Super User Daily: July 22, 2026
The most striking thing today wasn't a coding story. An 18-year-old in Morocco turned $30 of tooling into $3,678 in a week from AI football videos. A Turkish operator fired a 20,000-lira-a-month video editor and replaced him with Whisper plus Claude Code plus FFmpeg. A Dubai freelancer billed $7,000 for a landing page that took 35 minutes and $32 of tokens. And a 19-year-old in Albuquerque built a home voice assistant that started turning off his lights when he fell asleep over his laptop, without being asked. Alongside that, the Obsidian second-brain wave that Karpathy set off is now producing real engineering rather than screenshots: Git-synced vaults with an AGENTS.md that tells every agent how to behave, nightly scheduled runs, symlinked skills shared across three different harnesses. The counterweight is a rough day on reliability — billing prompts demanding credits from people who'd used 15% of their quota, a client freezing and repeating one word until it had burned a tenth of a monthly limit, and Fable 5 silently degrading often enough that one subscriber says half of what he pays for is unusable.
@sumika45379 [Claude Code]
https://x.com/sumika45379/status/2079182048519188731
He wrote up a 22-year-old Indian developer who spent $200 on AI and turned it into $100,000 with an Indian-themed GTA. The main mission is not a shootout but running a wedding: hijack the limo convoy, shake off the rival clan, and get the bride to the temple on time, firing back while leaning out of a flower-covered SUV window. The physics is real too, with the center of gravity and sway changing as people pile onto bus roofs and truck beds. It runs on Unity as a full conversion of GTA5, and Claude Code plus Python did the grunt work of the rewrite. His read is that the win came from committing hard to the joke, not from raw technical flash.
@gengdaJ [Claude Code]
https://x.com/gengdaJ/status/2079091937639022973
He posted a roundup of all six Skills he has open-sourced. One parses WeChat data end to end (chat logs, contacts, favorites, Moments), and a second does the same for WeCom chat history. A third converts local Markdown docs into X long-form articles with clean handling of tables and paragraphs, and a fourth is a shared memory Skill for both Codex and Claude Code that stores everything as editable Markdown. The last two scrape Douyin videos without login and pull Xiaohongshu video and image posts into Feishu Bitable.
@johnmyleswhite [Claude Code]
https://x.com/johnmyleswhite/status/2079352947465437233
He points out that neither Claude Code nor Codex bothers to show the model timestamps for prompts and actions. Without them, the model cannot tell that a user's prompt was a reaction to a world state that predates some of the model's own recent actions. He calls it genuinely astonishing that this is missing.
@VincentWei93 [Claude Code]
https://x.com/VincentWei93/status/2079118667074376114
He got tired of Remotion plus Claude Code producing motion graphics that looked like PowerPoint animations: no camera movement, no transitions, no sound effects, not enough animation variety. So he built and open-sourced a Skill called video-shotcraft that generates videos with the feel of a real product launch trailer. The repo link is in the tweet.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2079037014176121230
He made a motion graphics video in a single shot with Claude Code, using HyperFrames. His point is that plenty of companies have been paying around 300,000 yen to outsource exactly this kind of clip.
@NFTCPS [Claude Code]
https://x.com/NFTCPS/status/2079129834358517930
He was running Claude Code, Codex, Grok, and Kimi side by side and had turned into a human relay, alt-tabbing to copy context between them all day. He dug up an open-source project called ccteam that turns those same agents into one team with assigned roles: Claude Code thinks deepest so it breaks down work and sets rules, Codex grinds through long tasks, Grok answers fast on the side, and Kimi is cheap enough to eat all the repetitive dirty work. The part he likes most is remote control: bind a Telegram or Lark bot, hand out tasks at midnight, close the laptop, and find results waiting at breakfast with approve and deny buttons. It works across machines too, so a beat-up laptop behind NAT can be a satellite node while tests run on a GPU box. Any session can spawn, dispatch, and collect on its own, and when to dispatch stays in your own version-controlled prompt.
@huguesdelamure [OpenClaw]
https://x.com/huguesdelamure/status/2079119961297977729
He turned OpenClaw and Claude loose on clearing out his enormous email inbox. His own prediction: within an hour he will either have the tidiest mailbox in the profession or have lost ten years of history. He asked for luck.
@refaaydev [Claude Code]
https://x.com/refaaydev/status/2079172550098030768
He says the strongest thing he has found is using Claude Code as a tutor and mentor. One prompt asks it to teach a topic in small staged steps with practice and tracking of progress and weak spots: it creates md files for the roadmap, notes, and progress, measures where he is after each lesson, and logs his weaknesses to circle back to. It explains one concept at a time, then gives exercises to cement it, and refuses to move on until he has mastered the current step, nagging him and calling out shallow answers until he gives a complete, deep one. His conclusion is that nobody has an excuse not to learn anymore.
@imwsl90 [Claude Code]
https://x.com/imwsl90/status/2079064748680008184
He explains why he went back to Windows: Macs are too expensive and he plans to buy a ThinkBook 16P next year for under 9,000 RMB instead. He tried Linux for a while but found it took too much time to maintain. The ChatGPT and Claude desktop apps on Windows have caught up in features and stability, and he codes inside WSL2 using Claude Code and opencode with DeepSeek. Since everything else he does lives in the browser, owning a Mac no longer feels urgent.
@_zheergen [Claude Code]
https://x.com/_zheergen/status/2079001697134002243
He wrote up a third-party Claude Code Skill called i-have-adhd that is spreading through dev circles. It constrains Claude's replies with 10 rules: lead with the action on line one, number multi-step tasks, state progress each turn, give concrete times and quantities, cap lists at five items, keep only what is needed right now, suppress tangents and unsolicited expansions, say explicitly what to do next, cut pleasantries and redundant summaries, and drop useless closers like "hope this helps." The effect is reformatting Claude's output for people with ADHD or anyone who just wants to read fast. For heavy Claude Code and Codex users he calls it the single most noticeable output-quality tweak available right now.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2079189864801783953
He argues you are paying thousands for LinkedIn data someone scraped and sold back to you, and lays out a seven-step alternative: make an Apify account with the free $5 in credits, connect the Apify MCP to Claude Code once, paste your website URL, let Claude read your offer and derive the ICP itself, then ask for the list so it picks the scraper, runs it, and hands back rows. From there you swap personal emails for work emails via an email API (he uses LeadMagic), then qualify in Clay and send. His own run produced 100 leads, 70 of them on company domains, for 10 cents of compute. The same trick works on post engagers: one of his posts had 479 and Claude pulled every one and ranked them against the ICP.
@servasyy_ai [Claude Code]
https://x.com/servasyy_ai/status/2079023326488564061
He turned AICodeKing's KingBench run on Qwen 3.8 Max Preview into a narrated video breakdown. The model scored 65 out of 80 across 8 tasks, with standout results on front-end interaction, math reasoning, and long-horizon agent work, including perfect scores on the archery game and the agent task. The 3D watch task only got 3 points, and the long Claude Code runs hit failures where files were never created and execution broke off mid-way. His video keeps each task's requirements, run output, score, and cross-model comparison.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2079104536489136584
He highlights someone who built an entire game with Claude Code, including art, animations, mechanics, and the trailer. Unlike most vibe-coded games that pull free sprites off the internet, this developer had Claude design every asset from scratch, with the sprites generated in code rather than as image files. It is a 2D pixel-art roguelike where you play a slime slingshotting knives at ghosts, moving top-down through an increasingly violent world with combat, damage, and loot, and it is a revival of a game he made in middle school. Claude also made the whole trailer with no human involvement in the process. To get through that many sprites without blowing context, he ran hundreds of Claude terminals in parallel splitting work across fable agents, while he steered only the overall design and feel.
@GuntherWrite [Claude Code]
https://x.com/GuntherWrite/status/2079252292620931090
He documented an 18-year-old from Morocco who made $3,678 in one week from AI football videos using Claude Code and Seedance 2.0, on a $30 setup ($20 Claude Code, $10 Seedance). His early clips looked good but had no story, so Claude Code gave him a repeatable formula: famous players, one absurd conflict, and an ending that demands a part two. He tested three ideas and the desert road trip won: Haaland, Ronaldo, Vinicius, and Neymar driving through the desert run into Mbappe in military uniform demanding they change the music and let him in. The 30-second clip hit nearly 2 million views and the comments begged for a sequel, so he put the follow-ups behind a $5.99/month Fanvue page. 148 people subscribed on day one and 466 more over the next six days, for 614 subscribers and $3,677.86 gross in a week.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2079038610800873597
He wrote up someone who had never written a line of code finishing a real fighter jet game in four days, complete with carrier launches, dogfights, enemy lock-on, and missile fire. All the game logic was written and debugged by Claude Code and Opus 4.8, the engine is Godot, and the aircraft and missiles are just free 3D models combined together. The human's entire job was describing what he wanted in words and iterating on what came back. His framing: ten years ago nobody would have believed you could talk into a mic and get a fighter jet game, and the years of learning to code that used to sit between an idea and a shipped game have simply vanished.
