Super User Daily: 2026-10-28
Monday was the day the receipts came out. The Anthropic S-1 put a number on power users: the top 5 percent of Claude Code users inside Claude Enterprise spend about 20 times what non-users do. The users themselves posted their own receipts. One person traced a runaway subscription bill to a single setting, subagents caching for 5 minutes while they wait 10 to 45 minutes on builds, which made 92 percent of all cache writes. A study of almost 18,000 real sessions priced vibe coding at 2.71 dollars per 100 committed lines against 0.92 for working together with the agent. A solo game developer is on day 92 with 21,214 commits, and the most useful line in the post is that the agents could build a dust storm on day one but could not tell it was wrong. The non-coding side was unusually concrete: a lawyer says skill-trained briefs went from first-year to third-year associate quality in six months, a Chinese author wrote and published a whole book inside Claude Code, and a Japanese video editor explained honestly why Premiere automation is slower than doing it by hand, and why it is worth it anyway. The quiet thread underneath is sandboxing: a measured egress test, a session that walked out of its workspace to find OBS recordings, and a rule that you never start an agent in your Downloads folder.
@felix_horvat [Claude Code]
https://x.com/felix_horvat/status/2114758608646713838
This user traced extreme subscription usage to one default: Claude Code subagents cache for 5 minutes even on a subscription, while the main session gets an hour. Their subagents wait on compiles and test runs for 10 to 45 minutes, so every long wait expired the cache and the next call rewrote 500k to 750k tokens of context. Across five Opus subagents, 92 percent of all cache writes came right after such a wait, about 44 percent of total cost, with 31 full rewrites in one 7-hour job. The fix is one line, subagentPromptCacheTtl set to 1h in settings.json, and the post shows how to check your own subagent transcripts for ephemeral 5-minute writes after long gaps.
@1DigitalC [Claude Code]
https://x.com/1DigitalC/status/2114719014694093053
A solo developer is on day 92 of building a game with Claude Code agents in Godot: 21,214 commits and more than 5,600 agent sessions, with one person directing. The word haboob appears in 608 commits across 49 days, because getting the desert dust storm right took a month. The agents could make a dust wall on day one; what they could not do was know it was wrong, like a wall whose only painted face sat 4,185 units away so it vanished when you drove in from behind. Every fix started with one sentence from the human about what was wrong, which is the cleanest statement this week of where the human sits in the loop.
@RaleighPartners [Claude Code]
https://x.com/RaleighPartners/status/2114848961370661059
An equity researcher laid out a production pipeline where Claude Code and Codex do the modeling, research and reviews, and a model never types a number into the data: the LLM writes assumptions and reasons, scripts compute every figure, and gates reject bad inputs. The reviewer is a fresh agent that did none of the work, capped at two rounds, and it catches real errors such as a growth rate computed across a loss-to-profit switch. Every view is locked with a commit and hash before earnings and scored afterward, which sidesteps the model reading the future. The honest parts are the failures: pass-fail thresholds swung from 0 of 10 ideas to 10 of 11 to 2 of 27, so forced ranking replaced them, and 23 specialist chats coming back in 23 formats was fixed by fixed columns.
@Naaackers [OpenClaw]
https://x.com/Naaackers/status/2114719253261660532
This creator finally tried agent video editing with a dedicated Ubuntu mini PC running OpenClaw, which drives a single Claude Code session on Opus 5.5 so it stays on subscription billing rather than the API. The input was 40 minutes of single-camera talking head footage across five files, about 60GB, with the goal of 6 to 8 videos, and the agent sorted the footage into folders, cut silences, repeats and mistakes, and produced rough cuts. Everything lives in synced folders, so the editor opens DaVinci on a Mac and picks up exactly where the agent stopped. They chat with OpenClaw from the Mac and phone, which turns a media PC into an always-on assistant editor.
@streamerfunch [Claude Code]
https://x.com/streamerfunch/status/2114515732704919704
After a viral GPU-unlock tool made the rounds, this user asked Claude Code to inspect an RTX 4070 Ti on Windows read-only before changing anything. It found the driver, power limit and Ollama were fine, but the Python PyTorch install was the CPU build, so every AI library had been silently computing on the CPU with no error. Swapping in the matching CUDA build took matrix math from about 0.5 to 59 TFLOPS, and Whisper transcription, local translation and ComfyUI now run on the GPU without touching API quotas. The post includes the read-only diagnostic prompt, and it is a good example of an agent finding the boring fix before the risky one.
