918759961919703
openskycam is an open-source sky camera: a 4K camera records the sky around the clock, detection models sort objects every morning, and only the good meteors get texted to the owner, with footage kept on local disk. Claude Code runs the setup and reviews detections, and when the model was raising false star alarms about half the time, Claude built a dataset from approved detections and retrained it. False alarms dropped to 7 percent while 90 percent of meteor detections were kept. It needs an RTSP camera, a Raspberry Pi, a machine with a big disk and a Cloudflare account. This is the agent doing the boring ML maintenance work, not just writing the app.
@fotoexamen [Claude Code]
https://x.com/fotoexamen/status/2115439925034139908
Setting up phone calls for an AI agent in Spain meant Twilio numbers, verification paperwork and per-minute billing, so this developer built agentring. Their agent, whether Claude Code, Codex, OpenClaw, n8n or a script, gets a call tool, and when it needs a decision, like whether to roll back a failed migration or fix and retry, the phone rings even when locked and a voice AI explains the situation. The conversation is transcribed back to the agent, which continues with the answer, and the author can also call the agents. It runs over VoIP from their own server with no phone numbers and no per-minute fees. They use it to get called on critical vulnerabilities, server security problems, critical emails and outages, on the view that a text message is not enough when something is actually wrong.
@illyism [Claude Code]
https://x.com/illyism/status/2115484780674556055
Inspired by marclou's MailCheap, this developer built a headless newsletter and lifecycle email engine on a Hetzner server in about 30 minutes, controlled entirely from Claude Code and Grok Bot over MCP. It has zero npm dependencies: Bun's server, SQLite in WAL mode and raw fetch calls to Resend's batch API. 6,134 subscribers were imported across 9 Resend audiences, marketing mail runs on a separate subdomain so it never touches transactional reputation, and a nightly backup goes to Cloudflare R2. The post includes the full build prompt with the table schema, which makes it easy to copy.
@illyism [Claude Code]
https://x.com/illyism/status/2115357045264425386
People kept pasting this developer's sites into Claude Code and bragging about scraping the whole database, and when the sites returned 403s the agents spoofed a Googlebot user agent, probed the API routes and spun up Playwright. So the author built a six-layer agent firewall that stops Claude Code, Cursor and Codex on the first turn while keeping Google SEO intact. It serves HTTP 200 decoy pages instead of 403s so agents do not retry, verifies real Googlebot by IP and ASN, and embeds a stop directive that makes the agent pitch the product to the person trying to scrape it. It even red-teams itself by spawning subagents to attack the URLs until they fail on turn one, and the SKILL.md is open source.
@0xMovez [Claude Code]
https://x.com/0xMovez/status/2115130015260291145
This builder made an AirDrop-style app to send files between a Mac and a Windows PC with no cloud, account or cable, in one afternoon on Opus 5.5 without writing a line. Opus picked a single Electron codebase, used UDP announcements every 1.5 seconds for discovery and direct streaming for transfer, and wrote a test first in which two fake devices find each other and send a folder plus a 3MB file byte for byte. Work was split across parallel agents, and separate reviewer agents caught a shader bug that would have rendered black on the Mac, which the Windows-only author would never have seen. GitHub Actions builds the dmg and exe on every release, and when the first Mac launch only bounced in the dock, Opus added a startup log and a fallback that forced the window open.
@yetone [Claude Code]
https://x.com/yetone/status/2115516478254624914
This developer vibe-codes with Herdr and, when out of the house, used a remote-control app to reach the home machine, which meant constant zooming and painful text and image input. So they built a native mobile Herdr client, currently for Claude Code only, and folded it into their Alma iOS app, which already had an end-to-end encrypted relay server. The app now includes a full Herdr client and remote desktop, and detects when the phone is on the same LAN to switch to a direct connection automatically. The author plans to open-source the relay so people can host their own. It is one more data point that the mobile control layer for coding agents is where people are building right now.
@hudson_gri [Claude Code]
https://x.com/hudson_gri/status/2115537588442505256
This founder's remote dev setup went through five versions in a year. It started as Claude Code on a Linux PC at home reached by Tailscale, Termius and tmux from a phone, which worked but was painful and kept dropping on bad Wi-Fi; a Vultr VPS was not much better. After quitting their job to start a company, they rebuilt on DigitalOcean, still stuck with the same tmux pain and unable to use Claude Code remote control because of account switching and Codex. The team then moved its boxes to a hosted service, built a web terminal that is actually good on a phone and added Herdr, and now runs a terminal that hops between machines and shows every session. The closing line is the lesson: the terminal keeps changing, the boxes do not have to.