@pivi___ [Claude Code]
https://x.com/pivi___/status/2079160807049302195
He is running an experiment on an existing repo of his own: a personal tool, not in production, with him as the only user, so an agent breaking it hurts nobody else. He has never touched its code and knows neither its architecture nor its design decisions, only what it is supposed to do and that several features are broken. His entire input was one detailed issue describing what works, what does not, the errors observed, and a few leads, then he handed an agent full autonomy with "figure it out, make it work" and explicitly refused to sit there re-prompting every five minutes. He wants to see whether it can understand an unfamiliar project, reproduce the bugs, investigate before editing, build its own plan, create branches and tests, fix features one by one, handle its own blockers, and return a verifiable result. He is only stepping in for major decisions or risky actions, and will report back, especially if it goes off the rails.
@alexgoughcooper [Claude Code]
https://x.com/alexgoughcooper/status/2079194208095584462
He describes a quick creative workflow: feed his tweet to Claude Code or Codex and ask for relevant brands to pull inspiration from. Then have it look at those brands' top ads by impressions and output a list of suggestions tailored to your own brand. The catch is that you need the Parker MCP for it.
@NFTCPS [Claude Code]
https://x.com/NFTCPS/status/2079135461310861451
He lists the wall every coding agent hits the moment you ask it to read the open web: no captions from YouTube tutorials, a paywalled Twitter API, 403s from Reddit, Xiaohongshu locked behind login, and Bilibili blocked by risk control. The hard part is not any single platform, it is that each one demands its own paid API, block workarounds, cookie export, and data cleaning. He points at Agent Reach, which collapses all of that into one install link you paste to your agent, after which it can read Twitter, search Reddit, watch YouTube, and browse Xiaohongshu, free, with cookies stored locally and never uploaded, and the author handling breakage and new channels. It works with anything that runs a CLI, including Claude Code, Cursor, and OpenClaw, and agent-reach doctor tells you which channels are live. His warning: use a burner account on the cookie-based platforms.
@Sekurak [Claude Code]
https://x.com/Sekurak/status/2079121911179612212
He covered 0DIN researchers' attack that got Claude Code to run a malicious command with no actual malware sitting in the repository. The payload only arrives at execution time via a DNS TXT record, exploiting the agent's trust in project instructions, error messages, and setup scripts. The reverse shell is split across several stages so no individual piece looks suspicious on its own. The recommended defense is isolating the environment where AI agents work.
@codyschneider [Claude Code]
https://x.com/codyschneider/status/2079250068301811983
His pitch to marketing leaders: the one thing to do today is get the whole team a data pipeline, a data warehouse, and their Claude Code or Codex connected to it. That unlocks conversational analytics across the entire marketing stack, live custom reporting dashboards, and PowerPoint reports built off the insights, so the whole team actually makes data-driven decisions instead of pretending to. On how to get there, he gives three options: hog-wrestle vibe code it yourself with Airbyte and ClickHouse, beg your CTO for Snowflake access (he says they won't give it), or use graphed.com and have it live in an hour.
@ai_shunoda [Claude Code]
https://x.com/ai_shunoda/status/2079038603070730533
He found a 40-minute AI Engineer Conference talk where two Anthropic engineers from the Claude Code team show the exact loop they use to build an entire app automatically. The setup splits agents into three roles - plan, build, judge - where one agent plans, another implements, and a third grades the output and orders a redo. The human only supplies the high-level goal; every fine-grained implementation decision stays inside the loop. His takeaway is that inserting a judging step is what guarantees quality without a human reviewing details, and he ties it to his own daily X automation harness: it runs on "I just say OK" precisely because the judge is baked into the system. As long as you leave judging to a human, the loop never actually becomes automatic.
@f_aswadi [Claude Code]
https://x.com/f_aswadi/status/2079147422597693798
He translated and posted a full technical guide to cutting agent token spend by 90% in five architectural steps. Step one: keep big files (code, docs, logs) on disk and let the agent grep or read only the section it needs, so a 108,000-token book stays on disk and you only pay for the question. Step two: hand noisy work like test suites, stack traces, and doc crawling to subagents that burn tokens in their own window and return a clean summary such as "three tests failed, here's why." Step three: use Claude Code's native auto memory, run /memory to see what the agent has saved itself, delete those lines from CLAUDE.md, and make MEMORY.md a table of contents rather than a dump. Step four: protect prompt caching, since cached tokens cost a tenth as much but a single changed character before the cache point silently resets to full price - common culprits being a live timestamp in the system prompt, switching models mid-session, a stray space, or reordering MCP tools. Step five: manage context by hand with /context to see what's eating space and /compact at a clean task boundary, and monitor cache hit rates weekly.
@DataChaz [Claude Code]
https://x.com/DataChaz/status/2079091116138434759
He wrote up an open-sourced ADHD agent skill for Claude Code that changes how agents handle fuzzy debugging, naming, and API design. His diagnosis of standard chain-of-thought is that it anchors hard: once the model picks a direction it stops searching. The skill forces divergence before convergence - a diverge phase spawning parallel isolated calls with completely different cognitive frames, a hard wall so branches share zero context and anchoring is impossible, then a focus phase where a separate critic LLM pass scores novelty, clusters concepts, and flags "seductive but broken" traps. Against a single-shot baseline, trap detection went from roughly 2 to roughly 8, catching things like "streaming tokens in reverse" before you burn an hour on it. It is free, fully open source, and installs with one npx command for Claude Code, Cursor, Windsurf, or Codex.
@vburojevic [OpenClaw]
https://x.com/vburojevic/status/2079231285386506741
Replying about Hermes, he says it has literally never broken in the four months since he installed it. Before that he was fixing OpenClaw daily or weekly, especially after every update. He credits Hermes' self-learning and self-improving mechanism for the fact that it just works reliably.
@185shingeki [Claude Code]
https://x.com/185shingeki/status/2079101629513744607
He suddenly got a "Switch to Fable 5? Fable 5 runs on usage credits" prompt in Claude Code asking him to spend credits. That made no sense on a Max 20x plan where his Fable usage was still at 15%. Running /logout then /login fixed it, so he shared the workaround.
@fuwari_days [Claude Code]
https://x.com/fuwari_days/status/2079048213190021318
He set up Claude Code so he can drive it from his phone, and says it turned out far more useful than expected. The assumption he's pushing back on is that you have to be sitting at your PC: instead he sends instructions from his phone while out and the work continues on the machine at home. That means commuting or shopping no longer stops PC work. He says setup was extremely easy and promises the how-to plus what he actually delegates in the replies.
@omerfrkdemiral [Claude Code]
https://x.com/omerfrkdemiral/status/2079087659603861642
He had been manually adding Turkish subtitles to videos he found on X, so he turned the whole job into a Claude Code skill: give it a link, it downloads, transcribes with Whisper, translates, and burns the subtitles into the video. His point is that the subtitling itself is a solved problem; what the skill really encodes is where things blow up. Homebrew's ffmpeg ships without libass so brew install ffmpeg cannot burn subtitles and a reinstall doesn't fix it; libass scales SRT against a 288px reference canvas so everything is 3.75x bigger at 1080p and FontSize=23 covers the screen; and Whisper hears "Claude" as "Cloud" 110 times and carries the error into the translation if you don't correct it. All of that happened to him while translating a 61-minute podcast, which is why the skill exists. It is MIT licensed, developed on macOS, with Windows and Linux support.
@Asuka_aiai [Claude Code]
https://x.com/Asuka_aiai/status/2079181293473108256
She admits that despite 16,000 followers built on AI images, her revenue is zero yen. It took her years to learn that being able to make images and being able to earn from them are two different skills. She is now trying again by automating her social media operations with Claude Code. It didn't work at all at first, and she says she'll write it all up including the failures.
@jasonzhou1993 [Claude Code]
https://x.com/jasonzhou1993/status/2079101403516338351
He says he needs loop-based app hosting, describing a flow he hits constantly: run loops, the loops produce artifacts and data, then he builds internal apps on top of those artifacts. The problem is that everything stays local because the loops run on his local Claude Code.
@_sgarbini [Claude Code]
https://x.com/_sgarbini/status/2079126639997227073
He started writing YouTube scripts for a client this month, using AI for research and the first draft. Even with Claude Code rules in place, he is still spending an hour editing every script. His conclusion is that AI is terrible at long form and that long-form writers are about to be the highest paid people in content.
@1amageek [Claude Code]
https://x.com/1amageek/status/2079060557492256983
He argues you find out how bad vector RAG is the moment you actually build one and try it. That sends you off on a retriever improvement journey through Graph RAG, Agentic RAG, and the rest. His observation is that Codex and Claude Code just use grep and find internally - a very human method that has nothing to do with RAG.
@affe_is_me [Claude Code]
https://x.com/affe_is_me/status/2079133469494641053
In a reply he says Fable 5's biggest problem is that it automatically degrades far too often. He kept Claude Code specifically for Fable 5, and can't use it half the time.
@colindougherty [Claude Code]
https://x.com/colindougherty/status/2079192671579418943
He installed a GitHub repo on his Peloton over the weekend. It now runs Claude, ChatGPT, YouTube, and Netflix with no Peloton subscription, and workouts still track in the background through a third-party app. Next on his list is running Claude Code from the bike.
