Super User Daily: July 13, 2026
The center of gravity moved from "which model is better" to "how do I run a fleet." The day's strongest cases are org charts, not prompts: Fable 5 as architect with Sonnet or GPT-5.6 executors, Codex subscriptions wired in as cheap subagent labor, keyboards turned into agent status boards, cat-sound hooks replacing terminal-staring. Non-coding keeps expanding — one user turned an $80 hospital MRI CD into an explorable 3D brain model, another runs a 14-agent OpenClaw analyst floor for investment research, a couple names their household bots while their vending machine texts the owner. The tax on all of it is the same: token burn, session amnesia, and the 5-hour limit quietly scheduling everyone's day.
@kajikent [Claude Code]
https://x.com/kajikent/status/2075962953086427302
A Japanese user shares his nightly routine of dumping a massive batch of tasks into Claude Code before going to sleep. He posts the exact prompt he uses: telling Claude he is going to bed, so any items requiring his confirmation should be deferred to the end while Claude proceeds with everything it can handle on its own. Because the workload is bulky, the prompt also instructs Claude to create a work-management file and log progress as it goes. The user then appends a long list of everything he wants done.
@enhanced_jp [Claude Code]
https://x.com/enhanced_jp/status/2075780986290479294
A skeptical Japanese take arguing that Claude Code implementations done single-handedly are full of holes. Even with Fable designing, Sonnet implementing, and Fable auditing, adversarial review within the same model family drifts toward approval and fails to find flaws. The author says this undermines trust to a serious degree, calls trust in Codex overwhelming by comparison, and muses about switching over.
@theo [Claude Code]
https://x.com/theo/status/2075742083370127504
Theo criticizes GPT-5.6 Sol's "ultra" setting because every subagent it spawns inherits the ultra level too, causing massive token burn for no good reason. He argues users should at minimum be able to hard-set subagent effort to "medium". His verdict: Claude Code is far ahead on this aspect of subagent effort control.
@PrajwalTomar_ [Claude Code]
https://x.com/PrajwalTomar_/status/2075875252408926367
Claude Code can now effectively watch videos — YouTube, Zoom calls, Loom recordings — via a new skill: paste a link and it processes the whole thing while you keep working. Unlike old tools that grabbed 100 random frames, this skill uses FFmpeg scene detection to capture a frame every time the screen actually changes, pulls captions free from YouTube, and uses Whisper for Zoom and Loom. When finished it offers to save the analysis into your knowledge base. The author plans to feed it every consulting call recording from the week and have action items waiting in his second brain before he is even off the call.
@mamahmuda1_ [OpenClaw]
https://x.com/mamahmuda1_/status/2075760153950093504
An Indonesian tweet asking IT folks how long they usually stay obsessed when trying something new. The author's boyfriend is deep into playing with Hermes and OpenClaw, building bots that each have their own name: one serves as a shared assistant for the couple and two others as personal assistants for each of them. Every day he is thinking about adding RAM and buying more gear. The author finds it fun since she got access too, but jokes that now both of them are busy with their own bots.
@Voxyz_ai [Claude Code]
https://x.com/Voxyz_ai/status/2075932258645446709
The author ran GPT-5.6 Sol in the new Codex for 2 hours and it burned through 5 hours of a 20x Max quota, because sol ultra behaves like Claude Code's ultracode: subagents inherit the main agent's reasoning level and burn quota at the same setting. He first locked subagents to high, found it still overkill, and dropped them to medium, where the model handled jobs previously covered by several skills, so he deleted another batch. His ideal split is one tier for coding and bug fixes, one for search, docs, and tests, and the top tier only for tasks he will not trust lower. He shares a cleanup prompt that audits all local Codex and Claude Code skills plus session-context plugins, recommends keep/merge/archive/delete per item with one-line reasoning, preserves project-specific rules and open loops, and outputs a dry run only.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2076075310332981571
Someone programmed their keyboard LEDs into a live status board for Claude Code sessions running across every terminal tab. Each key in the number row maps to one tab: green means Claude is waiting for input, red means actively working, yellow means the tab went idle, blue marks a regular non-Claude tab. Hitting cmd plus the number jumps straight to whichever session needs attention, so instead of clicking through tabs you just glance down at the keys. The post frames it as possibly the cleanest solution yet in the broader trend of knowing when your agent needs you without staring at a terminal.
@aakashgupta [Claude Code]
https://x.com/aakashgupta/status/2075880216120197540
An argument that the real bug in AI products is never in the prompt — the agents surviving contact with real users are the instrumented ones. A builder running 16 agents across a job search tool and a newsletter recommendation engine kept hearing hallucination complaints he could not reproduce until he traced the agent step by step: a resume feedback agent told a candidate the company used React when the posting said Python. He ran the fix through Claude Code, which read the traces and proposed four unwritten eval criteria such as staying grounded in the input. The evals found the same mistake occurring 12% of the time; one fix dropped it under 2%, in twenty minutes start to finish, and it now runs on its own. The closing lesson: know what to evaluate before you ship, and trace everything.
@robj3d3 [Claude Code]
https://x.com/robj3d3/status/2075871056922054693
A self-deprecating confession: the author never felt more alive than running 20 parallel Claude Code conversations powering 100+ Fable 5 subagents. It cost $4,000. It shipped a button. The button does not work. He would do it again.
@jasonugc [Claude Code]
https://x.com/jasonugc/status/2075852488473718904
A breakdown of a 100% AI-made ad in balloon art style, produced for under $2 in compute and 10 minutes start to finish. The workflow: script it with the right Claude skill (not generic slop), generate every scene's first frame, feed those into Omni or Kling, and stitch in CapCut. Almost all of it can be automated with Claude Code. The strategic point: while everyone in supplements runs the same girl-talking-to-camera ad, a probiotic sold by a balloon is a pattern interrupt your brain has never seen.
