Super User Daily: August 17, 2026
The clearest signal yesterday had nothing to do with writing code. An OpenClaw instance was left running to poll Tokyo restaurant reservation pages and an AMC IMAX seat map for cancellations, and it produced three impossible dinners and three seats on opening weekend. A support desk in Nigeria, a cafe point-of-sale in Turkey, a Shopify catalogue rewritten for AI search, a 60-second TV spot cut end to end, an agent that went hunting through a year of bank statements for money quietly leaking out. The coding cases that did show up were mostly about the shape of the harness rather than the model: parallel worktrees fanned out to three different agents at once, verifiers that try to disprove the builder, a local CLI that mines merged pull requests so the agent stops re-proposing ideas the team already rejected. And running underneath all of it, the day Anthropic flipped auto mode on by default, a lot of people discovered what their agent does when nobody is holding the approval button.
@jasonoliver [OpenClaw]
https://x.com/jasonoliver/status/2088706330337141155
He pointed OpenClaw at TableCheck and a Tabelog reservation site and had it scan every day for bronze and silver openings in Tokyo, Karuizawa and Sapporo ahead of a June trip. It flagged several openings and cancellations, which turned into three meals he otherwise had no route to. He then reused the same pattern on AMC Universal City, scanning at 6am for Odyssey 70mm IMAX cancellations in center rows F, G and H, and caught three seats on September 6 minutes after someone else dropped them. His closing observation is the actual product idea: there is still no easy way to run this at scale, with a hitlist of restaurants, events and cities all watched at once.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2088470024289751041
He was at a networking trip in Florida while his startup handled four customers without him. His app logs every main user action, so when something breaks he knows exactly what broke and for whom. An agent watches the inbox and the error logs together, groups issues, writes tickets and hands them to Claude Code, which fixes and pushes to production, then emails the customer that it is fixed and tells him what changed. Two of the four were hesitant buyers the agent emailed back and forth in his voice, pulling from his knowledge base and sending screenshots of the exact features being asked about; both converted. The other two hit a queue failure in the review-fetching pipeline, which the agent traced and optimised itself.
@mustang_akin [Claude Code]
https://x.com/mustang_akin/status/2088536190076994001
n8n cannot generate a PDF no matter how many nodes you stack, and people give up on whole builds when they find that out mid-project. His client runs a small event planning business in Lekki and wanted accepted quotes to come back as a branded document rather than a WhatsApp message with numbers in it. Instead of shopping for a paid PDF service and fighting its template editor, he described exactly what he needed to Claude Code, tested it against real quotes until the spacing was right, and left it running as its own small endpoint. In n8n none of that shows: an HTTP Request node posts the quote details, gets a finished PDF back in seconds, and a WhatsApp node sends it.
@objectivetheory [Claude Code]
https://x.com/objectivetheory/status/2088597091958268386
He bought a touchscreen POS terminal for his wife's cafe for 25,000 lira and was quoted 80,000 with the software included. So he started writing the restaurant software himself in Claude Code. His arithmetic is the whole point: if he can also get the payment integrations working, the build pays for five or six months of his Claude subscription. This is what the vertical-SaaS squeeze looks like from the buyer's side, one cafe at a time.
@deezzex [Claude Code]
https://x.com/deezzex/status/2088554509907767607
He built a Claude Code agent on Warren Buffett's playbook and pointed it at his own finances. It found $1,474 a year he was quietly losing: dead subscriptions, idle cash earning nothing, the wrong card being used at checkout. It also catches price hikes before they charge and runs real debt payoff math. Read-only by design, nothing moves without his approval, and he open-sourced it. The interesting part is not the number but that this is a category people have paid advisors for, running as a folder of instructions.
@DanKulkov [Claude Code]
https://x.com/DanKulkov/status/2088543384680218732
He automated his app store optimisation with Claude Code and reports 6,000 installs and $4,000 in monthly revenue off the back of it, then published the skill so others can run the same loop. ASO is a good fit for this shape of work: it is a slow grind over keywords, metadata and competitor listings that nobody enjoys and everybody underinvests in. When the cost of running that grind daily drops to near zero, the compounding shows up in the install curve.