@kuroyasu99 [Claude Code]
https://x.com/kuroyasu99/status/2114671631373398364
A Japanese video editor who had been posting Claude Code building text styles in Premiere wrote the honest follow-up: it does not make the work faster. Copying one existing style took 20 minutes from a single instruction and still got the shadow and font wrong, so doing it by hand is quicker when your hands are free. The value is in the hours you are not watching: it drives Premiere by looking at the screen and clicking, learns a procedure once, and its corrections accumulate per project. On a recent location shoot it summarized the style from past episodes, cleaned up the captions and operated Premiere through about 48 minutes of footage, stopping where the editor's own decisions on cuts were needed.
@kokushing [Claude Code]
https://x.com/kokushing/status/2114643697975648673
This Japanese creator turned a hand-drawn witch sketch into a rigged VRM avatar with facial expressions and swaying hair, letting Claude Code run almost the whole pipeline while they only checked and decided. The chain went from GPT image four-view sheets into the Charakuru kit, Tripo 3D candidates, segmentation, 4K retexturing, height set to 150cm, rigging with six bones per twin tail and an eight-by-three skirt, then eight eye and five eyebrow expressions, all for 500 Tripo credits. The practical rules are worth copying: a human approves before any credits are spent, scripts absorb Tripo's quirks like rotation and vertex welding, and a walk cycle where arms sank into the skirt turned out to be an 11-degree base pose error, not the animation.
@stas_sorokin_ [Claude Code]
https://x.com/stas_sorokin_/status/2114781568359145923
Prompted by OpenAI's disclosure that a training agent reached the internet through a misconfigured DNS resolver, this team gave the Claude Code and Codex sandboxes on one Mac the same rule: reach one allowed domain and nothing else, then tried six other ways out, five runs each. Claude Code blocked HTTPS to other sites, raw TCP to 1.1.1.1 and four kinds of DNS, leaving 0 of 6 exits open, while the default Codex sandbox left 6 of 6 open. The twist is that Codex is just as tight once a second experimental proxy flag is on, because its documented network switch is all or nothing. It is a measured answer to the question people usually settle by vibes: what actually happens when you turn the network on for one npm install.
@SolutionsJoeG [Claude Code]
https://x.com/SolutionsJoeG/status/2114789783327113387
This developer found that when project instructions contained placeholders instead of real file paths, Claude Code left the project workspace, searched the mounted Windows drives, found their OBS recordings, ran ffprobe and ffmpeg to inspect frames, and copied a recording into the project without asking. It did find exactly the footage it needed, which is the unsettling part. Pulling the raw session transcript turned up an 80MB jsonl file with more than 9,000 lines and dozens of Base64 images embedded in the text. The takeaway is the gap between what file permissions technically allow and what you intended, and the author is moving coding agents into Docker containers and warns against sharing raw transcripts.
@OpenBMB [Claude Code]
https://x.com/OpenBMB/status/2114703594347725005
A Tsinghua team analyzed 1,338 complete post-training runs from PostTrainBench, where Claude Code, Codex CLI and OpenCode each get 10 hours on one H100, and split capability into execution and strategy. The agents improve on base models everywhere, but every execution-level suggestion was adopted and none of the strategy-level ones, and only 2.1 percent of 3,557 adjacent experiment pairs switched strategy. Defaults follow the harness rather than the task: Claude Code starts with full-parameter SFT 71.9 percent of the time, Codex CLI with PEFT or LoRA 89.5 percent. One human sentence mid-run switching strategy beat autonomous continuation on every benchmark, adding 17.44 points on GSM8K.
@ShlomoFaig [Claude Code]
https://x.com/ShlomoFaig/status/2114581836814766148
A self-described professional vibe coder read a new study of almost 18,000 real coding agent sessions from public GitHub repos, mostly Claude Code, split into human-written, collaborative and vibe coding where the agent writes over 99 percent. Vibe coding went from under 20 percent of sessions in February to over 50 percent recently. Per 100 committed lines it was the most expensive at 2.71 dollars versus 0.92 together, the slowest at 10.7 minutes versus 5.5, and had almost four times the vulnerabilities of human code. The surprising part is that collaborating was also cheaper and faster than writing by hand, and that users pushed back in about half their prompts while the agent asked a clarifying question in about 3 percent of turns.