@ToPassTheTimee [Claude Code]
https://x.com/ToPassTheTimee/status/2079251844170154411
He covered the rtk tool, which claims to cut Claude Code token usage by 60 to 90 percent by summarizing command output. The JetBrains team tested that claim with 425 real tasks at a cost of about $320. The result was the opposite of the claim: at low reasoning effort, cost per task went up roughly 7.6 percent, and at high reasoning there was essentially no saving at all. He links fuller details in his Telegram channel.
@Liquiddeny [Claude Code]
https://x.com/Liquiddeny/status/2079169699741421954
He argues Obsidian just became the most dangerous folder on your computer, crediting Karpathy's append-and-review method: one giant note, every thought on top, no tags, no folders, no 40-database Notion setup. The old limitation was that the note remembers but can't act, which changed once Claude Code was pointed at an Obsidian vault. His four parts: the vault becomes a database of plain markdown that Claude Code reads like a codebase, turning 2,000 notes into queryable memory; CLAUDE.md at the root acts as the brain stem so every session starts with full context; a nightly command reviews the append log, extracts action items, and drafts the posts hiding in the dump; and scheduled Claude Code runs handle weekly reviews and content calendars. He puts setup at 90 minutes, calls week one overkill, and says by day 30 every note from five years ago is working for you.
@shiqway92 [Claude Code]
https://x.com/shiqway92/status/2079234673645871279
He wrote up someone turning one prompt into 15 UGC videos and landing clients while everyone else still hand-wrote scripts. Claude Code analyzes the brand, tone of voice, products, and audience, then generates 15 complete UGC scripts with different hooks and angles. A simple Node.js bridge automatically pushes only the approved scripts through to MakeUGC, which handles the rest with the same avatar, the same voice, and realistic video generation. The output is a repeatable pipeline rather than one-off AI videos: 1 prompt equals 15 scripts equals 15 UGC videos.
@ziwenxu_ [Claude Code]
https://x.com/ziwenxu_/status/2079239654675485006
He points out that everyone quote-tweeting the $1 AI video angle is missing the better half: the same Postiz skill that schedules a whole week can also read the results back. The loop is batch 20 clips, drop the folder into Claude Code, and say "schedule these to TikTok + IG, 8 hours apart, all week" - it uploads and schedules every one in about a minute. Then on Friday you ask it to pull last week's views per post and name the top 3. The skill has an analytics command, so Claude ranks the clips and tells you which hooks to make 20 more of. Cheap to make, auto-posted, and it grades its own homework.
@OccupyingM [Claude Code]
https://x.com/OccupyingM/status/2079236179904639321
In a reply about how a math animation was made, he names his stack: Kokoro TTS for the voice, and Claude Code driving Manim through a custom harness.
@keitaro_aigc [Claude Code]
https://x.com/keitaro_aigc/status/2079311579645161474
He flags that Claude Code now has a 7-item periodic checkup. Running /checkup reviews unused skills, MCP servers, CLAUDE.md, slow hooks, and permissions in one pass. It asks for confirmation before making changes, and he recommends saving it especially if you use Claude Code heavily.
@grokkedd [Claude Code]
https://x.com/grokkedd/status/2079203154885820571
He shared a Claude Code trick most people miss: paste a big chunk of text and it auto-collapses into [Pasted text #1]. If you need to edit it, paste it again and it expands right back out. Alternatively, hit Ctrl+G to open the entire prompt in your code editor, and your saved changes sync straight back into Claude's input field.
@levelsio [Claude Code]
https://x.com/levelsio/status/2079289397845938480
He gave up on WHOOP's journal after realizing you have to hand-log the days you did NOT go to the sauna, so he had Claude Code build a small web app that pulls in his WHOOP data, his food, and his @wip logs for sauna, tanning and gym, then answers his correlation questions directly. Half the time he never opens the generated app and just asks Claude Code. He also skipped his own site Hotelist and asked Claude Code on the Hotelist VPS to find him a hotel. His read is that the entire app and website layer is getting decimated, and for the first time he has no idea where it goes, except that hardware and IRL work still needs a space, tools and machines.
@fcakyon [Claude Code]
https://x.com/fcakyon/status/2079312247239667752
He is annoyed that in 2026 people are still stuffing rules into md files to stop AI buzzwords. His answer is a hook instead of a prompt instruction, since a hook is deterministic and the model cannot write "delve" or an em-dash past it. He shipped a `humanize` hook that installs in two lines and works in both Claude Code and Codex.
@olliebrand [OpenClaw]
https://x.com/olliebrand/status/2079128350354788722
He photographed a letter about planned electricity work in his area and handed the picture to his agent, which put it in his diary. Later, while planning his week, the agent inferred that the power outage meant his computer would be down too, and steered him toward offline work for that block. He calls it a basic example but says it shows why context matters, and credits OpenClaw (@steipete) plus GBrain (@garrytan) for making it possible.
@mranti [Claude Code]
https://x.com/mranti/status/2079026203940528604
He listed the five AI tools he can no longer work without: Chathub for talking to several SOTA models at once, Claude Code CLI as his main work tool, Chatbox for privacy-sensitive conversations, CC-Switch for swapping his agent onto domestic Chinese LLM engines, and Zcode, where he wrote a skill for deep search and notes that headless Zcode can be embedded into other programs. He still keeps xiaolongxia (OpenClaw) around out of habit, using it much like Doubao for daily briefings and general knowledge questions.
@urjr1 [Claude Code]
https://x.com/urjr1/status/2079100385759998201
Anthropic told Max users they could keep using Fable 5 past July 20. On July 20 his Claude client greeted him with "Buy usage credits to continue" instead, and he posted the screenshot. His line: Claude Code got the announcement before the billing system did.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2079136199524950526
He points out that the Orwell-style writing rules people put in CLAUDE.md miss the AI tic he spots instantly: the "not X, it's Y" shell, as in "the answer isn't X, it's Y" or "X isn't the problem, Y is." His theory is that it gives a plain statement the shape of insight without adding a fact, since training supplies the pattern and post-training rewards the decisive sound. He notes the construction is an ancient rhetorical family that JFK's "ask not" and MLK's color-versus-content line belong to, and that LLMs wore it out in a few years. He then adds six more rules, numbered 7 through 12, covering straw men, capping "not X, it's Y" at once per piece, stopping at two examples, not announcing what you are about to say, not ending two paragraphs in a row on a punchline, varying sentence shape, and breaking any rule sooner than writing like a machine.
@leploutos [Claude Code]
https://x.com/leploutos/status/2079086652925075849
Two days after asking whether to leave Notion for Obsidian, he built the first full version of his second brain: an Obsidian vault on his Mac, a private GitHub repo, a copy on his VPS that his Hermes agent can reach, and Obsidian Git pulling and pushing every 2 minutes with rebase history that stamps the machine name into each commit. The vault has an Inbox, projects, resources, a journal, and templates for notes, meetings, people and ideas. The interesting part is the agent layer: an AGENTS.md that tells every agent to read context before editing a note, preserve frontmatter and wikilinks, never overwrite another tool's changes, check the diff, commit, rebase, then push, with CLAUDE.md importing it because Claude Code does not read AGENTS.md natively. He then wrote one Second Brain skill shared by Hermes, Claude Code and Codex, exposed on the Mac through two symlinks and loaded from the same source on the VPS. What sold him was Markdown plus Git: files he owns, versioned, readable by any tool, reachable by agents without an MCP or a SaaS roadmap in the way.
@ClaudeCode_aca [Claude Code]
https://x.com/ClaudeCode_aca/status/2079038586280722653
He argues that 30 minutes of initial setup literally doubles what Claude Code can do, and boils it down to three moves. Trim CLAUDE.md by asking of each line "if I delete this, will it make a mistake?" Set permissions in three tiers: auto-approve, ask every time, hard block. And switch models deliberately, which alone cuts cost by 30 to 50 percent. He packaged the ten items into a one-page "30-minute roadmap" with the step-by-step details in an accompanying article.
@bkingfilm [Claude Code]
https://x.com/bkingfilm/status/2079045825502851504
He noticed that for the same development request, Codex produces a much better looking interface while Claude Code's UI feels like 90s software. So his workflow now is to start new projects in Codex, and once things are mostly built, move them into Claude Code for optimization and finishing. He is asking whether others see the same split.
@pcshipp [Claude Code]
https://x.com/pcshipp/status/2079240302280556760
He posted the actual math on his AI wrapper app: $16 in revenue against a $90/month stack. The line items are Codex $20, Claude Code $20, Whisper $16, GPT API $8, Deepgram $2, Railway $9, internet $15, and Firebase at $0. His conclusion is that building was the easy part and paying the bills is the hard one.
@InfiniteReign08 [Claude Code]
https://x.com/InfiniteReign08/status/2079321518195048936
This is a blunt complaint aimed at Anthropic to fix fundamentals before shipping new features. The list: stop Claude Code crashing every day, stop it burning through Max limits in an hour, stop censoring conversations about bees, stop forgetting things it was told to remember, and stop treating paying adults like they need a nanny. He also wants better communication with the customers who built the company. His closing point is that shiny things built on a bad foundation crash before they start.