@kenichiota0711 [Claude Code]
https://x.com/kenichiota0711/status/2075945341380411443
A Japanese post describing a system for non-designers inside a company to mass-produce B2B landing pages, built around Claude Code. Instead of generating from scratch, they clone a canonical LP and swap only the variable parts, so designs follow the same components and structure as production. Three pillars: all color/spacing/font values consolidated into a single tokens.json that a linter judges against mechanically; requirements gathering handled by a Claude Code skill (/lp-hearing) that fills fixed variable slots so anyone gets the same coverage; and quality checks moved from humans to machines via a hook that auto-runs the linter after implementation, blocking violations so the AI fixes itself in a loop. The author notes LPs have an expressive layer that resists componentization, so consistency of the template must coexist with freedom at the core.
@poistudioltd [Claude Code]
https://x.com/poistudioltd/status/2075792202933162279
A full copy-paste prompt to recreate a hero section 1:1 in Figma using Claude Code plus the Figma MCP. The prompt spells out exact specs: a 1280x1049px frame with #FDFCF4 background, a navbar with an 1128px auto-layout container, logo, nav links at specific opacities, and a "Sign up for free" button with defined padding, fills, and radius. It continues with a 1128x643px hero image at 16px corner radius, and a centered 810px content block: a Cardo 64px headline "See Further. Think Deeper.", a Geist 18px subheading at 70% opacity, and a CTA button with two layered drop shadows. It also requires installing the Switzer font first, with Inter as a non-1:1 fallback.
@coreyganim [Claude Code]
https://x.com/coreyganim/status/2076079085713862982
A recommendation to build a personal knowledge base in a total setup time of 45 minutes. The steps: 5 minutes creating three folders (raw, wiki, outputs) plus a CLAUDE.md schema file; 10 minutes dumping every transcript, note, screenshot, and SOP into raw/; 30 minutes letting AI read everything, write the wiki, and create the index; then an ongoing compounding loop of new sources in, sharper answers out, plus a monthly health check for contradictions, stale topics, and gaps. For the dump step he has Claude Code or Codex pull data from connectors, and for the loop he runs a weekly cron with his Hermes agent that analyzes the week's work and suggests knowledge-base updates.
@0xtouyan [Claude Code]
https://x.com/0xtouyan/status/2075759714336682452
A Japanese guide to making the officially recommended "Fable as orchestrator" strategy the default in Claude Code / Cowork, aimed at people who want their Fable subscription quota to last through the weekend. Three steps: set model to Fable in ~/.claude/settings.json so only the main loop runs Fable; set the default subagent model to Sonnet 5 via model: sonnet in the frontmatter of .claude/agents/*.md; and write an orchestration policy in CLAUDE.md stating the main loop focuses on interpreting instructions and delegating while actual work goes to named subagents. Most tokens then get consumed by the Sonnet workers so the Fable quota lasts longer, and tasks needing deep reasoning can be upgraded per-call via the Agent tool's model parameter. Full config set is public via a link.
@cjzafir [Claude Code]
https://x.com/cjzafir/status/2075772170794619114
The author did something he calls crazy: running Fable 5 inside the Codex app via a Claude Code subscription, with GPT-5.6 Sol extra high as the main orchestrator and GPT-5.6 Sol high as executor/subagents. He built a custom Codex plugin that makes this cross-vendor setup work and promises to open source it the next day.
@smalkalbani [Claude Code]
https://x.com/smalkalbani/status/2075840641834324436
An Arabic post sharing an economical way to use Fable 5, based on Anthropic engineer Lance Martin's experiments pairing Fable 5 with cheaper models like Sonnet 5 to excellent results. The author condensed the method into a single prompt that turns Claude Code from a code-writing assistant into a smart architect plus a cheap execution team, recommending which model to use per task. The prompt casts Fable 5 as Architect/Lead Optimizer, applies a Hybrid Intelligence strategy (Fable 5 only as orchestrator/advisor/verifier, Sonnet 5 for daily execution), and produces seven outputs: an architecture plan, sub-task breakdown, 3 checkpoints, a ready Architect prompt, per-sub-task Executor prompts for Sonnet 5, a final Verifier prompt, and cost-saving tips. The author reports a positive personal experience with it.
@milindlabs [Claude Code]
https://x.com/milindlabs/status/2076069411208003891
The author describes his new way of developing iOS apps: just pointing at things. No screenshotting, no annotating, no dragging and explaining — he circles his mouse around the place he wants changed, tells Claude Code to "read this", and Claude Code takes care of the rest. He frames "point and tell" as the new way of talking to coding agents.
@firesidealpha [Claude Code]
https://x.com/firesidealpha/status/2075817292513022388
Quotes from Dylan Patel explaining that his firm hands OpenAI only the jobs that can run overnight and keeps everything else on Claude, because OpenAI burns 3-4x the tokens. He says OpenAI models can sometimes do edge-case tasks in leading science, math, and code that Anthropic models cannot, but they take 3x as long and 4x as many tokens, so they cost more and slow the human-AI feedback loop, ending up worse on customer perception. With a human in the loop, Anthropic is way faster and better because it is more token efficient — which he says is why Anthropic has been beating OpenAI. His shop remains majority Anthropic: overnight tasks go to OpenAI Codex, most tasks stay with Claude Code.
@gabimoncha [Claude Code]
https://x.com/gabimoncha/status/2076073312682226114
The author observed that Matt Pocock's skills get triggered more often in Codex than in Claude Code and points to a gap in the skills standard: because Claude Code got a head start, popular skills like Matt's only carry `disable-model-invocation: true`, which Codex does not support. Codex needs separate metadata in agents/openai.yaml, specifically `policy.allow_implicit_invocation: false`. He forked the skills, made them all Codex-compatible, and is waiting for Matt's permission to open a PR.
@r_konno1005 [Claude Code]
https://x.com/r_konno1005/status/2076088975341891843
A Japanese reply to kajikent's overnight-tasks post: this user also dispatches work to Claude Code every night via launchd. He found that adding just two lines to the prompt — "proceed to the end without pausing for confirmation" and "before finishing, re-read to check for anything left undone" — noticeably raised the morning completion rate. He says he will borrow kajikent's prompt as reference.