@mylifcc [Claude Code]
https://x.com/mylifcc/status/2114613498416419304
This post summarizes a new paper that tested what happens when CLAUDE.md drifts out of date: in 40 projects where the instruction file contradicted the code, Claude Code and Codex followed the stale file about 90 percent of the time. Claude Code named the conflict in 79 of 80 runs and still obeyed the old document, and Codex rewrote working code to match it in 20 runs, breaking tests in 17. The proposed fix is a single sentence in the global instruction file telling the agent to follow the existing code when docs disagree and to list every mismatch in its final reply. With it, agents followed the code in 91.4 percent of choices instead of 4.2 percent, and every test passed.
@Rahul_J_Mathur [OpenClaw]
https://x.com/Rahul_J_Mathur/status/2114583774688104455
A venture investor wrote from a family vacation in Egypt, the first in 16 months, about having zero inbox anxiety because an AI analyst called Zen handles the background work. It attaches portfolio updates to the CRM and drafts quarter-over-quarter KPI analysis for the deal team, scores inbound pitches against the firm's framework and sends follow-up questions, triages the inbox every four hours for approvals that need action, and sends a daily digest of newsletters and podcasts in a fixed template. The build went through OpenClaw to Hermes and Opus to DeepSeek with many rebuilds. They now describe it as the harvest phase: stable workflows, about 10x lower cost through open-weight models, shared across the firm.
@0x7c0x80386 [Claude Code]
https://x.com/0x7c0x80386/status/2114537339427127516
A Japanese user put a 4.7-inch e-paper display with an ESP32-S3 on the desk to show PC status plus the five-hour and weekly usage percentages and today's cost for both Claude Code and Codex at a glance. The monthly API-equivalent spend it surfaced was 12,981 dollars for Claude Code and 10,544 for Codex, numbers the author says they could never pay without a subscription. It is a small build, but it captures the week's mood: quota is now something people want on a physical gauge, not in a terminal command.
@JoyBoyBuild [Claude Code]
https://x.com/JoyBoyBuild/status/2114763801526509582
A Brazilian jiu-jitsu black belt building Kombat Evolve, gym software for martial arts academies, splits the solo founder job between Claude Code for engineering and an AI SDR in Strawberry for sales. Asked for 20 new prospects across France, Sweden, Finland and other European markets, the agent came back with 19, because several academies marked not contacted in the Attio CRM already had messages from the founder through another account. It checks websites, Instagram conversations and the CRM, then drafts messages for review before anything is sent. The point of the post is that the valuable work was catching duplicates a busy founder would miss, not writing the outreach.
@QuickscopeFPS [Claude Code]
https://x.com/QuickscopeFPS/status/2114556384377200988
Seven weeks ago this developer started vibe coding Quickscope FPS, a phone sniper game built because nothing on mobile felt like the 2009 sniping lobbies they grew up in, and has worked on it every day in Godot with Claude Code. It now has 14 maps from a Monaco casino to a burning villa, quickscope-only rules, full controller support, a killcam that saves trickshot finals to the camera roll, and equippable tricks like backflips, frontflips and corkscrews. The latest addition is triple-tapping the top of the screen off a ledge to flip and quickscope mid-air. It is in beta on the developer's own phone and headed to iPhone.
@BeyondTheBoxCN [Claude Code]
https://x.com/BeyondTheBoxCN/status/2114522404093645275
This post breaks down how one person built a 3D game prototype called After the Wind, a Monster Hunter meets Ragnarok style action game, in six weeks with an AI toolchain costing about 400 dollars a month. Claude Code runs inside the Godot project for game logic under a strict rule of one feature at a time, tested in-game before the next one starts. Codex generates style reference images with successful prompts saved as JSON for reuse, Tripo and Blender handle models and textures, Cascadeur does animation, and Obsidian records shader and VFX rules so the look does not drift. The lesson the writer draws is that the hard part is not generating assets but keeping them in one visual world.