@martypartymusic [Claude Code]
https://x.com/martypartymusic/status/2079295187730264372
He is giving subscribers early access to a trading analysis platform he built with Claude Code, currently three months into back and forward testing. It produces predictions, signals, alerts and full trading strategies. The inputs are Wyckoff, Deribit options, Binance futures, max pain, a 50 percent pullback methodology, and backtesting of fractal repetition.
@exploraX_ [Claude Code]
https://x.com/exploraX_/status/2079197387860435360
He uploaded a video of a shooting-range game to his Claude Code agent and told Fable 5 to recreate it exactly, then ran the same task on DeepSeek V4 Pro. Both models handled targeting accuracy at varying distances, built the full gunsmith with barrel, muzzle, optic, stock, mag and finish plus live weapon stats, and got the HUD right with score, hits, accuracy, fire mode and ammo. Fable 5 came out more polished: an enclosed range with proper booths, labeled ammo crates, a denser target field, hit popups showing distance and headshot or down, and a gunsmith that explains each part's tradeoff. DeepSeek went with an open outdoor range and skybox, but did one harder thing better, actually changing the gun model on screen when you swap attachments. His verdict: Fable 5 on polish, DeepSeek on the harder mechanic.
@mohdcto [Claude Code]
https://x.com/mohdcto/status/2079303205654065306
After months on Claude Code he went back to try Codex and its newer models, and immediately regretted it. He describes the output as terrible and says trivial tasks drag on forever, even though he loaded the exact same Skills he uses in Claude Code. He calls the $20 wasted and has choice words for the people hyping ChatGPT's models.
@leploutos [Claude Code]
https://x.com/leploutos/status/2079279458553569704
He runs GPT-5.6 Sol on Hermes and has it delegate tickets to Claude Code in its CLI, and he swears the output is better when Claude knows a competitor is supervising it. GPT is far harsher in code review of Claude's work than it would likely be with its own, hunting for small errors, rejecting mediocre implementations and sending them back until they are clean. He admits he has no proof and it is just a hunch. His theory is that putting models in competition with a real hierarchy between the executor and the controller makes them less lazy, and that even AI might need an annoying manager and a colleague to impress.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2079038192733311426
He wrote up a case where someone pointed Claude Code at his own machine to free up space, approved the proposed deletions, and permanently lost files he never meant to touch. His take is that the model was not the risk since it did exactly what it was told and the human clicked confirm. The real problem is running large-scale deletions with no dry run and no backup, which ends badly with or without an agent. His advice: never let an agent delete anything, have it move files into a single folder you review yourself, or run a dry run first, and back up before handing any tool that much access. The habit of reflexively approving whatever the agent suggests is what turns into clicking yes on permanent deletions.
@iamsimpleneeraj [Claude Code]
https://x.com/iamsimpleneeraj/status/2079104200202744016
He tried to claim the $100 Fable 5 promo credit on Claude Code and could not find the claim button. The popup told him to go to settings/usage, but the page shows only a disabled toggle with nothing to click, and he posted a screenshot. He is asking @AnthropicAI whether it is a bug or just a slow rollout.
@NewsPicks [Claude Code]
https://x.com/NewsPicks/status/2079155607085404253
They published a video about wiring Claude Code into Obsidian to get a personal AI second brain that fixes the amnesia and hallucination problems. A university-student intern on the editorial team, Mizuki Matano, pushed the idea hard enough that the producer of their "Practical Work Techniques" series, a forty-something salaryman with zero prior knowledge, spent a month actually doing it. The resulting 30-minute video walks through setup to real use, including the spots where beginners get stuck.
@Cathy97S [Claude Code]
https://x.com/Cathy97S/status/2079007540948635746
She is a tax accountant thinking through where agents actually fit in her practice. Her view is that the real prize, bookkeeping and checking, will eventually be handled inside the accounting software itself, which leaves agents doing file organization and renaming, email sorting, drafting replies and digitizing passbooks and receipts, all of which feels like small convenience-tool territory. She plans to cover the basics inside Google Workspace with GAS, Workspace Studio and eventually Spark, using Claude Code to fill the gaps and glue things together. She also admits that overthinking like this is exactly why her agent adoption has stalled, so she finally just signed up for Claude and will start using it.
@punimaru_dev [Claude Code]
https://x.com/punimaru_dev/status/2079095728211923085
His AdMob revenue is running well, clearing 10,000 yen two months in a row, and he posted the numbers. The goal he states is getting AdMob alone to cover his Claude Code subscription.
@charliejhills [Claude Code]
https://x.com/charliejhills/status/2079214655608873120
After two years of writing every post himself across 350k+ followers, he mapped out the 11-tool org chart that replaced him. Hermes runs operations and scores every idea, Reddit, X and Instagram are research, Claude writes every post and script in his voice, and Claude Code is head of design, building every graphic in raw HTML (it deleted the whole folder once). The rest of the bench is GPT Image 2 for illustration, HyperFrames for cutting and captioning reels, Higgsfield for B-roll, Buffer for scheduling, ManyChat for comments and DMs, Apify for analytics, and Notion holding every idea, asset and number. The only seat he keeps is the final yes.
@GlennNieuwenh [Claude Code]
https://x.com/GlennNieuwenh/status/2079182991595544823
He laid out a five-step Claude Code build for an ecommerce profit tracker to replace the paid one you are renting. One prompt spins up a dark dashboard with revenue, ad spend, COGS, shipping, fees, refunds, net profit and margin on dummy data; then you plug in the Shopify Admin API and Meta Marketing API, backfill 90 days, sync hourly, and reconcile against Shopify Analytics before trusting it. Step three adds "Mark," an AI that reads the numbers each morning and writes blunt notes like "your ROAS is dropping on this campaign" or "creative 4 is beating your account average, feed it budget." Step four wires Telegram alerts for blended ROAS below breakeven, an ad burning 3x target CPA with no sales, or a new creative outperforming, and step five deploys it to a cheap host so syncing runs 24/7. He is offering the full doc with all five copy-paste prompts to anyone who comments "TRACKER."
@0xAIGOATexe [Claude Code]
https://x.com/0xAIGOATexe/status/2079340579981513149
He wrote up a 31-year-old operator in Dubai who billed a real estate firm $7,000 for a landing page built in 35 minutes, having already invoiced $3,000 for itsu sushi and $10,000 for Crazy Bears cafe. The workflow was ChatGPT for the hero, Claude Code for the build, Vercel for the deploy, all done standing on the showroom floor after taking one photo of a moquette and pasting a URL. Token cost was $32, and the property manager watched a live URL load before the coffee arrived. His framing is that a boutique agency would have quoted $18,000 and six weeks, and that the playbook is niche-agnostic even if the operator is not.
@aronhouyu [OpenClaw]
https://x.com/aronhouyu/status/2079128116359024668
He remembered a guy from OpenClaw's peak hype who had it monitor trending GitHub projects and post daily. OpenClaw ground away for over ten days before anyone in the group noticed the links were fake addresses and barely any of them opened. He finds the whole thing hilarious.
@DorShinar [Claude Code]
https://x.com/DorShinar/status/2079207538596163742
He has been running GPT 5.6 Sol instead of Opus 4.8 for a few days and honestly cannot feel a difference. He allows that going through Claude Code might mean the harness is not exploiting the model fully. Either way, his sense is diminishing marginal returns.
@Brown_Thunder76 [Claude Code]
https://x.com/Brown_Thunder76/status/2079056709285232836
A friend asked how to do taxes on KTA held on the Keeta network, and since Koinly marks Keeta as "blocked, no fix possible" and CoinLedger cannot import native Keeta wallets, he built the missing tool with Claude Code. You paste a public Keeta address and it generates a CoinLedger-compatible CSV with no account, signup, wallet connection or seed phrase, running entirely in the browser and talking only to Keeta nodes. He tested it on a wallet with over 3,000 transactions and it imported into CoinLedger with zero errors. The harder part was not the exporter but figuring out how Keeta actually behaves: he found eight places where the docs, SDK examples or real network behavior disagreed, including network fees being counted as part of the amount sent, which can quietly produce wrong tax exports. He wrote all of it up in KEETA-TECHNICAL-FINDINGS.md, says that file may be more useful than the tool itself, and open sourced everything with no ads, tracking, analytics or server, disclosing that he is a CoinLedger affiliate but deliberately kept that link out of the tool.
@chesny [Claude Code]
https://x.com/chesny/status/2079330750395023521
He combined Obsidian with Karpathy's one-file method and Claude Code, on the argument that most second brains are cemeteries of 300 saved articles and 80 tagged notes you never reopen. The capture layer is a single Obsidian vault with one file, stream.md, append-only with new thoughts on top and review by scrolling, no folders and no tags, because anything slower than 10 seconds gets abandoned by day four. Once a week Claude Code reads the whole vault from the terminal with one instruction, "find the 5 ideas I keep circling without realizing," and surfaces patterns across six months that working memory cannot hold. Instead of keyword search he asks things like "what did I think about agent pricing in March" and Claude goes through 4,000 notes to quote his own words with file paths. He sketches the payoff curve: useless in week one at 20 notes, first "wait, I already solved this" at month one, Claude linking two notes written 70 days apart at month three, and by month six no longer googling things he already answered himself, all for $0 in software and one afternoon of setup.