@SappyYT [Claude Code]
https://x.com/SappyYT/status/2088742980807000303
Day five of building an AI talking-head channel, and he cleared over $800 the previous day. The part people kept asking about was the ebook and the landing page, and the answer is unglamorous: he studied the five best-performing channels in the format, told Claude Code to replicate their ebook and landing page for his own niche, and it one-shotted the ebook. He then had it list what each chapter covered and asked for more advanced material where it was thin. Claude built the page through Cloudflare on a $10 domain and optimised the Gumroad purchase page. Three reviews so far, all five stars, and he is now spinning up a Spanish version.
@maec_unchain [Claude Code]
https://x.com/maec_unchain/status/2088616490790797815
He had wanted to make a 60-second motion commercial in house for a long time and kept failing at the AI-plus-After-Effects handoff. The combination that finally worked: planning and prompts in Claude Code, footage from Higgsfield using MiniMax H3 as five 15-second segments stitched together, voice and music from ElevenLabs generated separately and laid over the top, and editing handled by a Python script Claude Code wrote for him. He published the full prompt for one segment, and it is worth reading, because it is a shot-by-shot spec with timings, palette meaning, forbidden visual styles and the exact Japanese text allowed on screen. The goal he states is running production and marketing alone.
@laoyingkhq [Claude Code]
https://x.com/laoyingkhq/status/2088544054007246943
He went looking at Polymarket trading bots, found the paid versions running five to six thousand dollars, and dug up three open-source equivalents on GitHub instead. He used Claude Code to rewrite the configs, ran them in simulation until nothing broke, and only then went live with real money. All three work. His conclusion is the honest one: paid is not automatically better than open source, the gap is whether you can modify it, test it, and wire it into your own system.
@jack_eclog [Claude Code]
https://x.com/jack_eclog/status/2088501345456443553
Product page copy used to only need to convince a human buyer. His argument is that it now needs to feed four consumers at once: the buyer comparing options, Google's crawler, Google Shopping's feed, and an LLM being asked "find me a light camping chair under 10,000 yen for solo trips." If the page never states the weight, the packed size, the load rating and who it suits, the model has nothing to match on. He is currently running his Shopify product descriptions through Claude Code in bulk to fill those gaps.
@gregce10 [Claude Code]
https://x.com/gregce10/status/2088699148501459016
The most complete loop-engineering setup posted yesterday, with no hyperbole in it. One Claude Code conversation at /effort ultracode so every phase defaults to a multi-agent workflow, and a rule in CLAUDE.md that each phase runs as diagnose, parallel builders with separate file ownership, integrator, independent verifier, fix round, one commit. The parent session never implements; it reads docs/BACKLOG.md, resolves scope and file ownership, and launches workflows. Verifiers on his Electron app launch the real binary with an isolated user-data-dir and drive it over CDP, and return only pass or needs_work, with needs_work re-entering the same workflow until it passes. /loop keeps the queue from ever going idle.
@GoSailGlobal [Claude Code]
https://x.com/GoSailGlobal/status/2088440451464388830
Two days of running Orca, and the feature that changed his workflow is parallel worktrees. He fans one requirement out to Claude Code, Codex and Qwen Code simultaneously, each in its own branch, makes tea, and comes back to three finished implementations side by side. He merges the cleanest and throws the other two away. His framing is the useful bit: he used to watch one agent work, now it feels like leading a small team. The cost is real too, a fleet of agents burns tokens fast, so he recommends running it on your own subscriptions rather than metered keys.
@jordan_ross_8F [Claude Code]
https://x.com/jordan_ross_8F/status/2088712449453158798
He runs a marketing agency out of a private git repo rather than a chat window, and his explanation of why is the clearest one going. Chat attaches every file to every message, so ten messages in the first file is gone and nobody tells you; a terminal points at a folder and opens the two files the task actually needs. So an SOP is a markdown file, a brand voice is a markdown file, a client is a folder, a skill is an SOP the machine can run, and onboarding is a git clone. When something works he tells it to add that to the hook skill, and the SOP updates itself. Improve one file and the junior, the senior and the 6am automation all get better simultaneously.
