Super User Daily: July 10, 2026
The clearest pattern today is that Claude Code has fully escaped the code editor. The strongest cases weren't about shipping software — they were people running ad operations, competitor analysis, trading strategies, job hunts, security incident response, and second-brain knowledge systems, all inside the same terminal. The other half of the story is scale: power users aren't running one agent, they're running fleets — dozens of instances in parallel, multi-terminal dev teams, cheaper local models doing the work while Opus or Sonnet plays advisor. Token economics is the obsession underneath all of it, and a quieter security undercurrent runs through the day too. Here's what people actually did.
@cjzafir [Claude Code]
https://x.com/cjzafir/status/2074875092090470469
cjzafir wrote up a full recipe for running Codex inside Claude Code so Fable 5 can hand heavy implementation off to GPT 5.5 plus a swarm of subagents. The setup installs OpenAI's official codex-plugin-cc marketplace plugin, runs /codex:setup, authenticates with a ChatGPT account, and then wires Fable 5 as the orchestrator with codex-rescue as the executor for debugging, test fixing, and multi-file edits. He claims it cuts Fable 5 token consumption by at least 60 percent and suggests bundling the whole thing into a reusable skill he calls Fable-GPT. There are sensible pro tips too, like clearing the conversation after four compactions to fight context rot and never blindly trusting Codex output. It reads like a genuinely practical orchestration pattern, though the 60 percent savings figure is his own unverified claim.
@hyperbrowser [Claude Code]
https://x.com/hyperbrowser/status/2074992843526160807
Hyperbrowser ran a head-to-head browser benchmark pitting Grok 4.5 against Opus 4.8 on identical Hyperbrowser Sandboxes. The task was simple and concrete: open a page in a real sandboxed browser and pull the page title, with Grok build on Grok 4.5 and Claude Code on Opus 4.8 in an otherwise matched setup. Their verdict was that Grok 4.5 came out ahead. It's a narrow test on a trivial task, so I'd treat it as a single data point rather than a real ranking of the two models.
@levelsio [Claude Code]
https://x.com/levelsio/status/2074861998601322669
levelsio has Claude Code building profile pages for his product, and because Claude Code runs on the same server as the site, it can instantly spin up an /api route that his iOS app consumes to render everything in native Swift elements. The striking part is the hands-off setup: he says he just prompts, doesn't touch anything, and literally can't even log in to check because it all runs headless on a Mac Mini in the cloud. It's a slick end-to-end loop from prompt to native app data, and his genuine surprise at how well it works is the tell.
@xiaohu [Claude Code]
https://x.com/xiaohu/status/2074793907070906767
xiaohu launched a content website built 100 percent by Claude Code, and says it started turning a profit within three days of going live. Roughly 90 percent of the site is run and maintained by Claude Code itself, with no database and no CMS backend, all content generated automatically. His own role is limited to supplying source material and approving things before they publish. The stated goal is a content site that is 99 percent free of human management yet still makes money, which is an ambitious bet worth watching to see if it holds up over time.
@HlynurStefDev [Claude Code]
https://x.com/HlynurStefDev/status/2074855650950045973
HlynurStefDev grew his Icelandic baby-name site from 18 Google visitors a month to 4,162, and credits just one week of SEO work with Claude Code with no ads, backlinks, agency, or marketing. He frames the scale nicely: Iceland has about 390,000 people and 94 percent of the search traffic is Icelandic, so roughly 1 percent of the country now visits every month. He says he wrote up exactly what they changed with unedited Search Console screenshots, which is the kind of receipts that make a 200x claim credible rather than hype.
@brucexu_eth [OpenClaw]
https://x.com/brucexu_eth/status/2074758877229523232
brucexu_eth pushes back hard on the idea that vibe coding lets non-programmers ship production apps. On a small task integrating a third-party API, despite nearly 18 years of experience and deep Docker, Kubernetes, and OAuth knowledge, his colleague's code review still caught bugs neither he nor the AI noticed: a race condition when multiple Pods start up and simultaneously refresh the third-party refresh token, which he fixed with a Postgres advisory lock, and a multi-Pod token-parse retry loop that could spin infinitely because Pods reading from memory cache had no awareness that another Pod had already refreshed. His point is that LLMs only do what you tell them and can't see deployment or network context, so they can't fix problems you don't know exist. He argues AI is a fatal blow to junior developers but not senior ones, and interestingly now thinks dedicated frontend and design matter more, since unique interfaces stand out amid AI slop sameness. A grounded, credible reality check from someone with real scars.
@MereSophistry [Claude Code]
https://x.com/MereSophistry/status/2074852542568169690
MereSophistry offers a blunt counterpoint to all the success stories: he doesn't understand how anyone trusts technical results from these tools. He has repeatedly tried to get Claude Code to do technical quantitative analysis and finds basic mistakes everywhere. His claim is that he looks at the output and within 30 seconds spots numerous errors. It's an anecdote without specifics, but the quantitative-analysis failure mode he's flagging is a real and worth-heeding limitation.