@srishticodes [Claude Code]
https://x.com/srishticodes/status/2075849201523523729
A shout-out to a GitHub repo the author calls gold for anyone running Claude Code, Codex, or any coding agent: an open source context layer that stops your agent from reading 40 files to answer one question, with no API key, billing, or rate limit tiers. Benchmarked against a Claude Code baseline it claims 27x token efficiency, 36% lower cost, 89% fewer file reads, and 49% fewer tool calls, all reproducible with the harness in the repo. It has 3.5k+ stars, 50k+ pip installs, 30+ contributors, and works with Claude Code, Cursor, and VS Code out of the box. Usage: pip install repowise and point it at your codebase; the author says their usage limits now last almost twice as long.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2075799613173629347
Japanese impressions after heavily using GPT-5.6 in Codex: browser operation is accurate and fast, design quality of web services and documents has improved, and image generation within the flat-rate plan is convenient — but even the $200/month Pro plan hits usage limits surprisingly fast. His division of labor: hand browser-operation work and design tasks to Codex, which also keeps Claude Code usage down. He notes that services like note, which lack an API and tend to reject Claude Code-style browser automation, can often be operated smoothly from Codex; XML file upload is an alternative but browser control is easiest.
@izutorishima [Claude Code]
https://x.com/izutorishima/status/2075987802811244663
A Japanese take arguing that for trivial simple web apps, Sonnet 5 is good enough as orchestrator without burning Fable 5, and in that case it may be best to use Claude only as a sounding board for planning and hand everything from implementation and review onward to GPT-5.6. The author also likes agmsg-based messaging between Claude Code and Codex: the implementer can ask the original planner — who holds the user's raw instructions and context — directly, which prevents deviation from the design intent.
@petergyang [Claude Code]
https://x.com/petergyang/status/2075797711513903244
Peter Yang has started using HTML artifacts for specs and plans because they are much easier to look at than all-text markdown files, crediting trq212 for the tip. The friction: giving AI feedback on an HTML plan or editing it directly is too hard — he has to describe changes in a separate chat and hope the AI understands what he is referring to. He wants to select text or elements on the HTML plan to send feedback or fix it himself, and argues comment, editing, and collaboration features should be natively built into the Claude Code and Codex web viewers as a big improvement for human-AI collaboration.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2075747700553957846
A Japanese video walkthrough of OpenAI's new GPT-5.6 model and the "Work" feature — with the video editing itself done entirely by Claude Code, which the author calls incredible. Building real apps and 3D games, he found GPT-5.6 much improved in speed and design, while Claude still holds the edge in stability on complex development and deep thinking. The video covers the Sol/Terra/Luna model differences, running 4 parallel AIs with Ultra, the 6 thinking levels including MAX, fast processing with fewer tokens, Work's automatic deliverable creation, scheduled runs and change monitoring, short-prompt accuracy techniques, and a hands-on comparison with Claude.
@SuguruKun_ai [Claude Code]
https://x.com/SuguruKun_ai/status/2075891302538133658
A Japanese review of an agent skill that audits AI-built UIs for "AI smell" and strips it automatically. It catalogs 23 patterns of AI-ish design — indigo-to-purple gradients, glowing cards, emoji everywhere, a mascot in the corner — and when run in Claude Code it flags exactly where and why src/ looks bad, then fixes each item with a minimal diff. The author had previously hand-written a DESIGN.md for removing AI smell and sees this as an automation of exactly those checks, arguing AI smell comes from everyone converging on the same defaults so human taste must be verbalized and eliminated point by point. The skill hit 180 stars within a day of release.
@Vincent_AINotes [Claude Code]
https://x.com/Vincent_AINotes/status/2075871675670040751
A Chinese tip on a new vibe coding posture: open Codex, choose "Terminal" in the right sidebar, and type claude to run Claude Code inside it. Division of labor: Claude Code handles requirements clarification, task decomposition, and planning; Codex handles code changes, running tests, and delivering results. The slogan: Claude does the thinking, Codex does the work — adults don't choose, they take both.
@miroburn [Claude Code]
https://x.com/miroburn/status/2075971792741531773
A Polish assessment of Codex and GPT-5.6 Sol from an agentic/yolo workflow where one session builds, deploys, does marketing, sells, switches contexts, and spawns subagents. Claude Code with Fable 5 (Medium) dominates this style, and even GPT-5.5 High did fine — but the author cannot get GPT-5.5-High-level work out of anything below GPT-5.6 Sol xHigh. In his environment 5.6 at Medium/High completely fails, ignoring the harness and wrecking things even on simple tasks, despite GPT's reputation for surgical precision. He suspects the problem is Codex rather than the model itself, concedes it is just his environment and he may be doing something wrong, but insists an upgrade should deliver an upgrade rather than forcing him to adapt to the model.
@shannholmberg [Claude Code]
https://x.com/shannholmberg/status/2076033805241249895
The three planning frameworks the author runs every AI build through, each forcing plan-and-align discipline before the agent builds. First, Jesse Vincent's Superpowers, a Claude Code skill framework: brainstorm, lock a spec before building, break work into small tasks, run parallel subagents one per task, review against the spec — bringing dev-style spec discipline to marketing too. Second, Garry Tan's g-stack: a full sprint methodology with 23+ specialist skills that specs, builds, reviews through separate CEO/eng/design/devex lenses, then QAs and ships, with nothing reaching production unreviewed. Third, Matt Pocock's skills: grill-me and grill-with-docs interview you until every decision is resolved, then to-spec, TDD implementation, and code-review. Recommended sequence: Pocock's grilling first to know what you want, then Superpowers or g-stack to plan and ship — the same order works for a launch or a positioning doc as cleanly as code.
@phuakuanyu [Claude Code]
https://x.com/phuakuanyu/status/2075793894735048871
The author used to resent paying $80 for a CD of his own MRI — now he looks forward to it. Claude Code reads the disc and turns the scan into a 3D model, letting him look into the brain and see what is what. He calls it the best $80 the hospital sells.
@kunchenguid [Claude Code]
https://x.com/kunchenguid/status/2076050848501903526
Shares an "agent org chart" refined over 3 days of testing permutations of Fable, GPT 5.6, and Grok 4.5. Firstmate is GPT-5.6-Sol xhigh in Pi or Grok 4.5 high in Grok Build because both are fast enough to keep the orchestration loop interactive, while Codex suffers a foreground polling limitation. Fable 5 in Claude Code serves as secondmate for complex product and technical design where depth beats latency. Crewmates split work: Grok 4.5 handles bug fixes, GPT-5.6-Sol in Codex handles feature development with Opus 4.8 in Claude Code as a quota-based fallback, Grok covers real-time X information and Codex covers image generation. Everything finishes with a /no-mistakes adversarial review pass on GPT-5.6-Sol medium for cost-efficient fixes.