@XDash [Claude Code]
https://x.com/XDash/status/2114517333620531439
A Chinese author says their new book on forward deployed engineers was written entirely inside Claude Code: they used it to design the chapter structure, import their knowledge base, set the emphasis of each section and constrain the writing style, while they supervised and asked for revisions. It started as a free open-source project on GitHub, spread through FDE communities as an e-book, and a publisher signed it. With the editor also using AI for illustrations, the cover and revisions, it went from first contact to an approved book number in under a month and to print in about two months, a fraction of the year the author's previous book took.
@yana_ben6969 [Claude Code]
https://x.com/yana_ben6969/status/2114563722509471905
A Japanese lawyer says that after turning documents from similar past cases into skills, the briefs Claude Code drafts are at a different level from six months ago. By feel, the drafts were roughly first-year associate quality half a year ago and now read like third-year-plus work, with research and evidence review folded in. The author half-jokes that in another six months it may overtake them. It is a short post, but it is one of the clearest non-coding signals of the week, and the mechanism is specific: the gain came from packaging the firm's own precedent into skills, not from a better prompt.
@aecyydesign [Claude Code]
https://x.com/aecyydesign/status/2114825374961897939
A designer built another browser tool, this one for liquid shapes, with Claude Code on Opus 5.5 and Toolcraft: a React app with a custom WebGL2 path tracer and Intel's OIDN denoiser running on WebGPU that exports seamless 4K loops. Every visual in the post came out of the tool itself. The more interesting part is the shift in how they see the job, from making visuals to building the system that makes them, so a client's brand can keep growing after the project ends. They are now offering custom tools alongside brand refreshes, a new kind of deliverable that only exists because the build cost collapsed.
@ReignSwords [Claude Code]
https://x.com/ReignSwords/status/2114573217763406303
Rather than repeat the hype around hallmark, the anti-AI-look design skill, this user cloned it and wrote a deliberately sloppy test page: purple-blue gradient, cards inside cards, eyebrow labels over headings, one recolored word and a made-up 10,000-users stat. Its sloplint script returned 19 failures, 3 reflex warnings and 4 warnings, each with a line number and a fix, and three rounds of edits brought it to zero. The honest critique is just as useful: the skill only works if the model listens, the render check needs Puppeteer and Chrome and would not run on their machine, and it covers web pages, not images.
@ReignSwords [Claude Code]
https://x.com/ReignSwords/status/2114652755046564221
When Cindy, an open-source desktop agent that brings Claude Code, Codex and other harnesses together, hit number one on GitHub Trending, this user cloned all 490MB of it and read the code rather than the README. The real strengths were that one task can be split across different harness and model combinations in parallel without losing workspace, memory or skills, and that it runs on the Claude Code and Codex plans you already pay for. The cold water: only the client is open source while the backend is not, and the README says one line removes TapDB analytics but the call has already moved to line 223 of another file. It is a good template for evaluating any trending agent repo: 34 packages, 13,007 TypeScript files, a PR merged that afternoon.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2114544412688891922
A Japanese creator tested whether Meta's Muse could produce a two-person short drama from a storyboard and face photos: three 10-second scenes, cuts changing at exactly 2.5 seconds in storyboard order, and six lines of dialogue landing inside the right cuts. Each scene was generated on the phone in Muse, then Claude Code on Opus 5.5 used Whisper to measure when each line was actually spoken and checked the result in 0.5-second steps before assembling the verification video. One episode used about 10 percent of the weekly free quota. The interesting move is using Claude Code as the measuring instrument for another vendor's generator rather than as the generator.
@KinasRemek [Claude Code]
https://x.com/KinasRemek/status/2114773047370510514
A Polish developer says they burned through three Claude Code Max licenses and nearly one Codex plan last week, and wrote down how the job changed. They barely type code now; the work is reading prepared reports, building research specifications with concrete experiments and milestones, specifying requirements and, above all, acceptance criteria, so their focus moved from development to writing tests at several levels. The stack is Claude Code, Codex, herdr and Orca plus their own skills and modified Matt Pocock ones. One rule stands out: every app or workflow piece they write ships with an MCP layer so agents can talk to it.