@ProvableMat [Claude Code]
https://x.com/ProvableMat/status/2115130768859025839
This user puts three machines into one app: a MacBook Pro for daily work, a Windows PC for Blender renders and small local models, and a Mac mini running Hermes agents and automations day and night, alongside Codex, Claude Code and Hermes as harnesses. They had used Codex and Claude Code since launch, saw the desktop apps replace VS Code, and had been connecting machines over SSH with Herdr. Orca now pulls all of that together with Linear and GitHub tasks, and the mobile app lets them check and reply to agents away from the desk. The details they value say what multi-agent users need: a resource manager, an embedded browser, subscription usage tracking, and a keep-awake mode while agents work.
@DracowynX [Claude Code]
https://x.com/DracowynX/status/2115692297450835980
After more than a year of AI-assisted development, this developer shares the multi-model flow that has held up best. Opus 5.5 only writes task briefs and reviews, never the implementation; backend work goes through omp to GPT 6.1 Sol on high reasoning, which the author finds better than Opus at backend logic; frontend goes to Sonnet 5.5, which looks as good as Opus for less money; research and read-only tasks go to DeepSeek v4.1 Flash, which is not allowed to touch code. omp runs in its own git worktree with no commit rights, and patches must pass review before they merge. Unsatisfying results are not patched by hand: the task is re-dispatched with more reasoning or a different model.
@santtiagom_ [Claude Code]
https://x.com/santtiagom_/status/2115450361020317934
This Argentine developer says a large part of the job is now orchestrating agents, and shows the exact shape of a prompt: to add subscriptions to an app, send one agent to research the payment provider integration, another to review current permissions, a third to analyze failed payments and cancellations, then merge the findings into a plan before touching code. That one request defines a task per agent, what can run in parallel, what each should bring back, and a review step before implementation. After implementing, another agent reviews the change for errors. The honest part is the failure mode: sometimes the changes step on each other and it costs more time than a single agent would have, so the skill is deciding what to parallelize and making sure no two agents edit the same thing.
@ilyaraz2 [Claude Code]
https://x.com/ilyaraz2/status/2115303805550055598
This researcher interviewed about two dozen researchers, engineers and managers on how their teams actually use coding agents, to separate practice from what gets sold on social media. The first finding is blunt: fancy workflows do not really work, so use stock Claude Code or Codex, perhaps with subagents in rare cases, and loop through design doc, plan, implement, merge. The second is a rule for when skipping code review is acceptable: one-off demos that only need to work on one input, verifiable work like GPU kernels, or code that only needs to be directionally correct, like research and data work. Anything maintainable, reliable or customer-facing still needs a human reading the code and steering the tests.
@mardehaym [Claude Code]
https://x.com/mardehaym/status/2116115115456229753
This agency founder says the model is the least important decision in an AI project. In the last couple of months the team switched its default coding agent between Codex and Claude Code more than once, and nothing built for clients had to be redone. What carried over was the work around the model: rules written in code, real examples tested against every release, and a record of what the agent actually did. A model swap takes an afternoon; that surrounding work takes months and is the part a team actually owns.
@shotatykr [Claude Code]
https://x.com/shotatykr/status/2115979116742844695
This Japanese user recommends a habit before debugging any AI setup: ask Claude and Codex to map how your environment is actually configured and draw it, then fix bugs. The shared prompt is careful: read-only, no config dumps, no secrets in output, and it covers every layer, models and roles, shared and project instructions, skills, agent definitions, hooks, MCP and plugins, memory locations, scheduled jobs and hosting. It asks the agent to tell apart a symlink to the same file from a copy with the same content from a different version, and to distinguish a configured connection from one that actually authenticates and returns current work. The output is an overall diagram plus per-layer diagrams with broken references, stale copies, conflicting instructions and sync failures flagged with evidence.
@dotey [Claude Code]
https://x.com/dotey/status/2115505947187585302
This Chinese builder recommends running /autocompact 300k in Claude Code, down from the original suggestion of 400k, because 300k has been enough in daily use. On 1M-context models the default only compacts around 967k tokens, and since every message carries the full history, waiting that long means every late turn re-sends a huge context. Even with a warm cache most of that is cheap cache reads, but the volume still counts against usage limits. Pulling the trigger forward keeps history lean and stops single requests from ballooning.