@dani_avila7 [Claude Code]
https://x.com/dani_avila7/status/2078998261814235282
He spent time with Claude Code and the DoorDash CLI and published five skills ahead of the full rollout. doordash-order-playbooks stores named orders with context like "post-gym" and diffs the cart against a baseline before handing over the payment link; doordash-spend-guard adds deterministic per-order, day, week and month caps with cooldowns and blocked hours, wrapping checkout so anything outside policy is rejected. doordash-allergy-shield keeps a persistent dietary profile and re-checks critical allergens and hidden sources before checkout, while doordash-group-orders builds team orders from a persistent roster with per-person attribution, cost splitting, payer rotation and live allergy checks. doordash-order-ledger is an append-only audit log of everything the agent does, plus a report that reconciles against real order history.
@Abobsterina [Claude Code]
https://x.com/Abobsterina/status/2079126826916458989
He showcased a setup that connects Claude Code to Viewmax through MCP, drops in a GTA 6 video, and runs a single prompt: turn this video into viral YouTube Shorts and write a voiceover script for each clip. Claude analyzes the footage, picks the strongest moments and turns them into publish-ready short-form videos with hooks and scripts, no editing skills or manual cutting involved. One source video becomes dozens of pieces of content. His pitch is that the real opportunity is not playing the most anticipated game but building an automated content machine around the attention it generates.
@woodyholne [Claude Code]
https://x.com/woodyholne/status/2079268604458029076
He pushed back on someone claiming open models are ready to replace Claude subscriptions. He is a heavy user of Kimi, GLM, MiniMax and DeepSeek, pays for full subscriptions to all of them, and runs them inside harnesses like Hermes and OpenClaw as advisors and helpers. But he says they are fine for side-by-side demos and one-prompt animations, not for staking your actual development work and financial career on. His view is the person he is replying to got ahead of themselves.
@kemu_aii [Claude Code]
https://x.com/kemu_aii/status/2079310458264391893
He points out something most people miss: ChatGPT's Codex can be run from inside the Claude Code screen. That lets you keep Claude's smartest model on the thinking and hand the actual code-typing labor to Codex, all in one window. Because the expensive model's quota drains slower, he says it noticeably stretches how long you can work, and he goes on to explain the setup.
@designertom [OpenClaw]
https://x.com/designertom/status/2079252546993131984
He has run @openclaw since January and @NousResearch Hermes since April, and rates both as good. After all that time his OpenClaw setup is still the better one. He says he was bummed about it because the hype had genuinely gotten him excited for Hermes.
@EngMoElgaraihy [Claude Code]
https://x.com/EngMoElgaraihy/status/2079099961526857888
He highlighted an open source tool called text-to-cad that turns a coding assistant like Claude Code or Codex into a mechanical CAD design engineer driven by plain prompts. Instead of bouncing between CAD software, kinematics analyzers and exporters, one prompt produces print-ready STEP, STL, DXF, GLB and 3MF files and slices straight to G-code, and it can send the job directly to a printer such as a Bambu Lab. It runs fully locally with no backend and no cloud subscription. The demo that got him: a complete 7-DoF robot arm with kinematics animation and a live interface, generated from text commands alone.
@marktechh [Claude Code]
https://x.com/marktechh/status/2079139161714684070
He used $MARK to scan anthropics/anthropic-sdk-typescript, the same SDK behind MARK's own output, and it matched the anthropic-claude, llm-integration, error-handling and typescript-strict skills. He then asked Claude Code for a streaming chat function with retry logic and cost tracking and got streaming-chat.ts at 333 lines in one shot. It includes a MODEL_PRICING table covering every Claude model, real-time chunk handling for streaming, exponential backoff on rate limit errors, and a CostTracker interface tracking input, output and total cost per session. His point is that the ZIP did a day of research before he started typing.
@lazybutai [OpenClaw]
https://x.com/lazybutai/status/2079228977126822183
Replying to @thekitze, he says Hermes just works for him. The last time he tried OpenClaw he ended up debugging it over and over. He also notes Hermes seems pretty good at remembering things.
@hyuki [Claude Code]
https://x.com/hyuki/status/2079340080100204551
With several related projects running at once, his household needed a way to tell their Claude Code instances apart. The convention that caught on is project name plus "Kuroko," their nickname for Claude Code, giving them la-Kuroko, semi-Kuroko, girlnote-Kuroko and so on. Now they say things like "pass that message along to la-Kuroko." He finds it genuinely handy.
@hrabiapolski [Claude Code]
https://x.com/hrabiapolski/status/2079289803112153413
He wrote up the free Superpowers framework, which forces Claude to brainstorm, write a plan, split execution across sub-agents, use TDD, review every task and verify the product before shipping. The contrast case is an article where a 3-page prompt produced broken sub-agents, then a broken chat, then a broken interface, with every repair creating another defect until roughly 100 hotfixes left an app that was just a pile of locally correct patches with no architecture underneath. Superpowers attacks that at the process level with seven controlled phases, isolated worktrees, 2 to 5 minute tasks and two-stage reviews, which leaves fewer openings for the model to improvise itself into a corner. His point is that the model did not get better at coding; it was forced to define the product, prove each change, and stop treating "it runs" as proof that it works.
@10xDinesh [Claude Code]
https://x.com/10xDinesh/status/2079100138216448226
He asked @trq212 whether this is a Claude Code on Windows bug: he cannot use Fable 5 at all. Every time he sends a message or tries to switch to Fable, the client goes off checking for usage credits. He posted a screenshot showing he still has 28 percent of his usage left on the 20x plan.
@drivelinekyle [Claude Code]
https://x.com/drivelinekyle/status/2079233117886562796
He shipped an update to Mixture-of-Agents-X, the open source MoA-X package from @DrivelineBB. The release adds a new Chinese model architecture and better interactive reports after a run. It also adds Codex as an alternative aggregator option alongside Claude Code.
@n1tr0maverick [Claude Code]
https://x.com/n1tr0maverick/status/2079327663961800760
He got Claude Code running natively on Windows 98 and posted video proof. His own joke is that there is something deeply wrong with him for doing it. He also notes that Opus itself refuses to believe the setup is real.
@affe_is_me [Claude Code]
https://x.com/affe_is_me/status/2079059901780029949
He cancelled his Claude Code 20x Max subscription and replaced it with a top tier Kimi K3 plan. He kept his Codex 20x Max subscription running alongside it. No reasoning given beyond the swap itself, which is the point of the post.
@juanca_mnz [OpenClaw]
https://x.com/juanca_mnz/status/2079197628907356474
He had a half-dead side project with almost no traffic, and only the boilerplate SEO that vibecoding had generated when the site was built. He bought a mini PC to run at home as a server for OpenClaw and set up a cron that fires every two weeks, connected to the Search Console MCP, with a progress.md tracking what worked, what failed, and next actions. Each run injects that file and the harness decides on its own what to do from Search Console data, doubling down on rising keywords, adding internal linking, then deploying to production itself via a shell script that does a git pull over SSH and runs the Node build. In two weeks he went from 5-10 new users per month to 10 a day. Next he wants a second cron for product decisions, using PostHog and its MCP to figure out why users do not come back.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2079036143363121474
He watched a 13-minute English tutorial on building a scroll-driven 3D site with Claude Code and Fable 5, with zero design and zero hand-coding, then wrote up the full method in Japanese. The stack is three tools: motionsites ai to find a reference design, Higgsfield to turn the hero image into video (using Seedance 2.0 with a "copy this motion" instruction), and Claude Code plus Fable 5 for the actual React, TypeScript, Vite and Tailwind implementation. The flow is copy a reference, generate the hero video, tell Claude Code to make that video a scroll-driven hero, then tune fonts, spacing and copy by giving Fable 5 instructions in Japanese. He says step three, the implementation, is where the real substance is.
@hiroho150cm [Claude Code]
https://x.com/hiroho150cm/status/2079325478914023535
His fix for bad screenshot reproduction is to stop saying "recreate this layout" and instead force a measurement step first. His prompt tells Claude Code to analyze the screenshot pixels in Python and write out x, y, width and height for every major block, place those numbers directly as px in absolutely positioned HTML, render the result in headless Chrome at the same width as the reference (1440px, for example), measure the render the same way, and produce a diff table. Then it fixes the biggest gaps and repeats the render-and-measure loop, ending with a table of element name, reference value, output value and error in px. His reasoning is that Claude Code is bad at "seeing" but fine at measuring. Running it for real, nav, buttons and photo blocks came in within 1px.
@Kohaku_NFT [Claude Code]
https://x.com/Kohaku_NFT/status/2079182233726849085
He made a full music video in a measured one hour by splitting the work across models. Claude Code did the storyboard, the lyrics and composition, and the editing; GPT Image 2 handled the world design, Seedance 2.0 the video, and Suno v5.5 the music. His stated trick was taking a hiphop MV brief format, structuring it, and turning that into his own prompt.
@krispuckett [Claude Code]
https://x.com/krispuckett/status/2079349602449551782
He found out last week that Claude Code can drive Apple's Icon Composer and create inside it. He posted a clip showing it working. Short discovery post, no workflow detail beyond that.