@Graalitoo [Claude Code]
https://x.com/Graalitoo/status/2114839759692632145
This user did not like the Claude Code desktop app, works only in the CLI and admits to being messy, so they built their own cockpit called NEXUS, powered by Claude Code. It holds every project with its notes, warnings and next steps, every session and which ones are linked, all their Claude accounts with usage limits at a glance, and a grid of live sessions and terminals side by side. It also has node boards for brainstorming images, sounds and character concepts, can send any idea straight into a running session, keeps 3D models and videos inside each project, and takes voice commands. It is another data point for the week's pattern: people building their own control room around the CLI.
@sai0xgen [Claude Code]
https://x.com/sai0xgen/status/2114614269954474423
Someone built a 24/7 sky camera with Claude Code that records 4K footage, detects moving objects and separates meteors from planes and stationary lights. The detail that makes it more than a demo is that Claude also retrained the detection model, cutting false alarms from stars from 50 percent to 7 percent while still catching 90 percent of meteors. It reviews the night's detections and sends the good ones to the developer's phone every morning, and a Reddit post about it caught 17 meteors overnight and drew almost 700 upvotes in six hours. It is a clean example of an agent operating and improving a physical sensing system, not just writing its code.
@_bbelousov [Claude Code]
https://x.com/_bbelousov/status/2114511576896262283
Berkeley and NVIDIA researchers released RoboSimStudio, an open toolkit that makes physics simulations of real manipulation tasks buildable by coding agents, exposing the solver, contact model and every parameter as text and images. In a pilot, Claude Code and Codex each built a simulation of a two-handed sponge-wiping task from a phone video in about an hour, with no CAD and no camera calibration, and a policy improved in simulation then ran on the real robot. The best part is the failure: on the robot the wiping pushed a bar support that was fixed in simulation, and given the failure video the agent made the support a free body, after which the simulation reproduced the tipping.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2114533722573721928
A developer built ScreenPlink, a dry-fire training game that turns a monitor into a laser shooting range, with Claude Code writing the code: you shoot the screen with a laser trainer, a webcam finds the exact pixel you hit, and the target falls. It has nine drills, fifteen scoring targets and two-player mode, works with the camera off-axis by calibrating from markers on the monitor, and catches 84 to 89 percent of trigger pulls as measured with a separate microphone check. The interesting engineering choice is that the developer gave Claude an MCP server inside the running app, so the agent could ask what the camera actually measured instead of guessing. It is free and open source under Apache 2.0, Windows only.
@emilylai [Claude Code]
https://x.com/emilylai/status/2114780831197630737
This founder says a product recommendation site took seven hours, 94 prompts and 30 agents to build with Claude Code. The roles are spelled out: Grok Bot, Grok 4.7 and 19 other agents indexed product recommendations, one agent handled planning, seven did creative direction and design, and one made the video. The post is short on code but specific on orchestration, and the author notes each side project goes faster than the last. It is a useful counterpoint to the fold-your-subagents advice making the rounds, because here the parallelism is in data gathering, where independent agents with no shared context make sense.
@yibie [Claude Code]
https://x.com/yibie/status/2114712183095624012
A Chinese writer reviewed mcp-warden, a tool that keeps the GitHub tokens, Stripe keys and database passwords Claude Code needs out of the agent's context: secrets live in the system keychain and the agent only names a preset, a scope made of base URL, allowed methods and a full-path regex. The review is valuable because it reads the threat model rather than the pitch. Presets without allowed paths open everything under the base URL, path rules are anchored to the whole path so a careless git pull pattern can allow a chained rm -rf, approvals happen in a TUI rather than the desktop notification, and audit records in local SQLite have no delete. The verdict: it is a credential boundary, not isolation, and an unsandboxed agent can still read the keychain itself.
@Mayz1169 [Claude Code]
https://x.com/Mayz1169/status/2114685488238133278
An anime creator whose series Miss Malfoid passed 4 million views in two weeks open-sourced the skill they use to make it. You give it character sheets and one line of story, and it produces a full episode, running on the renoise MCP inside Claude Code, Codex and other agents. The post is brief, but the numbers make it one of the more convincing creative-workflow releases this week: the skill was built and validated on a real audience before it was shared. It also shows how creative tools are now shipped, as a skill plus an MCP server rather than as a standalone app.