@froessell [Claude Code]
https://x.com/froessell/status/2115730370477838635
This indie game developer shipped Bolt Pit, which averaged 2 minutes 17 seconds of playtime on CrazyGames with almost nobody returning the next day: the game worked, but the core loop ran out fast. The lesson they took is about what to hand the agent, not how to code. Next time they plan to give Claude Code a structured list of game mechanics, like the free Gameplay Atlas of 280 mechanics, each broken into its core loop and a small prototype to build first, before the first milestone. Agents can build a working game quickly; deciding whether the loop is worth replaying is still the human's job.
@asynkimo [Claude Code]
https://x.com/asynkimo/status/2115422615439392937
This developer built a platformer with Claude Code, SpriteCook and Godot, every sprite and animation included. The key step is giving Claude access to game asset generation and letting models like Opus 5.5 and Astra quality-check their own art: they look at every asset, check that it fits the game, and regenerate what does not. The workflow becomes a conversation about the game while the agent does the building. Self-review of visual assets is the part that used to need a human artist in the loop.
@sohail_beates [Claude Code]
https://x.com/sohail_beates/status/2115077718446587967
A self-described non-coder from Multan, Pakistan, built a product called beates with Claude Code and was interviewed live on ABN News about it. The author counts 317 hours of work so far and says it is still being built, with an opening date of November 5. There are no technical details, but the hours figure is the honest part: this is not a weekend vibe-coding demo, it is a sustained build by someone who could not write code before. It is a useful counterweight to the posts claiming finished businesses in an evening.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2115706645762420968
This Japanese developer's tool for adding motions to Live2D models passed 220 GitHub stars, and its scope has grown well past the original idea. It now adds expressions that can be reused on other models, loop motions that run for as long as the character is speaking, motion capture from a camera saved straight to a motion file, and streaming to OBS with motion switching from other apps. The usage model is agent-first: tell Claude Code or Codex where the model file is, give an instruction, and it starts working. It is a niche creative tool designed to be driven by an agent rather than a menu.
@Av1dlive [Claude Code]
https://x.com/Av1dlive/status/2115929453859881169
This user kept hitting Grok Bot usage limits, so they built a desk that routes the heavy work to Claude Code running on their own subscription. Grok Bot stays as the front end while the expensive coding work burns a different quota. The result, by the author's account, is that the limits now last the whole week. It is the cheapest form of multi-agent architecture people are actually adopting: split by who pays, not by role.
@yoshio_nocode [Claude Code]
https://x.com/yoshio_nocode/status/2115035492014244220
This Japanese builder's preferred setup is to install the herder agent manager on Grok Bot's cloud computer and have Codex, Claude Code and Cursor implement features as a team. Asking the bot to add a feature is enough: Claude Code on Opus plans and directs, Grok models implement in parallel, and Codex reviews and runs tests with evidence. The bot has learned the pattern and assembles the orchestration on its own, saying in effect that this is the usual routine. Handing the multi-agent platform to an AI bot, rather than operating it directly, is what makes it easy.
@EXM7777 [Claude Code]
https://x.com/EXM7777/status/2116202243401871444
This operator runs a team of agents on Grok Bot, Claude Code and DeepSeek, all connected to one Obsidian vault, following Andrej Karpathy's LLM wiki idea. The vault holds what the business sells and to whom, the house writing style and banned words, winning hooks and ads from the niche, step-by-step job procedures, landing-page swipe files and every piece of content ever written. Every agent reads it before working, and the agents keep it updated themselves. The practical payoff is the line about corrections: fix one agent once, and the whole team remembers.
@FaztTech [Claude Code]
https://x.com/FaztTech/status/2114953702003855530
A Spanish-language developer educator relays a point from CSS teacher Kevin Powell, asked why anyone should learn CSS when Claude Code builds the site: without specific requests you get the same site as the last 5,000 people, so the Bootstrap era has become the Claude Code era where everything looks alike. The author extends it to all of software. Ask for a backend and you get one that works, but it may lack authentication, input validation, speed or documentation, and you will not notice unless you knew to ask. Learning did not die, it changed: not memorizing syntax, but knowing what a REST API, a JWT or a cache is so you can ask for it and check it.
@matthewcanham [OpenClaw]
https://x.com/matthewcanham/status/2115103609474154750
For six months OpenClaw ran huge parts of this user's daily life, and after forcing a 10-day trial of Grok Bot they moved everything over. The reasons are specific: voice calls with full context, to the point of running daily standups with the bots by voice, each bot with its own voice; multi-threaded conversations, replying to one topic while the bot replies to another, which removes the urge to spin up duplicate agents; and specialists that collaborate, like an inbox-triage agent consulting others before deciding whether to escalate an email. It is the most detailed migration story of the week, and it lists exactly which features OpenClaw would need to win users back.