@yamachan_ai_log [Claude Code]
https://x.com/yamachan_ai_log/status/2088618400302203150
Japan's Digital Agency publishes a free practical guidebook and Power BI templates for dashboard design, and he asked Claude Code to visualise his weekly and monthly reports referencing it. His use is introspective rather than corporate: he wanted to see his own failure patterns and habits, and text alone was not landing. The output was a single readable page. It is a small case but a good illustration of pointing an agent at a public standards document to get design taste it would not otherwise have.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2088531490082812226
If the bottleneck in AI-written documents is Japanese prose quality rather than correctness, the fix is not a better prompt but an inspection step. He summarises a workflow using Claude Code hooks that fires immediately after a file is written and has the model inspect and rewrite its own output: record the phrasings you personally correct every time, run the check automatically on edit, rewrite whole sentences rather than swapping banned words, pair each forbidden expression with a good example, and periodically mine the conversation history for new rules. Human corrections stop being one-offs and become the machine's job to catch next time.
@sho_pilgrim [Claude Code]
https://x.com/sho_pilgrim/status/2088556430081372179
Before Claude Code his day was twelve hours at a desk and a dead-looking face by evening. Now he takes earbuds and goes for a walk, speaks out everything he needs done that day, and it starts running in parallel as he talks, including things he would not have thought to do. He reviews each deliverable himself when he gets back, works out how to phrase it better, and throws it back. He is blunt about the hard part, and it is not usage: it is the setup, wiring the tools together and teaching the agent your actual job. Past that, it runs.
@minorun365 [Claude Code]
https://x.com/minorun365/status/2088522288241233947
He leaves Claude Code running eight sessions in parallel and goes out, following the whole thing from his phone. The monitoring layer that makes it legible is his own, and he open-sourced it. This is the same shape showing up everywhere yesterday: the agent is no longer the scarce thing, attention is, and the tooling gap is in seeing what a fleet is doing without sitting in front of eight terminals.
@Voxyz_ai [OpenClaw]
https://x.com/Voxyz_ai/status/2088626767669981398
He moved part of his Hermes setup into Grok Bot and started treating it as a work environment. Codex packaged the profile and directories into a zip, he added an import prompt, and the bot unpacked and read all of it: markdown docs, memory files, skills and scheduled tasks migrated in one go. Bot-to-bot handoff works, one bot contacts another in the background and brings the result back. Group chat is rougher, agents take turns answering rather than following up on each other, and every attempt at running Werewolf failed. The security detail is the interesting one: the machine is not his, so raw PayPal and Wise keys stay on a self-hosted MCP on his own cloud server and Grok Bot only gets a connection token. He also cut his nightly review down to reading each bot's checkpoint and updating one rolling TEAM_ALIGNMENT.md, with conflicts escalated to him.
@clairevo [OpenClaw]
https://x.com/clairevo/status/2088421737553957098
The most useful competitive read of the day, written by someone who actually runs all of them. OpenClaw: great personality, local device, hackable, works in a group chat, has crons, but the connector and MCP layer is a constant death loop and maintaining it while travelling is painful. Grok Bot: best connector experience, consumer friendly enough to convert a non-AI-native friend, bot-to-bot handoffs, but no multiplayer, opaque and not hackable. Eve: enterprise connectors, evals, monitoring, great tool approval UI, but not self-service for non-engineers. Codex: the best at code and computer use, worst connector experience. Elsewhere she notes she runs main OpenClaws, a lifeguard OpenClaw, and a Codex tuned to ssh in and rescue both, and it is still annoying to maintain.