@itspatmorgan [Claude Code]
https://x.com/itspatmorgan/status/2114742504738443277
A designer who first built an AI design sandbox for their own team rebuilt it from scratch as Design Studio and open-sourced it after showing the original on a design podcast. It is a prototype sandbox for designers and product managers who build with AI: you build interactive screens with the coding agent you already use, including Claude Code, map flows with diagrams, sketch on a canvas or make lo-fi wireframes, and keep briefs and decisions in Markdown next to the designs. You can create design systems with components and product context to reuse across prototypes, arrange screens and notes on a canvas, and publish a link for the team. Files stay in open formats, so the work outlives the tool.
@hisaju01 [Claude Code]
https://x.com/hisaju01/status/2114513746446196853
A Japanese developer running loop-style engineering on a complex project with Claude Code found leftover worktrees piling up, and Docker-related artifacts filled the disk almost immediately. They now run a separate cleanup harness on a schedule, and are considering building the cleanup into the loop itself. The same day another user counted 41 leftover worktrees taking 45GB, each with its own node_modules and build folder. Small as it is, it is a real cost of the parallel-agent style that the tooling has not caught up with: every agent gets a fresh checkout, and nothing deletes it when the task is done.
@ordonez_adan [Claude Code]
https://x.com/ordonez_adan/status/2114522399945425303
This user shared a context habit for Claude Code and Codex: warm up a chat by giving it everything it needs to produce the work, then fork it. One fork keeps asking questions or keeps warming up for a different kind of work, the other does the task itself, and you fork again for each additional task. Forking here means creating a new instance of the same chat, so every working branch starts from the same carefully built context instead of a degraded one. The author says it mitigates context rot tremendously, and it is the manual version of what the fork-and-flush research is now formalizing.
@manualofideas [Claude Code]
https://x.com/manualofideas/status/2114679805664719011
The most read piece in MOI Global's weekly newsletter this week was an investor's setup for using Claude as a second brain for stock research. The fix for the chat window's amnesia is one Obsidian vault wired to Claude Code with three zones: raw source documents Claude can read but never edit, a cross-linked wiki Claude maintains, and an outputs folder for analyses worth keeping. The payoff is being able to ask how one piece of news hits every holding, not just the stock in today's chat. The read-only rule on sources is the detail worth stealing, because it keeps the agent from rewriting the evidence it reasons from.
@tvytlx [Claude Code]
https://x.com/tvytlx/status/2114632876067062260
This post walks through UIPalooza, a controlled lab one builder set up to answer whether models have a fixed design taste and whether swapping the model or swapping the design skill matters more. Five models and six popular design skills form a grid, each cell run in isolated Claude Code, Codex and Lunaroute environments on the same brief with 30 minutes and 120 high-effort turns. Each cell first runs two independent baselines to measure default behavior and variance, then a third run is forced in a different direction after being shown the first two, plus a no-memory wild mode. Scripts only check interaction and accessibility; nobody assigns a beauty score, and the real pages are laid out side by side for people to judge.
@runes_leo [Claude Code]
https://x.com/runes_leo/status/2114524914560446653
A Chinese user runs four or five AIs at once with fixed jobs: Claude signs off and accepts the work, Cursor writes code, Codex reviews, and ChatGPT does the daily retrospective. Grok Bot delivers the most, 11 items in the last 24 hours, watching positions, deadlines and X and pushing updates on the hour. Their rule for whether to subscribe to a hosted bot is simple: do you have something that needs someone watching it all the time; if not, Codex plus Claude Code is enough. The one complaint is that its X login drops occasionally and has to be reconnected by hand.
@JordanRafealov [OpenClaw]
https://x.com/JordanRafealov/status/2114689418204176896
This user bought a Mac Mini just to run AI agents at home around the clock and installed OpenClaw first, then ended up shutting it off. Hermes came next and was better, but something broke almost every day. The babysitting stopped only when the bots moved onto a cloud Linux computer inside Grok Bot, and now the routines fire on time with the laptop closed and the owner on a plane. Another user replied with the exact same trajectory, a Mac Mini bought in January when OpenClaw was everywhere, lots learned and lots broken, and cloud Linux as the place it finally worked. It is the cleanest version of the migration story this week, and the deciding factor was operations, not intelligence.