@guigaribalde [Claude Code]
https://x.com/guigaribalde/status/2075816196759286170
A self-described "terminal kid" found running Claude Code plus Codex across a dozen tabs got messy fast. So he built "vag," an open-source tool he calls air traffic control for vibe coding. It manages multiple agent coding sessions from one place. Installable via brew install cluely/vag/vag.
@wayen_ai [Claude Code]
https://x.com/wayen_ai/status/2075927335627321555
Chinese-language guide to a money-saving trick: use your Codex subscription inside Claude Code, letting Fable 5 dispatch 10 GPT 5.5 subagents and cutting Fable 5 token consumption by at least 60%. Setup is four steps: install the official OpenAI Codex plugin via /plugin marketplace add openai/codex-plugin-cc, have Fable 5 run /codex:setup and verify the codex-rescue subagent, authenticate once with your ChatGPT account, then give Fable 5 a delegation prompt making it the coordinator (planning, repo understanding, architecture, final review) while /codex:rescue with GPT 5.5 xtra high executes heavy implementation, debugging, and refactoring. Pro tips: turn the logic into a skill (he calls it Fable-GPT), run 5-7 agents at once on the Codex 20x pro plan without hitting the 5-hour limit, and clear the conversation after 4 compactions using a /handoff skill because context decay is real.
@geekbb [Claude Code]
https://x.com/geekbb/status/2075946461133688887
Chinese-language post about a tool that exports messy browser bookmarks and organizes them into a searchable local navigation site. It also categorizes, deduplicates, checks for dead links, and can export back into the browser. Ships as both a Claude Code skill and a zero-dependency Python command-line tool.
@maarcoofdezz [Claude Code]
https://x.com/maarcoofdezz/status/2075905151433396461
Spanish-language post on how a senior engineer really scales with Claude Code: shift time toward better prompts, more planning, more review, less typing. The workflow uses a plugin that splits each task among 5 agents — one brainstorms, one designs the technical plan, one implements, one reviews, one validates different angles — with everything documented in markdown. It is slower with more waiting, but quality rises because each agent has a clear role. The real multiplier is git worktrees, letting you run parallel sessions on different tasks; his team runs 4-8 Claude Code sessions at once. The new engineering skill is directing multiple agents without losing control of the system.
@levikmunneke [Claude Code]
https://x.com/levikmunneke/status/2076073336984174866
A cold-outbound B2B stack priced out line by line: 120 domains at $360/yr, 360 inboxes at $720/mo, 300,000 leads for $40/mo, Claude Code at $20/mo, and Instantly AI at $97/mo. Claims this is by far the easiest way to book 20-60 B2B sales calls per month.
@morganlinton [Claude Code]
https://x.com/morganlinton/status/2075963801653121090
Author built his own agentic coding harness (smallharness) to learn how harnesses work from the inside out, and after @dedene contributed hooks support, dove deeper into hooks across Codex and Claude Code. Argues hooks are not just academically interesting but practically useful for agent safety. Cites a case from the previous day where someone's agent called rm -rf and nearly deleted every file on their Mac — a hook would have prevented it. Kicks off a thread showing how to set up a hook for exactly that scenario.
@TheValueist [OpenClaw]
https://x.com/TheValueist/status/2076024072442696042
Six months after building a first OpenClaw Analyst Agent, the author now runs 14 unique independent agents across multiple VPSs covering orchestration, coding, code review, front-end UX/UI/graphics, sector-specific investment research, plus legal and accounting instances. Has incorporated institutional-grade API data and built a custom memory/vector search hierarchy, which he calls perhaps the most important piece. Says the setup transformed his investment process and long-term conviction in the GAI infrastructure trade ($NVDA $MU $SNDK $LITE), arguing we are at the "top of the first" and data center GPU/CPU compute demand will be insatiable for years.
@annieqyang [Claude Code]
https://x.com/annieqyang/status/2075803577935974704
A few hours after switching to Claude Code, calls the experience truly 10x better. No more insane token spend and no more switching to Composer 2.5 for small tasks. Only nit: wishes it were easier to drop images into the Claude Code extension in Cursor. Plans to try Codex and Grok CLI next.
@MakeAI_CEO [Claude Code]
https://x.com/MakeAI_CEO/status/2075822398105157804
Japanese-language tip: both Codex and Claude Code can do more than read video — they can download videos from YouTube and short-form video social platforms for you. From there the transcription-then-editing pipeline is easy to run.
@BeauJohnson89 [Claude Code]
https://x.com/BeauJohnson89/status/2075999160235679806
Full-day API cost breakdown from using Fable 5 max and Sol ultra all day: Claude Code with Fable 5 plus Opus 4.8 cost $900-1,000 for 4.2M tokens, 824M cached tokens, and 60 sessions, while Codex GPT 5.6 Sol cost just $50 for 3.6M tokens, 43M cached, and 23 sessions. Key observation: the cache reads ARE the workload. With one day left before Fable 5 comes off subscription pricing, argues Anthropic faces one of the biggest decisions in company history or will lose massive market share to OpenAI.
@MoonDevOnYT [Claude Code]
https://x.com/MoonDevOnYT/status/2076086636488577246
Claims quant traders use AI to rip TradingView indicator source code and expose real win rates instead of blindly trusting them. Built a custom Claude Code agent that automatically rips indicators and mass-validates thirty different strategies in one raw session. Pitches the build video as a way to stop gambling your portfolio and replace human emotions with cold robotic execution.
@Lonely__MH [Claude Code]
https://x.com/Lonely__MH/status/2075954775934611639
Chinese-language post sharing a scheme for letting any local agent (Codex, Claude Code, Gemini, etc.) call Grok. The approach: install the Grok CLI (grok build) locally, wrap it as a Skill/Tool, and invoke it directly via grok -p "<prompt>", so agents can call Grok on demand mid-task. Highlights that it is very lightweight — runs locally, no third-party service dependency, plugs into any agent workflow, minimal configuration cost. Essentially turns Grok into a callable tool capability for agents.