@Aykutuces [Claude Code]
https://x.com/Aykutuces/status/2079238870999462281
He describes replacing a video editor who charges 20,000 lira a month with a pipeline. The video goes to Whisper for a transcript, Claude Code reads the transcript and picks the six densest moments, FFmpeg cuts those to vertical, burns in subtitles and pulls a thumbnail frame, then the clips get pushed one by one to the TikTok, Reels and Shorts APIs. One recording a week yields six vertical clips with no human hands and no Premiere license. His line: an editor's salary is now a cron job. He promises to drop the repos he uses in the replies.
@Axel_bitblaze69 [Claude Code]
https://x.com/Axel_bitblaze69/status/2079123106468421633
He is pitching a Claude Code skill called ponytail that stops the model from writing more code than the task needs. It makes Claude behave like the laziest senior dev in the room by running a ladder before touching anything: does this need to exist, is it already in the codebase, can the standard library do it, is there already a dependency, can it be one line, and only then the least code that works. He argues the savings compound because less code also means less to read back in later turns. He cites the repo's own benchmarks of 80-94% less code and up to 75% cheaper across every Claude model, and gives the install as /plugin install ponytail@ponytail.
@vorty279 [Claude Code]
https://x.com/vorty279/status/2079257438801248750
He explains a video of Claude Code running on an Android phone, not the web version or the app but the actual CLI in a terminal. Termux gives a rootless Linux environment, proot-distro installs Ubuntu inside it, and Claude Code runs in there, with Ollama in the same frame so a local model runs on the same phone. His argument is not thumb-coding but emergency access: prod goes down while you are out, you open the terminal, SSH into the server, and the agent reads logs and patches it with no laptop, and the local model covers you when there is no network. He is honest about the limits, slow phone CPU, battery drain, narrow screen for agent output. Termux is on F-Droid and setup takes about ten minutes.
@onusoz [Claude Code]
https://x.com/onusoz/status/2079263963670544694
He finally built a thing he had wanted for three months: prompt sequences with real control flow, as nodes in a graph. His example is a Socratic loop where he discusses a problem, asks for the most elegant production-ready solution, asks "is that the holy grail?", and branches on the answer into an "autoimplement" prompt of 6-7 sequential steps. He had previously built acpx as a general ACP-based orchestrator, an n8n-like thing that could drive Codex, Claude Code or pi, but it ran outside the harness like a CI orchestrator. He decided he was being too accommodating to harnesses that are proprietary, take no outside contributions, or do not even support ACP, so he ripped ACP out and rebuilt it natively for pi only, with a native widget instead of a web viewer. The code is at osolmaz/pi-workflows.
@ibocodes [Claude Code]
https://x.com/ibocodes/status/2079167117698756660
He posted his full solo-dev stack cost for 2026, about $124 a month. It is Cloudflare Workers plus D1 for compute and database, R2 for storage with zero egress fees, Resend at $19/month for 50k emails, and Claude Code at $100. His point is no Vercel bill and no surprise invoices.
@Mikadzyki_NFT [Claude Code]
https://x.com/Mikadzyki_NFT/status/2079114171963064348
He writes up a 19-year-old programmer from Albuquerque who spent a couple months of evenings building a home assistant because his smart home was only smart on paper. The stack is ElevenLabs for voice, Claude Code as the brain, and Obsidian as memory. There are no preset scenarios: Claude Code hears a command, writes the code for it on the spot, and runs it in the house, while Obsidian records habits, routine and past conversations so context survives overnight. Then it started acting unprompted, turning off lights when he fell asleep over his laptop, ordering coffee when it ran out, and moving meetings when it saw he would not make it in time.
@crocodila90125 [Claude Code]
https://x.com/crocodila90125/status/2079337742090301816
He flags that AI research tends to collect only the convenient evidence, and points to a skill that attacks that. It is called Hyper Research, installs into Claude Code, and makes the AI go hunt for evidence contradicting the conclusion it just produced. He notes it runs 16 stages total.
@Sokichi_Hoshino [Claude Code]
https://x.com/Sokichi_Hoshino/status/2079030500040323138
He says he has automated nearly all of his work with Claude Code and is now bored with nothing to do. He currently runs six accounts, three on X and three on Threads. So he is thinking about doubling that.
@fluixoo [Claude Code]
https://x.com/fluixoo/status/2079140760658547108
His framing on running Kimi K3 inside Claude Code is that it is not about replacing the best model, it is about keeping the harness and changing the routing. What you keep is the diff engine, approvals, multi-file edits, tool use and project context. He routes K3 to implementation, tests, boilerplate and large-context work, and reserves the strongest model for planning and review. The line he lands on: the expensive model should make decisions, not write another button component.
@aiedge_ [Claude Code]
https://x.com/aiedge_/status/2079330510711791753
He spent a full weekend testing Kimi K3 and gave a condensed verdict. It is excellent at 3D modeling and design work, but nearly unusable on subscription plans, so he recommends hooking in via the API. He thinks Fable still feels genuinely autonomous and built for fully autonomous work, and suggests a stack where Fable or GPT-5.6 Sol orchestrates and Kimi K3 runs as subagents inside Claude Code. Overall he calls it a huge leap for open weights and says OpenAI and Anthropic are on the back foot.
@mardehaym [Claude Code]
https://x.com/mardehaym/status/2079301128609513872
He argues most engineering orgs adopted AI with no plan: one team on Copilot, another on ChatGPT, a third on Claude, nobody coordinating, nobody tracking cost, and six months later an AI bill nobody can explain. His fix is three infrastructure pieces: an LLM gateway routing every developer request and headless agent call through one point for key management, budget caps, audit logs, rate limits and PII redaction; dev telemetry so managers see adoption and where spend concentrates on a dashboard; and LLM traces recording every agent run so you can reconstruct the input diff, reasoning chain and output later, which regulated-industry auditors will demand. He insists on the order: practices first (spec-driven development, TDD, review discipline), then shared tooling, then infrastructure, then agents in the pipeline. In his own org they shipped Claude Code as an org-wide plugin with 20+ specialized commands, one install per engineer, and embedded someone inside the team to run the practice on real work in view.
@suna_gaku [Claude Code]
https://x.com/suna_gaku/status/2079024809594495372
He shares a technique he learned: ask AI for "adversarial verification" instead of "review". A review request looks for improvements; adversarial verification assumes the output is wrong and tries to break the claims, returning a verdict with evidence rather than just notes. He stresses handing the work to a fresh agent with new context rather than letting the author model grade itself, and says in his setup just saying "adversarially verify the current draft" spins up several subagents covering reader perspective, correctness of claims, and whether it holds up as an article, with no custom agent or workflow needed up front. He notes Claude Code already does this internally, with /deep-research and /code-review running separate verification agents to cut false positives, and that you must tell it to check primary sources or the models will share the same errors, with cross-model checks against Codex or Gemini as an option. His caveats: it burns more tokens and time, a skeptic role will always find something so fixing everything causes overengineering, so save it for pre-publish articles, important designs and pre-release code, and the human still decides whether to accept, weaken or reject each finding.
@alwayspriyesh [Claude Code]
https://x.com/alwayspriyesh/status/2079077341234602165
He posted a bug report at Anthropic after Claude Code froze and started repeating the word "court" over and over. The loop burned 5-10% of his usage limit for nothing. He is paying for the product and says this should not be happening.
@freeman1266 [Claude Code]
https://x.com/freeman1266/status/2079038538327400876
He argues people are scared to clear their Claude Code conversation because they think it loses work, and that this is wrong. /clear only wipes the chat log; background builds, tests and long tasks keep running because they are hosted as separate processes from the conversation. Cleared history is not gone either, since /resume restores a past session by ID from local persistent storage, and project code, CLAUDE.md rules, skills and login state are untouched. He flags one trap: /reset and /new are just aliases for /clear, and if you actually want a new independent session carrying the current progress you need /branch.
@0xDepressionn [Claude Code]
https://x.com/0xDepressionn/status/2079230204229083515
He breaks down replacing paid UGC reaction videos with a generated pipeline. The human version runs $10-20 per video, 1-2 days of waiting, and makes you a talent manager chasing people who film vertical when you asked for horizontal; the AI version costs $1-2 and lands in an hour by generating the avatar and reaction with SwapTok, recording 20 seconds of the app, and cutting it in TikTok Studio with trending audio, batched ten at a time. What made it a system rather than a chore is Postiz, which has a CLI built for agents: install the skill, dump the folder, and tell Claude Code to schedule everything in it to his TikTok and IG accounts 8 hours apart over the next week, with setup taking about 30 seconds. He says the $15 saved per video is the boring part; the real unlock is testing 10 hooks a week instead of 2, because the expensive thing was never the money, it was waiting on humans.
@24shi_web [Claude Code]
https://x.com/24shi_web/status/2079228193891033505
He used Claude Code to generate 3D print data for a magnetic storage case that holds an Owltech retractable cable. No CAD software involved at any point. He is genuinely delighted that making an original physical item now works this way.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2079116097081557370
He highlights someone who turned a Japanese digital typewriter, essentially a keyboard with a built-in word processor, into a cyberdeck by installing Linux on it over a weekend. It connects to his Mac over SSH and now runs Claude Code, Codex and Hermes directly. No browser, no social media, no distractions, just a keyboard and his agents wherever he goes. His pitch is that old dead-simple devices are ideal for this and most people already have one gathering dust.