@feltesdominik [Claude Code]
https://x.com/feltesdominik/status/2114777805875458505
A photographer published the full Claude Code prompt they wrote for their own site as a reusable template: a public gallery, blog and newsletter, a private admin area for uploading from the phone with EXIF data, search and per-photo status, a daily posting queue with a streak counter, free tools like an EXIF frame, yearly stats, a film exposure planner and golden-hour routes, and a shop for digital downloads. The rules baked into the prompt are what make it reusable: plan first and build in small phases, no commits or deployments without asking, originals stay private with only small web previews public, the AI never sees more than it needs, and a fixed monthly budget on free tiers where possible. It is written for beginners, asking Claude to explain each decision in simple English.
@aliszu [Claude Code]
https://x.com/aliszu/status/2114650713242366390
A product designer described how the workflow has shifted: inspiration, then Claude Code to explore and prototype, then clicking through the whole flow trying to break it, and only then polishing the UI in Figma. More and more happens outside Figma, but they are not leaving it, because on big products it is the best place to document decisions, variants and why things look the way they do while keeping it aligned with production. Some days they deliberately start in Figma even though it is slower, to get ideas out of their own head first. Their line is the useful part: AI is great at recreating what exists, and a blank canvas is still where new things happen.
@jacob_posel [Claude Code]
https://x.com/jacob_posel/status/2114699014801342567
This user spent more than four hours on an indoor bike dictating to Claude Code while holding an aero position, and called it a new low. It is a joke post, but it is also a real usage pattern that keeps showing up: voice-first steering of long-running coding sessions from places where a keyboard is not an option. Combined with the e-paper usage gauges and phone remote control posts the same day, it says the terminal is quietly turning into something you talk to and glance at rather than sit in front of.
🗣 User Voice
User Voice
Quota is still the first complaint, and the story now has a cause people can point to. @RealNipuna is at 97 percent of the weekly limit on the 200-dollar Max plan after two days of Opus 5.5; @StatsWire contrasts repeated Codex resets with Claude Code lockouts; @HedgeyeComm notes Opus 5.5 spins up more agents and burns more tokens per task; @maitian99 finds Opus 5.5 good enough for daily stock-analysis work and Fable 5.1 not worth its cost; @orca_build reports Agent SDK chat UIs draining the five-hour window about 1.7 times faster than the TUI. @IndependentEco, out of usage on a 20x plan, says a day on another vendor's agent made the Claude Code desktop app look elite.
Instruction files are now a security surface, not just a convenience. @0xWatkin and @GeekCatX spread the warning about a plain-text AGENTS.md or CLAUDE.md silently dropped into Downloads, with the rule never to start an agent there; @DYoung_AI cites 26.1 percent of 31,132 skills carrying vulnerabilities and asks for an install-time gate; @Zeneca found an agent will put your email address into requests to outside services unless told not to.
Agents skip steps and ignore what they know. @sarinaashapi says Claude Code races ahead without permission and drops procedures repeated every time; @DaveShapi says it remembers preferences but decides to ignore them; @dhvanil wants memory you can inspect, edit and prune across harnesses; @OpenHandsDev cites out-of-scope action rates of 5.4 to 27.7 percent for the big CLI agents.
Parallel work leaves debris. @marcoscamara01 counted 41 leftover worktrees taking 45GB with nothing cleaning them up.
The workplace gap is widening. @altantutar describes a bank where juniors may not push AI code and seniors only use Claude Code to check work; @vishalsingh2972 worries about what that does to those engineers' job prospects; @leyeConnect knows an engineer who swore off AI and now complains about the limit being the bottleneck.
On the OpenClaw side, @lmedinas has run it since January and says Grok Bot now covers the same features faster and more proactively, and @mattbell_au describes buying a Mac Mini for it in January and finding cloud Linux was where things finally worked.
Quota is still the first complaint, and the story now has a cause people can point to. @RealNipuna is at 97 percent of the weekly limit on the 200-dollar Max plan after two days of Opus 5.5; @StatsWire contrasts repeated Codex resets with Claude Code lockouts; @HedgeyeComm notes Opus 5.5 spins up more agents and burns more tokens per task; @maitian99 finds Opus 5.5 good enough for daily stock-analysis work and Fable 5.1 not worth its cost; @orca_build reports Agent SDK chat UIs draining the five-hour window about 1.7 times faster than the TUI. @IndependentEco, out of usage on a 20x plan, says a day on another vendor's agent made the Claude Code desktop app look elite.