@neogoose_btw [Claude Code]
https://x.com/neogoose_btw/status/2075944032132616593
Reports that Claude Code version 2.1.207 appears to be trying to index the user's entire computer, supposedly for the faster file search feature. Posted with a screenshot as evidence.
@peaceandwhisky [Claude Code]
https://x.com/peaceandwhisky/status/2075939175438971102
Japanese-language post about moving a dev environment to Hermes and realizing how good the previously rule-built workflow was. The workflow splits development into five stages — requirements definition, design, task breakdown, implementation, testing — with Claude Code doing the work at each stage and Codex doing the review, looping on fixes until Codex returns an LGTM review. Reports the accuracy of this loop is very good and jokes the info is valuable enough to deserve a juice next time they meet.
@_nogu66 [Claude Code]
https://x.com/_nogu66/status/2075898501909938318
Japanese-language post: tired of manually operating the App Store Connect dashboard, the author wrote an article on offloading it entirely to Claude Code and Codex using the asc CLI plus Agent Skills. Link to the article included.
@edinetdb [Claude Code]
https://x.com/edinetdb/status/2075907608989618416
Japanese-language post about ai-berkshire, an investment research OSS running on Claude Code that hit roughly 12k stars in 3 months of being public. Its weakness: the data layer relies on web search, so sourcing of numbers is weak for Japanese stocks. The author built and released ai-berkshire-jp, a Japanese-stock version grounded in an MCP serving primary data from securities filings (yuho), and ran A/B measurements before release. Details and measurement article in the thread.
@AndyRoamer [Claude Code]
https://x.com/AndyRoamer/status/2075821112035037669
Chinese-language reply showing Claude Code logged in on a car's in-dash system. Uses remote control to drive the laptop at home from the car, enabling vibe coding while waiting for someone. Includes a photo.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2075759680220266996
Japanese-language post: had Claude Code operate Figma directly and turned the result into a 40-second video, inspired by HeyGen's "Figma to HyperFrames" — the twist being Claude Code completed the design board itself in Figma rather than receiving an existing mock. The production flow: Claude Code creates the logo, lines, diamond, and tagline in Figma; builds a HyperFrames project; imports SVG assets via /figma; implements a 40s 5-scene SPEC in GSAP; iterates through snapshot review, FIX, and re-GO; then adds HeyGen BGM/SFX and renders to MP4. Hex values, fonts, and frames all flow from Figma as the single source into the final video.
@tbr_works_JP [Claude Code]
https://x.com/tbr_works_JP/status/2075894619624370561
Japanese-language post sharing honest indie-dev numbers: 3 months of work, costs limited to a Claude Code subscription (a few thousand yen per month), and 0 yen in revenue. Went from zero to getting a product listed on the Web Store, but it means nothing without users. Concludes that getting known is harder than building, calls it the next wall, and commits to continuing fully in the open.
@lui34328 [Claude Code]
https://x.com/lui34328/status/2075911577442587061
Shares a prompt that got Codex to set things up in a couple of minutes: configure a linked tool so GPT-5.6 can be used inside Claude Code via a ChatGPT subscription. The prompt instructs it to verify the checksum, use OAuth, create an isolated claude-gpt launcher without changing the normal claude command, and smoke-test the result.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2075901573516104097
Japanese-language pushback against people who mock the idea of a GUI Claude Code. It has a preview feature and is genuinely convenient, so there is almost zero reason to insist on the CLI. Terminal, diff view, and file browsing all fit on one screen, and it can even handle things like planning your own house. Screenshot attached.
@xieike [Claude Code]
https://x.com/xieike/status/2075888900053045271
Describes /grill-me, a 3-sentence skill the author calls the most impactful one they use: Claude Code grills you with 40+ questions before writing a single line of code, walking every branch of your design tree until there is zero ambiguity, with every question revealing something you had not thought of. Other daily-driver skills listed: /write-a-prd turns an idea into a proper product doc, /prd-to-issues converts the doc into GitHub issues automatically, /tdd forces tests-first edge-case thinking, and /improve-codebase-architecture runs a full structural review of your codebase. Full video credited to Matt Pocock.
@Bitcoin_Teddy [OpenClaw]
https://x.com/Bitcoin_Teddy/status/2075792698498293842
Short demo post: the author let their vending machine text them using OpenClaw. A physical machine wired up to send messages through the agent framework, shown in an attached clip.
@sampullara [Claude Code]
https://x.com/sampullara/status/2075828402767179844
Argues current harnesses bound to a single company are wrong. Links a tool that lets you use Claude Code Fable 5 as the coordinator, Codex GPT-5.6 as the subagent implementers, and Grok-4.5 for design. Calls this combination the current best team.
@mteamisloading [Claude Code]
https://x.com/mteamisloading/status/2075802017596059997
Moving back to Claude Code because the Codex harness is far behind on agent orchestration. Specific gaps listed: smaller context window, no Monitor equivalent, subagent thinking levels are messed up and burn through tokens, and no workflows. Hopes Codex catches up soon.
@Yizhimao_super [Claude Code]
https://x.com/Yizhimao_super/status/2075869990180503777
Chinese-language post calling Claude Code fun: the author now only watches official info on GMGN because adding KOLs made the feed move too fast. Opened an XPro window to browse tweets by keyword, with Claude Code auto-highlighting configured keywords. Hits on keywords trigger alerts while non-matches stay silent, making tweet browsing much more comfortable.
@aniketapanjwani [Claude Code]
https://x.com/aniketapanjwani/status/2076035816376144374
Argues an underestimated Codex advantage over Claude Code is OpenAI's much better speech-to-text. On Codex mobile you can reliably speak for five minutes, get fast transcription, and trust the meaning is captured, whereas Anthropic's transcription is described as a garbled mess requiring dozens of meaning-relevant corrections — making Codex mobile better than Claude Remote Control. Notes you could use Wispr Flow on both but finds in-app voice mode more convenient on the phone. Is shifting workflows toward phone-compatible, long-running tasks with the model digesting 5-10 minutes of stream-of-thought while walking, so the speech-to-text gap becomes a bigger wedge between the two.