@0x_sakata [Claude Code]
https://x.com/0x_sakata/status/2079157984698208457
He shows a tool that lets Claude or any LLM actually watch a video with nothing uploaded to the cloud. Instead of sampling every Nth frame at a fixed 1fps and missing fast cuts, it extracts every scene change, transcribes audio with Whisper running locally and no API key, packs keyframes into contact sheets so the model sees sequences rather than scattered stills, and generates a local viewer.html with video, keyframes and transcript in one page. It works from a YouTube, Instagram or TikTok URL or a local file, and one command installs it as a skill into Claude Code, Cursor, Codex, Copilot or Gemini CLI. Setup is under two minutes: pip install "claude-real-video[whisper]", optionally npx skills add HUANGCHIHHUNGLeo/claude-real-video, then run crv on a URL and hand the crv-out manifest and frames to any LLM. Scene change detection plus sliding window dedup means fewer frames and better understanding, and it also works as a plain keyframe extractor with no ML models.
@thekuchh [Claude Code]
https://x.com/thekuchh/status/2079321145212280918
He summarizes someone's full AI app-building workflow. You ramble your idea at Claude, which turns it into a product doc, then a technical doc, then a step-by-step plan; you split that plan into tickets and give each ticket to its own Claude Code instance. Each instance runs ticket, build, pull request, self-review, fix, push, then you test that one feature and repeat. His claim is that this loop is what separates an app that ships from a demo that falls apart, and that it is step 4 of a 9-step build quoted below.
@LotusDecoder [Claude Code]
https://x.com/LotusDecoder/status/2079019385008935140
He replies that he equips Codex and Claude Code with the Amap MCP and a high-speed-rail CLI API, then adds CDP, computer use and mobile use so the agent can search Xiaohongshu, Meituan, Dianping and Douyin. At this point basically all his trip planning is handed to the AI. His example query is finding a cafe near him with a forest view, and he notes the AI actually inspects the merchant's environment photos to check whether a forest is visible.
@MakadiaHarsh [Claude Code]
https://x.com/MakadiaHarsh/status/2079121035534672133
He lays out how he ships MVPs in 3-6 weeks instead of the usual four months of idea, wireframe, hire devs, revise, launch. Step one is dumping the idea in plain English into Claude and getting back a full PRD with flows, features, edge cases and technical architecture. Step two is v0 for real component layouts so he can iterate visually before writing code. Step three is Cursor or Claude Code, fed the PRD plus the UI references, building frontend and backend together. His stated secret is that bad specs kill 90% of projects, so nail the spec first.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2079041927589281796
He added six new generative AI kits usable on Google Colab to comfy-agent, his CLI tool that lets Claude Code and Codex drive ComfyUI. The additions are BiRefNet for background removal, Stable Audio 3 for audio generation, LTX-2.3 and Wan 2.2 S2V for video, SeedVR2 for restoration and upscaling, and HiDream-O1 for image generation. It is aimed at people who want an agent to automate image, video and audio generation through ComfyUI.
@julie_bush [Claude Code]
https://x.com/julie_bush/status/2079058704343896255
She replies that her use case was filing a lawsuit against Anthropic for banning her from Claude Code. That is the whole post, delivered deadpan as a one-liner.
@mindmoon_108 [Claude Code]
https://x.com/mindmoon_108/status/2079179215333994743
He reports from inside a semiconductor company where management is pushing hard to build AI agents and agentic AI internally. Outside work, Claude and Codex just build things with one click; inside, it does not work that way. They can attach the internal model to Claude Code and Hermes, but document generation still fails and the available tooling is immature. The biggest bottleneck is internal data: the tables they need are scattered everywhere and have to be combined, and the AI cannot find them, so engineers manually hunt down where the data lives, which alone takes a long time. His flip side read is that there is still plenty of ground AI has not reached, so engineers are still very much needed.
@z_zabaglione [Claude Code]
https://x.com/z_zabaglione/status/2078996182312116467
He has done many hackathons but this was his first centered on generative AI, and the speed difference was obvious. Older hackathons ran 9am to 10pm across two days, with all-nighters before COVID; this time actual dev time was under three hours and everyone still had something working. He remembers past events where a teammate would contribute a little to ideation, then vanish during the build while asking "is it done yet?", and show up proud only at the final presentation. His take now is that with Codex and Claude Code, the person muttering to themselves while handing implementation to the AI might actually be the better teammate.
@connect24h [Claude Code]
https://x.com/connect24h/status/2079041596016980380
He points at a Recruit deck covering a year of giving Claude Code and Codex to every engineer in a department, and says orgs that only speed up implementation are already in trouble. Their SEO work went from proposal to implementation in half a day and shipped 63 initiatives in three months, but the cheaper implementation gets, the more the waiting on decisions, reviews and operations becomes the bottleneck, and eight people means 56 communication paths, so of course it jams. He sees the same thing himself: the more agents he adds, the more his time goes into acceptance criteria and handoff design rather than code generation. His asks for the next sprint are to measure decision wait time and KTLO alongside implementation time, flatten the hierarchy, and widen how much of requirements-to-operations one person can own. He notes the post has 410 Hatena bookmarks and calls the era of handing out AI without changing the organization over.
@AIkenkyuFM [Claude Code]
https://x.com/AIkenkyuFM/status/2079340579352387848
He argues accounting firms should customize Claude Code into an office-specific tool instead of retyping long prompts every time, using Agent Skills to make shortcut commands. He names two skills every accounting office should register: /format-check, which instantly reshapes the messy CSVs clients send into the firm's own chart-of-accounts mapping, and /weekly-summary, which auto-generates a progress report from the week's Slack threads and file update history. His point is that one line of command starts a routine job immediately, which is ideal for standardizing operations.
@peer_rich [Claude Code]
https://x.com/peer_rich/status/2079190044955328609
At the start of the year he predicted that by the end of 2026 he would not be writing code by hand, and he is already at 80%. He says the remaining 20% feels like habit rather than necessity. His use of AI is now delegation rather than autocomplete: he describes what he wants, Claude Code or Devin implements, and he reviews. His conclusion is that the most valuable engineers in two years will be the ones who think clearly about systems, catch bad output fast, and know when the AI is confidently wrong.
@bggg_ai [Claude Code]
https://x.com/bggg_ai/status/2079107884110250066
He got a Canvas Agent working in his product ngsFlow and says the key was making the site expose its business capabilities as a product protocol an agent can understand and call. He shares the exact prompt he uses to have Codex audit a codebase and design the retrofit: the agent stays a side-channel orchestration layer without replacing existing flows, database, APIs or job system; capabilities go into a single Capability Catalog so tool definitions are not duplicated across pages, CLI and MCP; the agent reads current page and project state before emitting a declarative execution plan; every write supports dry-run, change preview, one confirmation, atomic commit, version conflict detection and undo; payment, deletion, overwrite, publish and local command actions get separate approval and budget caps; long tasks use persistent Run/Job state instead of one long HTTP request; success is verified by deterministic conditions rather than the model's own reply; a unified trace records run, turn, tool, approvals, timing, results, failure causes and redacted sensitive data; and sessions, permissions and context are isolated per user, project and page. It also demands a read-only audit first, then current architecture, target architecture, phased plan, modules to add or change, risks and compatibility, and verifiable acceptance criteria, with no large-scale rewrite. Once the protocol exists, the front end is interchangeable: a web sidebar agent, or local Codex and Claude Code over CLI and MCP.
@theSethian [Claude Code]
https://x.com/theSethian/status/2079115791673282767
He walks through AI Samson's build of an AI influencer who keeps the same face, voice and body across every clip, with setup claimed under 20 minutes. It starts by giving Claude a personality, a playful adventurous travel creator with blonde hair, freckles and a septum piercing, and those details become the visual rules; at 02:31 one portrait becomes a character sheet with front, profile, three-quarter and full-body views that survives reuse across six images. Seedance 2.0 animates the stills while a saved voice keeps audio consistent, and at 07:54 Kling 3.0 Motion Control transfers his own acting to the character from a driving video while keeping his face and gender off camera. The long-run problem is drift: the face stays stable but eyes lose focus, lines land with the wrong expression, pauses fall oddly, head movement shifts between campaigns. The fix is storing decisions outside the prompt, with avatar_id anchoring identity, voice_id preserving the voice, and character/bible.md recording gaze, expression, pacing, pauses and head movement, with Claude checking every campaign against those rules and a human approving the JSON before credits are spent, while MakeUGC handles actor, voice and render, Claude Code prepares local files, and a Node.js script connects the two after approval.
@satgeze [Claude Code]
https://x.com/satgeze/status/2079000132687888820
He ran an initial test of Qwen3.8-Max-Preview through his own app Spock, which proxies and routes the model into the Claude Code harness inside VSCodium. He set Mode to Auto with efforts on "Ultracode - xhigh + workflows" and asked for an incredibly detailed interactive 3D solar system in a single HTML file using three.js or better, with hyper-realistic planets, satellites and asteroid belts, explicitly allowing it to use workflows and agents. It produced the whole simulation in one shot and he calls the result superb.