@designertom [OpenClaw]
https://x.com/designertom/status/2115665640166187238
A year-long OpenClaw user shares the structure that carried over into a multi-bot setup. Create an orchestrating chief-of-staff bot that routinely checks each bot for drift and performance and routes passive events like emails to the bots that need them; section the bots by function, such as growth, studio, social and product; send all coding through the Claude Code CLI. Add a security bot for daily scans of X and of devices, repos and the network. And keep an external vault in Git, Obsidian or Notion as shared memory for every bot, with a top-level index that all agents must read before querying it.
@alexeheath [OpenClaw]
https://x.com/alexeheath/status/2115571474878894249
Shopify CEO Tobi Lutke synced an OpenClaw agent with a Fitbit and asked it to identify which meetings were the most stressful, then told the people involved. According to the clip, they did not think it was funny. It is a playful example, but the data join is real: physiological signals matched against a calendar to rank meetings by stress. Personal agents with wearable data are going to surface uncomfortable truths about work, and who gets to see them is about to become a real question.
@hardiksd [OpenClaw]
https://x.com/hardiksd/status/2115894638724932061
This developer tried using OpenAI's Dot as the main coordinator for all software work and lists why it fell short. It frequently loses the permissions it was given, which stalls overnight work; it does not coordinate well when asked to run reviews and analysis while assigning tasks to both a local and a cloud computer; and it cannot run third-party harnesses or third-party models through Codex, which the author relies on via a Codex router with DeepSeek v4.1 Flash when limits run out. Between those issues the author keeps falling back on OpenClaw. The wish is for Dot and OpenClaw to talk to each other, with Dot as the coordinator.
@charlesmendez [OpenClaw]
https://x.com/charlesmendez/status/2116145208932893038
Four months ago a friend of this author who runs a roofing business asked for an OpenClaw install; the friend was already a heavy ChatGPT user and set out to build an app for the company from scratch. For months the calls were excited progress reports. Then the calls turned into requests for help, with the AI hallucinating and unable to get things right, and yesterday the friend stopped after spending thousands of dollars on tokens. The author hears the same story from companies that bought agents promising production systems built from scratch. It is the clearest cautionary case of the week: a non-developer can get surprisingly far, and the last stretch is where the money goes.
🗣 User Voice
User Voice
Rules that arrive from the server with no changelog are the new trust problem. @chasmmmmmmmmmmm surfaced a server-pushed instruction that stops Claude Code from recreating specific existing games, with no setting to turn it off; @Vincent_AINotes relays a Claude for Startups member whose board-game digitization business stalled under that rule and moved to another model; @babayagatwt puts it simply: paying 200 dollars a month and waking up to new rules nobody agreed to.
Usage limits still dominate, and the explanation is now specific. @bcherny says the team went through every shared session ID, found a few small token leaks being fixed, and traced most fast burns to one or two heavy sessions with xhigh effort, /goal, Workflows or Agent Teams, with subagents inheriting effort by default; @AlexLu7jx suggests screenshotting the usage page on the same day each week so complaints have numbers; @Im_IrushiK compares a 20-dollar Claude plan lasting hours with a Codex plan gone in under 20 minutes.
Session management is the everyday friction. @binghott now opens a new Claude Code session just to search for the old session the built-in search cannot find; @arvidkahl asks why Claude Code cannot see its own weekly usage percentage; @theaidesk_ wants multiple accounts in the desktop app; @omarvvvr runs Codex and Claude Code on the same repo and finds each has no idea what the other just did.
The agent toolchain is now an attack surface. @cursor_ai disclosed a tampered Claude Code extension on its marketplace affecting up to 5,000 users; @IntCyberDigest warns that updating does not remove the planted SSH key and startup scripts; @takapoyo_ai explains why launching an agent in the Downloads folder lets a planted AGENTS.md take over; @tetumemo lost an entire home folder, including Claude Code settings and skills, to a delete command from another bot, and concludes that written rules do not stop a command, only per-command deny settings do.
On the OpenClaw side, @JeffNylen likes the last month of updates but wants to be able to turn off the redaction filter, and @ExWhyyyyy asks why it climbs to 2GB of RAM.