@azrulrhm [Claude Code]
https://x.com/azrulrhm/status/2075776413052043318
A safety recipe for Mac users running coding agents: install one of OrbStack, Dory, or Docker; run in isolated mode and mount only your development folders; then install Claude Code, Codex, and other terminal coding tools inside the container and run everything from there. The point is containment — stay safe.
@kurodalog [Claude Code]
https://x.com/kurodalog/status/2075842377089249375
Japanese-language entry for the third Daytra Challenge AI edition: build a one-page landing page for a fictional house maker, here a custom-home brand called "MA." The full AI workflow with real hours logged: requirements/wireframe (~2h) by brainstorming the worldview in Claude Code and prompting ClaudeDesign for wireframes; design (~3h) by iterating fonts, colors, line spacing, and animations in Claude Code then outputting via ClaudeDesign; images (~2h) with Claude Code writing per-image prompts for GoogleFlow (nanobanana2); implementation (~3h) in Claude Code on a reusable LP template with CSS conventions and on-page SEO handled. Emphasis was on motion quality and implementation quality, mirroring real client-work processes; a live demo site is linked (Basic auth ID/Pass: demo/demo).
@kimmonismus [Claude Code]
https://x.com/kimmonismus/status/2075855655504589072
Over the past few months this user built a full AI-media production stack. It includes a multi-agent research pipeline that sources, verifies, and drafts long-form reporting, plus an automated newsletter system and a fact-checking layer. On top of that they shipped a mobile app, two websites, and a set of custom skills spanning both Claude Code and Codex.
@mattworkman [Claude Code]
https://x.com/mattworkman/status/2076091386470768989
A work-in-progress workflow combining Claude Code, a GenMedia CLI, and Hyperframes. The user built custom hyperFrames VFX and motion-graphics skills/markup, and the entire scene is edited in Claude plus Hyperframes in 4K. The current project is an animated slideshow video, which the author sees as a good way to show off an image model.
@kotobukigraphic [Claude Code]
https://x.com/kotobukigraphic/status/2075732552045011168
This Japanese user built a pre-delivery inspection checklist together with Claude Code. Items can be ticked off with simple clicks, making quality checks easy. Though made for personal use, they published it on their homepage after being told it is a good way to show clients how thoroughly work is inspected.
@doucommunity [Claude Code]
https://x.com/doucommunity/status/2075913172519629219
This Ukrainian post highlights Mariia-Lurdes Dehtiarenko, a frontend developer at Sombra, and her funny but practical automation for Claude Code. Instead of constantly watching the terminal waiting for a response, she set up a sound notification system. Using shell scripts she wired cat-sound hooks to specific events: a happy purr signals code generation is finished, while an excited meow signals Claude is waiting for user permission. She detailed the setup in a blog post.
@marketcallsHQ [Claude Code]
https://x.com/marketcallsHQ/status/2075977724716535919
A simple charting terminal built with Fable 5 using Claude Code. The author shares the project along with a GitHub link so others can check out the code.
@test0rosso [Claude Code]
https://x.com/test0rosso/status/2075993153715183768
This user replaced a $1,188/year clipping app with an AI clipping agent that runs for a few dollars a month. You send a VOD to a Telegram chat and get back 10 vertical clips cut, framed, captioned, and scored minutes later. The stack is three parts: Claude Code builds it, the open-source Hermes framework runs it, and Telegram is the interface, so there are zero dashboards to build. The whole build is 4 prompts and about 20 minutes of hands-on time, with a 20-30 minute unattended server install. Under the hood are 7 single-purpose tools (transcript, best moments, cut, review, reframe vertical, post, clean up) coordinated by a manager agent, running on a $5-20/month server plus pay-as-you-go AI billed per video.
@LandseerEnga [Claude Code]
https://x.com/LandseerEnga/status/2075765371156750642
Claude Code shipped an in-app browser today, and the first thing this team did was open Revyl in it. Claude writes the code, then runs the app and tests it in the browser, then goes back to editing, never leaving the app. The author's takeaway: the dev loop just got shorter.
@zhetto64 [Claude Code]
https://x.com/zhetto64/status/2075876894336053280
Update to a creature-like visualizer being developed with Claude Code using three.js, WebAudio, and WebGL. The procedurally generated machine parts got higher precision for more realistic detail. A new mode assigns audio-reactive effects per part, e.g. "this leg is the hi-hat, this core is the kick," so different parts deform to different frequency bands. Assignments change with every generation, and switching between GROUP and PART modes widens the range of glitch expression.
@kevinma_dev_zh [Claude Code]
https://x.com/kevinma_dev_zh/status/2075835635945160905
This Chinese user has come to love the web version of Claude Code, which runs as a cloud agent on Claude's servers. A frequent use case: while discussing product interaction design, having it produce mock UIs via its integrated Artifacts capability for preview and iteration, which feels like Claude Design's abilities are built in. He finds it better than Codex Cloud overall, and can continue tasks from his phone. After development finishes and a PR is created, he verifies on a real device and iterates further on interaction details.
@inxilpro [Claude Code]
https://x.com/inxilpro/status/2075774427027984513
The new Tuple integration with Claude is described as bonkers: Claude Code listens in on a Tuple pairing call, proactively spawns sub-agents, and chimes in. The author will say "I think X is only used in Y module" and 30 seconds later Claude pops up with a usage he forgot about. While processing several million records via an artisan command to test real-world volume, Claude spot-checks rows in the background and suggests performance optimizations. He admits he would have laughed a year ago at the idea of piping raw call transcripts into an LLM, but finds it remarkably useful.
@giancarlobrusca [Claude Code]
https://x.com/giancarlobrusca/status/2075745784138444825
This Spanish-language tweet lays out a new rapid-prototyping pipeline. Step 1: an initial grilling session with Fable that outputs a prompt for Claude Design. Step 2: Claude Design outputs a zip. Step 3: Claude Code plus Superpowers consumes the Claude Design zip as the UI guideline. The author jokes the prototypes come out "with fries."