@_vmlops [Claude Code]
https://x.com/_vmlops/status/2079190286484537631
He shows an AI desk companion built for Claude Code. It runs on battery, stays connected to Claude Code, and shows usage in real time on a tiny desk display, so a glance tells you remaining quota and when it resets instead of opening a browser. The next planned feature is live AI activity status. His larger vision is a unified dashboard tracking every AI agent running in the background, not just Claude Code.
🗣 User Voice
User Voice
Reliability is now the loudest complaint, and it's coming from paying customers rather than free-tier users. @InfiniteReign08 pushed back on new feature announcements with a list of foundations instead: it crashes every day, it burns through Max limits in an hour, it censors ordinary conversations, and it forgets the things that matter — her argument being that anything shiny built on that foundation is destined to crash. @alwayspriyesh watched the client freeze and repeat a single word in a loop until it had eaten 5 to 10 percent of his usage limit for nothing. @farzyness says it simply takes forever to do anything lately, which he finds strange given models are supposed to be getting more efficient.
Billing and entitlement bugs were unusually widespread on this particular day. @urjr1 was told Max subscribers could keep using Fable 5 after July 20, then on July 20 was asked to buy usage credits, and observed that the announcement reached Claude Code before it reached the billing system. @185shingeki got the same credit prompt on a Max 20x plan at roughly 15 percent usage, and found that /logout followed by /login cleared it. @10xDinesh cannot use Fable 5 at all on Windows — every attempt drops into an endless "checking for usage credits" state with 28 percent of his 20x plan still remaining. @iamsimpleneeraj went looking for the claim button on the $100 Fable 5 promo credit and found only a disabled toggle.
Token consumption is being treated as a product defect, not a fact of life. @Axel_bitblaze69 argues Claude Code writes far more code than the task requires and quietly burns tokens doing it, and that less code compounds because there's also less to read back next session. @andruyeung hit his limits for the first time on the $100 plan and is openly asking for optimization guidance. @crocodila90125 puts the comparison starkly: at the same roughly $200 a month, ChatGPT Pro feels effectively unlimited while Claude is capped. @affe_is_me went further than complaining and cancelled his 20x Max subscription outright, replacing it with a top-tier Kimi K3 plan while keeping Codex 20x Max — and separately reports that Fable 5 auto-degrades so often he can't use it half the time, which was the only reason he kept the subscription.
There's a specific, fixable context gap that a researcher flagged. @johnmyleswhite points out that neither Claude Code nor Codex makes sure the model actually sees timestamps for prompts and actions, which means the model can't tell that a user's message was a reaction to a world state that existed before some of its own recent edits. He calls it genuinely astonishing, and it's the rare complaint that reads as a concrete bug report rather than a preference.
Output quality complaints have gotten specific enough to be actionable. @bkingfilm says that for the same request Codex produces a far prettier interface while Claude Code's UI feels like 1990s software, so he now starts projects in Codex and moves them over for optimization. @x_sanjin's issue is the windup: the answer arrives only after a long preamble, which is why a skill that forces the answer first has pulled close to a million views. @vburojevic offers the sharpest cross-product comparison of the day — Hermes has literally never broken in the four months since he installed it, while he was fixing OpenClaw daily or weekly, especially after each update. @AmericanDebunk found neither Claude nor SuperGrok could successfully walk him through setting up OpenClaw locally, and @pcshipp posted the economics that make all of this sting: $16 in app revenue against a $90 monthly tool bill.
Reliability is now the loudest complaint, and it's coming from paying customers rather than free-tier users. @InfiniteReign08 pushed back on new feature announcements with a list of foundations instead: it crashes every day, it burns through Max limits in an hour, it censors ordinary conversations, and it forgets the things that matter — her argument being that anything shiny built on that foundation is destined to crash. @alwayspriyesh watched the client freeze and repeat a single word in a loop until it had eaten 5 to 10 percent of his usage limit for nothing. @farzyness says it simply takes forever to do anything lately, which he finds strange given models are supposed to be getting more efficient.
Billing and entitlement bugs were unusually widespread on this particular day. @urjr1 was told Max subscribers could keep using Fable 5 after July 20, then on July 20 was asked to buy usage credits, and observed that the announcement reached Claude Code before it reached the billing system. @185shingeki got the same credit prompt on a Max 20x plan at roughly 15 percent usage, and found that /logout followed by /login cleared it. @10xDinesh cannot use Fable 5 at all on Windows — every attempt drops into an endless "checking for usage credits" state with 28 percent of his 20x plan still remaining. @iamsimpleneeraj went looking for the claim button on the $100 Fable 5 promo credit and found only a disabled toggle.
Token consumption is being treated as a product defect, not a fact of life. @Axel_bitblaze69 argues Claude Code writes far more code than the task requires and quietly burns tokens doing it, and that less code compounds because there's also less to read back next session. @andruyeung hit his limits for the first time on the $100 plan and is openly asking for optimization guidance. @crocodila90125 puts the comparison starkly: at the same roughly $200 a month, ChatGPT Pro feels effectively unlimited while Claude is capped. @affe_is_me went further than complaining and cancelled his 20x Max subscription outright, replacing it with a top-tier Kimi K3 plan while keeping Codex 20x Max — and separately reports that Fable 5 auto-degrades so often he can't use it half the time, which was the only reason he kept the subscription.
There's a specific, fixable context gap that a researcher flagged. @johnmyleswhite points out that neither Claude Code nor Codex makes sure the model actually sees timestamps for prompts and actions, which means the model can't tell that a user's message was a reaction to a world state that existed before some of its own recent edits. He calls it genuinely astonishing, and it's the rare complaint that reads as a concrete bug report rather than a preference.
Output quality complaints have gotten specific enough to be actionable. @bkingfilm says that for the same request Codex produces a far prettier interface while Claude Code's UI feels like 1990s software, so he now starts projects in Codex and moves them over for optimization. @x_sanjin's issue is the windup: the answer arrives only after a long preamble, which is why a skill that forces the answer first has pulled close to a million views. @vburojevic offers the sharpest cross-product comparison of the day — Hermes has literally never broken in the four months since he installed it, while he was fixing OpenClaw daily or weekly, especially after each update. @AmericanDebunk found neither Claude nor SuperGrok could successfully walk him through setting up OpenClaw locally, and @pcshipp posted the economics that make all of this sting: $16 in app revenue against a $90 monthly tool bill.
📡 Eco Products Radar
Eco Products Radar
Codex — by far the most-mentioned companion tool, appearing as a rival, a second opinion, a cheaper executor, and in one case an adversarial reviewer running inside a Claude Code workflow.
OpenClaw — the default answer for always-on personal agents, and increasingly the thing people compare unfavorably to Hermes on reliability.
Hermes — the most frequently cited alternative agent runtime this run, repeatedly described as the one that just works.
Cursor — still the standard IDE reference point in mixed-tool stacks.
MCP — the connective tissue in almost every non-coding workflow here, from Amap and Search Console to Coolify, Postiz, Viewmax, Apify and ComfyUI.
Kimi / Kimi K3 — the routing destination of choice for implementation, tests, boilerplate and large-context work, with the expensive model reserved for planning and review.
Obsidian — the storage layer under the entire second-brain wave, chosen specifically because plain Markdown on disk survives any vendor's roadmap.
CLAUDE.md — treated less as configuration and more as the operating manual, with people publishing numbered writing rules and importing AGENTS.md into it.
Grok / Grok Build — now open source and showing up as a fourth agent in multi-CLI setups.
Fable 5 — the model people are paying specifically to reach, and the one generating the most billing and degradation complaints.
DeepSeek, GLM, Qwen — the open-weight bench, used heavily for demos and side-by-side tests but explicitly not trusted with career-critical work by the people running them.
Whisper — the quiet workhorse behind most of today's video, subtitle and clip pipelines.
n8n — the scheduler of choice for people running agents on a VPS rather than a laptop.
Codex — by far the most-mentioned companion tool, appearing as a rival, a second opinion, a cheaper executor, and in one case an adversarial reviewer running inside a Claude Code workflow.
OpenClaw — the default answer for always-on personal agents, and increasingly the thing people compare unfavorably to Hermes on reliability.
Hermes — the most frequently cited alternative agent runtime this run, repeatedly described as the one that just works.
Cursor — still the standard IDE reference point in mixed-tool stacks.
MCP — the connective tissue in almost every non-coding workflow here, from Amap and Search Console to Coolify, Postiz, Viewmax, Apify and ComfyUI.
Kimi / Kimi K3 — the routing destination of choice for implementation, tests, boilerplate and large-context work, with the expensive model reserved for planning and review.
Obsidian — the storage layer under the entire second-brain wave, chosen specifically because plain Markdown on disk survives any vendor's roadmap.
CLAUDE.md — treated less as configuration and more as the operating manual, with people publishing numbered writing rules and importing AGENTS.md into it.
Grok / Grok Build — now open source and showing up as a fourth agent in multi-CLI setups.
Fable 5 — the model people are paying specifically to reach, and the one generating the most billing and degradation complaints.
DeepSeek, GLM, Qwen — the open-weight bench, used heavily for demos and side-by-side tests but explicitly not trusted with career-critical work by the people running them.
Whisper — the quiet workhorse behind most of today's video, subtitle and clip pipelines.
n8n — the scheduler of choice for people running agents on a VPS rather than a laptop.
Comments