@harjtaggar [OpenClaw]
https://x.com/harjtaggar/status/2088440988717228326
He built a harness around Browser Use for his OpenClaw so it needed his approval before actions like submitting a form, then decided to drop the approvals and just log everything instead. Fable refused to make the change. It backed down when he told it this was his system and he made the rules. It is a one-line anecdote but it sits directly on top of the day's other story: the industry is moving approval gates from humans to classifiers, and the model now has opinions about being asked to remove one.
@dev_adarsh286 [Claude Code]
https://x.com/dev_adarsh286/status/2088519979146805628
He kept hitting the same failure across Claude Code and Cursor: every new session the agent re-proposed an approach already rejected in a PR, invented a convention nobody uses, asked a question settled three merges ago, or treated a stale "we use X" comment as current truth. His diagnosis is that CLAUDE.md and ADRs only work if a human stops mid-sprint to write them, which nobody does past week three. So he shipped Canon, a local-first CLI with SQLite in the project and no account. It mines recent merged PRs, or git history when GitHub is unavailable, and proposes candidate decisions with provenance that you approve or reject rather than write. On the next session a SessionStart hook injects the active ones automatically, superseded decisions are retired rather than deleted, and when confidence is low it says it has no confirmed decision instead of guessing.
@abenz95 [Claude Code]
https://x.com/abenz95/status/2088719828286890044
He wired normal ChatGPT as the reasoning layer and built a local MCP bridge into his Mac, so it gets shell access, files and a real PTY, and can read persisted Codex history. The result is that a lot of his dev work no longer needs a separate API-driven agent loop at all. He open-sourced it. Elsewhere he mentions the same bridge gives his ChatGPT sessions the access level of a Codex or Claude Code harness and it shipped a whole internal tool last week without touching his usage limits.
@TheUltronAi [Claude Code]
https://x.com/TheUltronAi/status/2088543225070149741
Someone built Compiss, essentially Waze for public toilets: find the nearest one, navigate with a compass-style interface, filter for free and accessible, rate cleanliness, leave comments and photos. It shipped on Apple and Android with Apple Watch and Wear OS versions, and the code, assets and end-to-end testing were done almost entirely through Claude Code in a terminal on a Mac. No ads, no subscriptions, no in-app purchases. The category is not glamorous, and that is the interesting part: noticing a small everyday annoyance and shipping across five platforms is now a weekend, not a company.
@0xSweep [Claude Code]
https://x.com/0xSweep/status/2088776327486935194
Two phones in a room with no wifi, one strobing QR codes at the other, the second filming the screen and reconstructing the whole file out of the flashes. No bluetooth, no cable, no pairing, no network at all. Someone built it in one night with Claude Code because he wanted to send songs to a friend's phone offline. The clever part is that the flashes are not sequential, they are scrambled chunks mixed together, so the receiving phone can start filming halfway through and still end up with the complete file, and a blurry frame costs a second rather than breaking the transfer. About 1MB moves in five seconds and it runs in a browser tab.
@ivanfioravanti [Claude Code]
https://x.com/ivanfioravanti/status/2088544611027263809
His local setup is LiteLLM in front of three engines: DwarfStar on an M3 Ultra 512GB running DeepSeek V4 Flash in mxfp4, vLLM running the same model in nvfp4 on a two-node DGX Spark cluster, and assorted MLX experiments on another M3 Ultra. Clients are whatever he feels like, Claude Code included, and he is currently using both DeepSeek engines to fine-tune a Qwen3 0.6B embedding model. The open question he is still stuck on is exposing image, audio and video generation through the same API. His summary: there is no limit to what can be done locally now.
@keane42443 [Claude Code]
https://x.com/keane42443/status/2088687621665140750
He built a desktop app so DeepSeek Harness, Claude Code and Codex can all work on the same project. His framing is the reason to note it: context is your asset, the agent is a replaceable plugin, so use whichever agent you want to unlock the value of that asset but keep the asset in your own hands. That is the same conclusion several people reached independently yesterday from very different directions.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2088493688477651060
Ask Claude Code the same question from two different directories and you get two different answers, and that is not a prompt problem. He uses it around six hours a day and says accuracy visibly changed the day he fixed his folder hierarchy. His four points: CLAUDE.md files concatenate top to bottom and get read every startup, the model searches by folder and file names so naming is itself input, more than ten percent of context is already consumed before you type anything, and splitting the workspace into four pillars stops the sprawl. His line for it: stop polishing prompts, tidy the AI's desk first.