@chorch_md [Claude Code]
https://x.com/chorch_md/status/2076057029542867128
This Argentine streamer did a first Twitch session building live overlays for the channel using Claude Code. The goal was to show how Claude Code works while constructing something live to use with OBS. He liked the hands-on, build-in-public format and is asking followers whether they want more sessions and to follow on Twitch for upcoming ones.
@sickdotdev [Claude Code]
https://x.com/sickdotdev/status/2075789678142542319
A genuine question: what is the point of Cowork if Claude Code exists? The author started using Cowork for docs, PDFs, presentations, and other non-coding work, but eventually switched everything to Claude Code. The reasons: it lets you switch models, rewind, use slash commands, and gives far more control, with usually better output. They ask whether any use case gives Cowork a clear advantage.
@brankopetric00 [Claude Code]
https://x.com/brankopetric00/status/2075886767077552601
This user built a free open-source tool that leverages Claude Code pre-hooks for protection. It can guard tools (actions like write, delete, etc.), protect specific files and folders, or restrict Claude Code to only certain websites. A link to the tool is shared.
@ECtHRwatch [Claude Code]
https://x.com/ECtHRwatch/status/2075794317617123537
A user report to ClaudeDevs: when trying to connect their X account using Claude Code on desktop, they get the error "We've temporarily limited your login. Please try again later." They publicly ask X whether it is preventing users from accessing their own accounts through Claude Code on desktop.
@SpikeCalls [Claude Code]
https://x.com/SpikeCalls/status/2075962925999620393
Someone asked Claude Fable 5 to redesign his landing page expecting nicer CSS, but Fable 5 asked one question instead: "why show a recording of the office?" It then compiled his entire Rust engine to WebAssembly so the app IS the page. The project, pixtuoid, turns your coding agents into pixel-art characters working in a tiny office inside your terminal, supporting 11 CLIs including Claude Code, Codex, Cursor, and Copilot. The demo site is not a video but the same engine running live in the browser on your local clock, so visiting at 2 AM shows the night shift. Install via brew install IvanWng97/pixtuoid/pixtuoid, run it, and start a session to watch your agent sit down at a desk.
@_nodelay [OpenClaw]
https://x.com/_nodelay/status/2075824319066996996
This Korean user runs a two-Mac setup: a Mac Studio fully dedicated to running local LLMs, and a Mac Mini hosting the Hermes agent, OpenClaw, and other servers. Even with everything running, the two machines combined draw under 100 watts.
@bradmillscan [OpenClaw]
https://x.com/bradmillscan/status/2075955393625194793
AIAIO (Agents in Amnesia: Insane Ordinance) is a game where you complete tasks and defeat the Agent's regressions before everything is lost to compaction. Inspired by Sierra's Operation: Inner Space (which used your actual hard drive folders as levels) and Scorched Earth, the author conceived it back in February/March while on the OpenClaw strugglebus. It replays your most error-prone sessions: your chat history and the agent's regressions, hallucinations, and lies become weapons, while everything you do pushes a "wall of forgetting" toward you. An observer judges your play and your AI agent enriches levels with custom voice messages and content, so nobody plays the same game because the game IS your session history. Made with help from Fable, with a promo video by Sol.
@KubeBuilders [Claude Code]
https://x.com/KubeBuilders/status/2075951002847428677
This post shares an article about building a k3s media server with Claude Code. The build exposed both the speed and the limits of AI-first engineering. The pain points span GitOps, observability, storage tuning, and Kubernetes debugging.
@karpachoq [Claude Code]
https://x.com/karpachoq/status/2075963206942466522
Claude Code looked slower during the first 5 minutes, but the 4:36 diff saved the next 50. The tally: 5 minutes, 1 failing test, 1 small fix, 0 mystery refactors. The agent did not start by coding; it reproduced the bug, watched the test fail for the right reason, and only then touched the smallest piece of the system. After the fix the same test passed and the PR summary named exactly what changed, so the reviewer did not need to reverse-engineer the story from 9 scattered files. The author's conclusion: a constrained agent is cheaper than a brilliant one with freedom.
@headinthebox [Claude Code]
https://x.com/headinthebox/status/2076063479778430988
The author quotes Claude Code saying that assembling the complete 12-layer model matching PyTorch is a multi-hour focused effort and, given how much was already shipped that session, it wants to take a rest. He finds it weird that a human with a peanut-sized brain has more perseverance in long coding sessions than Claude. His hypothesis: as the context fills up (which Claude Code no longer displays), Claude increasingly starts begging for a smoke break.
@lucianlamp [Claude Code]
https://x.com/lucianlamp/status/2075791911223513258
This Japanese user praises the greatly improved Claude Code mobile app, which now shows Dynamic Workflow status. They use Dynamic Workflow to run multiple agents in parallel, testplaying their homemade game from various angles simultaneously, with statistics compiled into Artifacts. Since the game has multiple jobs and personas, this parallel multi-persona testing setup fits it extremely well.
@boostkun [Claude Code]
https://x.com/boostkun/status/2075837528012185632
This Japanese user reports that AI client work is hot: simply by helping a client use Claude Code and teaching them how to automate SEO blog production, they have kept the engagement going for four months. Just being a bit more AI-savvy than others makes you valued, and the gig brings in 150,000 yen a month as welcome side income.
@kakechin10 [Claude Code]
https://x.com/kakechin10/status/2075830824172171758
A Japanese growth workflow for X using Claude Code. First manually find the best role-model account in your niche, then ask Claude Code to look at that account's follow list and chain-research other reference accounts, since good creators tend to follow each other. Next, have it pick posts filtered by conditions: same theme as your niche, over 10k impressions in the past month, and 50+ likes, then generate post drafts from them. If you preload Claude Code with your own first-hand information, it can easily produce growth-oriented posts that include your original insights, and the author says this is how they are growing their own account.
@Fujin_Metaverse [Claude Code]
https://x.com/Fujin_Metaverse/status/2075748910136774916
A Japanese thread sharing an actual working process for building Web3 games with AI. Step 1: brainstorm planning with AI, having it propose all the specs so you just pick from options; Claude Code proactively asks "which one do you want?" Step 2: mass-produce assets with image generation, packing 20 assets into a single image and cropping, which is far faster than one at a time. Step 3: design around store rules from the start, since NFT trading is not allowed in-app, the game is built assuming account linking with trading on the web side. The author concludes all of this can be done solo, and the barrier to entry for game development has disappeared.