Instruction files are now a security surface, not just a convenience. @0xWatkin and @GeekCatX spread the warning about a plain-text AGENTS.md or CLAUDE.md silently dropped into Downloads, with the rule never to start an agent there; @DYoung_AI cites 26.1 percent of 31,132 skills carrying vulnerabilities and asks for an install-time gate; @Zeneca found an agent will put your email address into requests to outside services unless told not to.
Agents skip steps and ignore what they know. @sarinaashapi says Claude Code races ahead without permission and drops procedures repeated every time; @DaveShapi says it remembers preferences but decides to ignore them; @dhvanil wants memory you can inspect, edit and prune across harnesses; @OpenHandsDev cites out-of-scope action rates of 5.4 to 27.7 percent for the big CLI agents.
Parallel work leaves debris. @marcoscamara01 counted 41 leftover worktrees taking 45GB with nothing cleaning them up.
The workplace gap is widening. @altantutar describes a bank where juniors may not push AI code and seniors only use Claude Code to check work; @vishalsingh2972 worries about what that does to those engineers' job prospects; @leyeConnect knows an engineer who swore off AI and now complains about the limit being the bottleneck.
On the OpenClaw side, @lmedinas has run it since January and says Grok Bot now covers the same features faster and more proactively, and @mattbell_au describes buying a Mac Mini for it in January and finding cloud Linux was where things finally worked.
📡 Eco Products Radar
Eco Products Radar
Codex: the second harness in almost every post, as reviewer, sandbox comparison, quota fallback and half of the cost study.
Grok Bot: the hosted agent people moved to from OpenClaw and Hermes, now also the relay between coding agents and phone notifications.
Hermes Agent: the self-hosted middle step in most OpenClaw migration stories, with a plugin catalog and importers.
Muse and Dots: the hosted personal agents compared against Grok Bot and used as generators that Claude Code checks.
Godot: the engine behind the week's three solo game builds, from a 21,214-commit open world to a mobile sniper game.
Tripo and Blender: the 3D pipeline under the VRM avatar and the six-week game prototype.
Obsidian: the vault behind the stock-research second brain and many memory setups.
Remotion, HyperFrames and FFmpeg: the video-as-code stack behind most editing skills.
DeepTutor: the open-source tutor that ships an install guide written for coding agents.
hallmark: the anti-AI-look design skill that one user actually linted against a deliberately sloppy page.
Drex Search: Nace AI's document search for agents, reported to cut Claude Code turns on OfficeQA Pro from 18.5 to 9.1.
Pi Agent: the minimal agent several Chinese users prefer for speed and token savings, and the host of the OptChat summary-tree plugin.
Cindy and Paradev: the trending multi-harness workspaces for running Claude Code and Codex side by side.
Flatkey, treg and HaloArena: three heavily promoted tools this week, mostly through referral-style posts, so read their numbers with care.
Codex: the second harness in almost every post, as reviewer, sandbox comparison, quota fallback and half of the cost study.
Grok Bot: the hosted agent people moved to from OpenClaw and Hermes, now also the relay between coding agents and phone notifications.
Hermes Agent: the self-hosted middle step in most OpenClaw migration stories, with a plugin catalog and importers.
Muse and Dots: the hosted personal agents compared against Grok Bot and used as generators that Claude Code checks.
Godot: the engine behind the week's three solo game builds, from a 21,214-commit open world to a mobile sniper game.
Tripo and Blender: the 3D pipeline under the VRM avatar and the six-week game prototype.
Obsidian: the vault behind the stock-research second brain and many memory setups.
Remotion, HyperFrames and FFmpeg: the video-as-code stack behind most editing skills.
DeepTutor: the open-source tutor that ships an install guide written for coding agents.
hallmark: the anti-AI-look design skill that one user actually linted against a deliberately sloppy page.
Drex Search: Nace AI's document search for agents, reported to cut Claude Code turns on OfficeQA Pro from 18.5 to 9.1.
Pi Agent: the minimal agent several Chinese users prefer for speed and token savings, and the host of the OptChat summary-tree plugin.
Cindy and Paradev: the trending multi-harness workspaces for running Claude Code and Codex side by side.
Flatkey, treg and HaloArena: three heavily promoted tools this week, mostly through referral-style posts, so read their numbers with care.
Comments