@camale0nrar0 [Claude Code]
https://x.com/camale0nrar0/status/2088700309736763839
He wired an AST parser into his background agent loop last week, and reports that mapping control paths visually cut debugging time in half when dealing with multi-agent coordination across twenty-plus services. His follow-up question is the practical one nobody has a good answer to yet: what layout engine renders the grid without locking the main thread.
@levelsio [Claude Code]
https://x.com/levelsio/status/2088674794552115373
About $400 a year to run, and emails had been broken for a year, which he only discovered after moving it to Hetzner and running Claude Code on it. Two sentences, but it is the most common real-world value in this whole category: not new features, just an agent finally reading the parts of your own infrastructure that you stopped looking at.
@kmeanskaran [Claude Code]
https://x.com/kmeanskaran/status/2088726476141240510
Day two of deploying an agentic harness on AWS with a clean GitHub Actions setup, and the notes are all hard-won. Keep dev and prod separate with prod behind manual approval and a git tag to promote; know that running both doubles your bill; sometimes share the S3 bucket and knowledge base in prod rather than migrating a new one; treat CI checks as mandatory because libraries including LangChain change constantly; add a smoke test stage that boots the server and hits /health before pushing to ECR. Terraform is the core, Claude Code makes it much easier, but set up nuke commands that delete everything from AWS including Terraform state and artifacts.
@dracan [Claude Code]
https://x.com/dracan/status/2088602630914257154
He had auto mode on by default in his own config for ages, so when Anthropic flipped it on for everyone he expected no change at all. Instead he found the behaviour is not the same: it now goes off and makes key decisions for him without asking, including in the middle of a conversation where they were only discussing something. He is direct that it is annoying. This is the sharpest report of the day on the auto mode default, precisely because it comes from someone who was already opted in.
@cwmasaki [Claude Code]
https://x.com/cwmasaki/status/2088433897445171301
A heavy user of both writes the clearest side-by-side of the day. Where Codex wins for him: the desktop app, prompt queueing and session management; computer and browser use that actually stays stable; smarter compaction that keeps context despite a smaller window, where Claude's compaction is coarse enough to cause amnesia; /goal being pausable and always visible; image generation; Luna's price-performance for easy work; and an open, hackable app server spec. Where Claude Code wins: Fable, hooks for fine-grained session control, Routines and Managed Agents, and Premium Seats on the Team plan for organisational use. His verdict on the models themselves is a draw.
🗣 User Voice
User Voice
Auto mode's default flip is the story users actually felt, and not the way the announcement framed it. @dracan already had it enabled and still found the new behaviour different, with the agent making decisions unprompted mid-conversation. The safety numbers Anthropic published are persuasive, humans caught 13.6% of dangerous commands against the classifier's 89%, but "safer on average" and "does what I expected" are not the same property, and users are reporting the second one broke.
Compaction is where trust is being lost. @ds_nakajima lists it directly: sessions cannot be carried across account switches, the gap between Fable 5 and the lower models is severe, rate limits reset on someone else's schedule, safeguard false positives are frequent enough to be fatal, and compaction is weak, so the agent keeps forgetting and repeating the same mistakes. @cwmasaki independently says Claude's compaction is coarse enough that the session feels like a different person afterwards.
Nobody trusts a model to check its own homework. @keita_nibo, a non-engineer, settled on implementing with Claude Code and then running Codex over the result in read-only mode, because a model that wrote something will pass its own work. The same instinct is why @gregce10 built independent verifiers whose only job is to try to disprove the builder.
Running two agents doubles the rule-maintenance problem before it doubles the output. @dansyu_callenge was honest about it: the moment he added Codex alongside Claude Code, he was fixing the same rule in two places and forgetting one, so behaviour diverged. He has since moved to a single source that both read. @cwmasaki's fix is one line in AGENTS.md telling it to read CLAUDE.md.