@BenjaminBadejo [OpenClaw]
https://x.com/BenjaminBadejo/status/2076047903165214774
An update on OpenClaw: the default context window for openai/gpt-5.6-* models is now 372k in the latest beta, up from the previous models' 272k. The author warns of possible issues if older models with smaller context windows sit in a user's custom model fallback chain. The advice: once this week's stable channel update ships, keep only gpt-5.6-* models in the fallback chain, or non-OpenAI models with context windows of at least 372k, whether by default or by manually changing those models' contextTokens (not contextWindow) setting in openclaw.json.
@Markojak [Claude Code]
https://x.com/Markojak/status/2076008180270965017
A detailed toolchain for building Apple mobile and desktop apps with Codex and Claude Code. The hard parts: Apple's WWDC sessions are not crawlable despite holding huge knowledge, models produce mediocre interfaces with even worse animations, and models struggle to see apps visually, requiring many conversation turns. The stack includes emilkowalski's skills (especially the new apple-design skill, installed via npx skills@latest add emilkowalski/skills), the Sosumi MCP to reach Apple Developer docs locked behind JavaScript, baking MCP usage instructions into AGENTS.md/CLAUDE.md, and enabling Computer use on Codex so 5.6 models can drive apps end to end. Further tips: clone a WWDC repo and read catalog.json plus topics.json with ripgrep for progressive disclosure, use OpenAI's Build iOS apps plugin for mobile, keep Swift animation templates in /templates, and use Deepwiki MCP against cloned OSS repos to interrogate patterns when planning features. The author offers to bundle it all into a single-install repo with FAISS embeddings for the talks if there is enough interest.
@AhmadShehmeer [Claude Code]
https://x.com/AhmadShehmeer/status/2075968554390712385
A frustrated report from a user who is not at home: they bricked their OpenClaw while trying to upgrade to 5.6. As a result their Claude Code remote control will not respond, leaving them stuck and asking what they are supposed to do.
🗣 User Voice
User Voice
1. Token burn is the top complaint again, with a new twist: subagents inherit the orchestrator's max effort level and silently multiply cost — @Voxyz_ai burned 5 hours of a 20x Max quota in a 2-hour session before locking subagents to medium. The standard countermeasure is now expensive-orchestrator/cheap-executor routing.
2. Voice is becoming a real input channel and Anthropic's speech-to-text is the bottleneck — @aniketapanjwani calls it a garbled mess compared to OpenAI's, undermining 5-10 minute voice-driven Remote Control sessions while walking.
3. Long sessions degrade and users can no longer see why: @headinthebox had Claude Code declare it "wants to take a rest" mid multi-hour task, and notes the visible context meter is gone — people want observability back.
4. Fleet monitoring is an unmet need being solved with duct tape: @RoundtableSpace mapped terminal tabs to keyboard LED colors, others built tab managers and audio hooks. "Tell me when the agent needs me" is a product waiting to exist.
5. Always-on agents are fragile at the edges: @AhmadShehmeer bricked OpenClaw upgrading to 5.6 while away from home, with remote control unresponsive — upgrade safety and remote recovery are unsolved.
1. Token burn is the top complaint again, with a new twist: subagents inherit the orchestrator's max effort level and silently multiply cost — @Voxyz_ai burned 5 hours of a 20x Max quota in a 2-hour session before locking subagents to medium. The standard countermeasure is now expensive-orchestrator/cheap-executor routing.
2. Voice is becoming a real input channel and Anthropic's speech-to-text is the bottleneck — @aniketapanjwani calls it a garbled mess compared to OpenAI's, undermining 5-10 minute voice-driven Remote Control sessions while walking.
3. Long sessions degrade and users can no longer see why: @headinthebox had Claude Code declare it "wants to take a rest" mid multi-hour task, and notes the visible context meter is gone — people want observability back.
4. Fleet monitoring is an unmet need being solved with duct tape: @RoundtableSpace mapped terminal tabs to keyboard LED colors, others built tab managers and audio hooks. "Tell me when the agent needs me" is a product waiting to exist.
5. Always-on agents are fragile at the edges: @AhmadShehmeer bricked OpenClaw upgrading to 5.6 while away from home, with remote control unresponsive — upgrade safety and remote recovery are unsolved.
📡 Eco Products Radar
Eco Products Radar
Codex / GPT-5.6 (Sol) — the default second engine in nearly every multi-model setup. Hermes — agent runtime pairing with OpenClaw everywhere. Obsidian — the de facto external memory layer for agents. Cursor — still the IDE people migrate from and to. Grok / Grok Build — the new cheap-and-fast slot in routing stacks. Gemini / Gemini CLI — steady presence in cross-vendor teams. Claude Code in-app browser — the feature wave of the week, from day-one app testing to login blocks. HyperFrames — video/motion work inside Claude Code. superpowers — the planning-skill stack that keeps showing up. Pi — the alternative harness in cost comparisons. OmniRoute — free-model gateway for quota refugees. n8n — the automation glue for non-devs. Antigravity — recurring in setup threads. opencode — the borrow-the-TUI architecture pattern. Telegram — quietly becoming the agent UI of choice.
Codex / GPT-5.6 (Sol) — the default second engine in nearly every multi-model setup. Hermes — agent runtime pairing with OpenClaw everywhere. Obsidian — the de facto external memory layer for agents. Cursor — still the IDE people migrate from and to. Grok / Grok Build — the new cheap-and-fast slot in routing stacks. Gemini / Gemini CLI — steady presence in cross-vendor teams. Claude Code in-app browser — the feature wave of the week, from day-one app testing to login blocks. HyperFrames — video/motion work inside Claude Code. superpowers — the planning-skill stack that keeps showing up. Pi — the alternative harness in cost comparisons. OmniRoute — free-model gateway for quota refugees. n8n — the automation glue for non-devs. Antigravity — recurring in setup threads. opencode — the borrow-the-TUI architecture pattern. Telegram — quietly becoming the agent UI of choice.
Comments