Rules that arrive from the server with no changelog are the new trust problem. @chasmmmmmmmmmmm surfaced a server-pushed instruction that stops Claude Code from recreating specific existing games, with no setting to turn it off; @Vincent_AINotes relays a Claude for Startups member whose board-game digitization business stalled under that rule and moved to another model; @babayagatwt puts it simply: paying 200 dollars a month and waking up to new rules nobody agreed to.
Usage limits still dominate, and the explanation is now specific. @bcherny says the team went through every shared session ID, found a few small token leaks being fixed, and traced most fast burns to one or two heavy sessions with xhigh effort, /goal, Workflows or Agent Teams, with subagents inheriting effort by default; @AlexLu7jx suggests screenshotting the usage page on the same day each week so complaints have numbers; @Im_IrushiK compares a 20-dollar Claude plan lasting hours with a Codex plan gone in under 20 minutes.
Session management is the everyday friction. @binghott now opens a new Claude Code session just to search for the old session the built-in search cannot find; @arvidkahl asks why Claude Code cannot see its own weekly usage percentage; @theaidesk_ wants multiple accounts in the desktop app; @omarvvvr runs Codex and Claude Code on the same repo and finds each has no idea what the other just did.
The agent toolchain is now an attack surface. @cursor_ai disclosed a tampered Claude Code extension on its marketplace affecting up to 5,000 users; @IntCyberDigest warns that updating does not remove the planted SSH key and startup scripts; @takapoyo_ai explains why launching an agent in the Downloads folder lets a planted AGENTS.md take over; @tetumemo lost an entire home folder, including Claude Code settings and skills, to a delete command from another bot, and concludes that written rules do not stop a command, only per-command deny settings do.
On the OpenClaw side, @JeffNylen likes the last month of updates but wants to be able to turn off the redaction filter, and @ExWhyyyyy asks why it climbs to 2GB of RAM.
📡 Eco Products Radar
Eco Products Radar
Codex, the default second harness, named next to Claude Code in a large share of posts.
Grok Bot, the personal agent people now put in front of Claude Code, often to route heavy work onto a separate quota.
Hermes Agent, the most-cited alternative harness for personal agents, usually alongside OpenClaw.
Cursor, mostly this week as the marketplace that served the tampered extension, and as a harness comparison point.
DeepSeek v4.1 Flash, the cheap fallback model for research, read-only tasks and running out of limits.
Meta Muse, the free personal agent that users compare against Grok Bot and Dots.
OpenCode, the open harness that now calls the official Claude Code through the Agent SDK.
Obsidian, the shared memory vault behind several multi-agent setups.
Blender MCP and Blender Skills, the main route for 3D work from Claude Code.
Aura Autopilot, an open-source agent that playtests games and files bug reports, launched widely this week.
Herdr, Orca and Magpie, the control layer for running sessions across machines, phones and accounts.
Copilot Fusion, GitHub's model router benchmarked against Claude Code at about half the cost per task.
Gemini API CLI, Google's tool that lets Claude Code call Gemini for images, video, speech and transcription.
Graphify and Impeccable, the knowledge-graph and anti-generic-design skills that show up in design setups.
ffmpeg-skill, Remotion, HyperFrames and Kino, the video-editing stack built around coding agents.
Codex, the default second harness, named next to Claude Code in a large share of posts.
Grok Bot, the personal agent people now put in front of Claude Code, often to route heavy work onto a separate quota.
Hermes Agent, the most-cited alternative harness for personal agents, usually alongside OpenClaw.
Cursor, mostly this week as the marketplace that served the tampered extension, and as a harness comparison point.
DeepSeek v4.1 Flash, the cheap fallback model for research, read-only tasks and running out of limits.
Meta Muse, the free personal agent that users compare against Grok Bot and Dots.
OpenCode, the open harness that now calls the official Claude Code through the Agent SDK.
Obsidian, the shared memory vault behind several multi-agent setups.
Blender MCP and Blender Skills, the main route for 3D work from Claude Code.
Aura Autopilot, an open-source agent that playtests games and files bug reports, launched widely this week.
Herdr, Orca and Magpie, the control layer for running sessions across machines, phones and accounts.
Copilot Fusion, GitHub's model router benchmarked against Claude Code at about half the cost per task.
Gemini API CLI, Google's tool that lets Claude Code call Gemini for images, video, speech and transcription.
Graphify and Impeccable, the knowledge-graph and anti-generic-design skills that show up in design setups.
ffmpeg-skill, Remotion, HyperFrames and Kino, the video-editing stack built around coding agents.
Comments