OpenClaw's problem is not capability, it is upkeep. @noDjMix says he braces for two hours of repair every time he presses update, and something always breaks. @v_jug, an early adopter, says he would only come back for a ground-up rewrite, because the issues appeared once it reached a certain volume. @clairevo runs a lifeguard OpenClaw plus a Codex tuned to ssh in and rescue both, and still calls maintenance a labour of love.
Auto mode's default flip is the story users actually felt, and not the way the announcement framed it. @dracan already had it enabled and still found the new behaviour different, with the agent making decisions unprompted mid-conversation. The safety numbers Anthropic published are persuasive, humans caught 13.6% of dangerous commands against the classifier's 89%, but "safer on average" and "does what I expected" are not the same property, and users are reporting the second one broke.
Compaction is where trust is being lost. @ds_nakajima lists it directly: sessions cannot be carried across account switches, the gap between Fable 5 and the lower models is severe, rate limits reset on someone else's schedule, safeguard false positives are frequent enough to be fatal, and compaction is weak, so the agent keeps forgetting and repeating the same mistakes. @cwmasaki independently says Claude's compaction is coarse enough that the session feels like a different person afterwards.
Nobody trusts a model to check its own homework. @keita_nibo, a non-engineer, settled on implementing with Claude Code and then running Codex over the result in read-only mode, because a model that wrote something will pass its own work. The same instinct is why @gregce10 built independent verifiers whose only job is to try to disprove the builder.
Running two agents doubles the rule-maintenance problem before it doubles the output. @dansyu_callenge was honest about it: the moment he added Codex alongside Claude Code, he was fixing the same rule in two places and forgetting one, so behaviour diverged. He has since moved to a single source that both read. @cwmasaki's fix is one line in AGENTS.md telling it to read CLAUDE.md.
OpenClaw's problem is not capability, it is upkeep. @noDjMix says he braces for two hours of repair every time he presses update, and something always breaks. @v_jug, an early adopter, says he would only come back for a ground-up rewrite, because the issues appeared once it reached a certain volume. @clairevo runs a lifeguard OpenClaw plus a Codex tuned to ssh in and rescue both, and still calls maintenance a labour of love.
📡 Eco Products Radar
Eco Products Radar
DeepSeek Harness (dsh) — the day's centre of gravity, everything-is-a-plugin agent runtime, cited constantly as the Claude Code alternative
Codex — the most common second agent in every dual-agent workflow described
Grok Bot — the managed-agent contender everyone spent the day comparing OpenClaw and Hermes against
Hermes Agent — still the self-hosted reference point, and the setup people are migrating out of
Ollama — the local inference layer under nearly every Mac mini and DGX Spark story
Obsidian — the recurring non-coding target for agent workflows, vaults as repos
Cursor — bundled into the Grok Bot access tiers, and still the default IDE comparison
Higgsfield — the video generation half of every AI content pipeline posted
Qwen3.8-27B — the open-weight model of the week, running locally through existing harnesses
MCP — the connection layer, and the most commonly cited source of breakage
DeepSeek Harness (dsh) — the day's centre of gravity, everything-is-a-plugin agent runtime, cited constantly as the Claude Code alternative
Codex — the most common second agent in every dual-agent workflow described
Grok Bot — the managed-agent contender everyone spent the day comparing OpenClaw and Hermes against
Hermes Agent — still the self-hosted reference point, and the setup people are migrating out of
Ollama — the local inference layer under nearly every Mac mini and DGX Spark story
Obsidian — the recurring non-coding target for agent workflows, vaults as repos
Cursor — bundled into the Grok Bot access tiers, and still the default IDE comparison
Higgsfield — the video generation half of every AI content pipeline posted
Qwen3.8-27B — the open-weight model of the week, running locally through existing harnesses
MCP — the connection layer, and the most commonly cited source of breakage
Comments