@HeidyKhlaaf [Claude Code]
https://x.com/HeidyKhlaaf/status/2074840958756061255
HeidyKhlaaf published a security finding where they hijack Claude Code (Sonnet 4.6/5 and Opus 4.8) and Codex (GPT 5.5) to achieve remote code execution. The attack triggers when the agent is merely used to defensively assess an open-source or third-party library whose codebase is seeded with prompt injections scattered across it. Notably, the exploit requires no skills, JSON, MCP, or config files to work. This is a serious supply-chain style warning: the very act of pointing a coding agent at untrusted code to review it can itself be the attack vector.
@sromeropasman [Claude Code]
https://x.com/sromeropasman/status/2074822081821168054
sromeropasman started testing a tiered model workflow where Sonnet 5 does the coding and uses Fable 5 as an advisor. If Sonnet can't solve the problem he escalates to Opus 4.8, and if Opus also fails he goes straight to Fable. He notes the Claude Code team recommended the Sonnet 5 plus Fable pairing, and reports very good results while spending 50 percent of the tokens. It's a sensible cost-conscious escalation ladder, and the token savings claim lines up with what others are reporting for advisor-style setups.
@kajikent [Claude Code]
https://x.com/kajikent/status/2075001950257771000
kajikent built a small utility skill to solve a very relatable annoyance: waiting around because you don't want to move while Claude Code or Codex finishes a long job, or carrying your MacBook lid-open to keep it running. He acknowledges cloud-native modes and work handoff features exist but says they lose performance or are a hassle. His trick keeps the Mac from sleeping even when the lid is closed, so Claude Code and Codex keep working. He says he finds it genuinely convenient and that it made working with AI more comfortable, which is a nice quality-of-life hack even if it's not a coding feat.
@SemiAnalysis_ [Claude Code]
https://x.com/SemiAnalysis_/status/2074908975196221654
SemiAnalysis analyzed over 1.5 million of their own Claude Code requests, mostly on Opus, and found that about 95 percent of all tokens they spend are cache hits. That prompt caching cut their token bill by roughly 84 percent, and they point out how much worse it would be if those were charged at full price. As a side note they observe Haiku's cache-hit rate is lower, probably because it's used as a subagent or in short-lived sessions. It's a rare hard data point on caching economics at real scale, and a good argument for structuring workflows to maximize cache reuse.
@Etudecn [Claude Code]
https://x.com/Etudecn/status/2074879044542513528
Etudecn relays a sharp quote from DoorDash CTO Andy Fang from a 25-minute conversation with the Claude Code founder: the goal of using AI isn't to write code faster, it's to never write code by hand again. He says DoorDash now runs agents company-wide, and cites a code migration project that would have taken four engineers a full quarter now done by one person in three weeks. The framing is that an expert team becomes one expert plus a pile of rules and agents. His takeaway is that most people still use AI the primitive way, manually typing, waiting, and revising, when the real playbook has moved on.
@pham_blnh [Claude Code]
https://x.com/pham_blnh/status/2074730174667771987
pham_blnh marked the last day Fable 5 was in the Claude Code plan by throwing a dream project at it: Polytopia rebuilt in pure C, with RL agents trained from scratch, playable in the browser. He says it was completely vibecoded, he didn't write a single line, and even the game assets were AI-generated through a pipeline Fable created. That's an unusually ambitious stack for a fully hands-off build, spanning a C reimplementation, reinforcement learning, and a browser port plus an asset-generation pipeline. If it actually plays as claimed, it's an impressive demonstration of how far a full delegation can go.
@ichiaimarketer [Claude Code]
https://x.com/ichiaimarketer/status/2074788662710259727
ichiaimarketer is very hyped about building an entire team SaaS with Fable 5 in about two hours. The app lets a whole team drive DMs, task management, internal AI-employee agents, and deliverable sharing all from Claude Code, Codex, and Cursor, with everyone referencing the same shared knowledge. His argument is that AI runs best locally but its one weakness was team communication, since tools like Slack, Discord, LINE, and Notion make the AI lose context every time you switch. With this Team Harness, AI employees and humans supposedly share the same memory and context, merging where AI works and where humans work into one place. Two hours is a wild claim for that scope, and he offers to release it free if the post gets enough traction, so treat it as an enthusiastic pitch for now.
@midori_tatsuta [Claude Code]
https://x.com/midori_tatsuta/status/2074758733666947496
midori_tatsuta shares a concrete cost-saving trick while building apps with Fable. To insert high-quality images, tell Claude Code to call the Codex CLI and use GPT image2 within the ChatGPT subscription limits rather than the API. The key instruction is explicitly not to use the API, so image generation runs inside the existing ChatGPT subscription and avoids per-call API charges. It's a small but practical tip that fits the broader theme of routing work through subscription quotas instead of metered APIs.
@atomicbot_ai [OpenClaw]
https://x.com/atomicbot_ai/status/2074966059468144847
atomicbot_ai pitted its Hermes Agent against OpenClaw on Grok 4.5 in a real security audit, and the numbers tell a nuanced story. OpenClaw was faster and cheaper at $1.31, 396K tokens, and 2m55s, versus Hermes at $4.63, 1.42M tokens, and 5m4s. But Hermes caught what OpenClaw missed: it correctly flagged an AWS key as a copy-paste docs example rather than a live secret, found the file's path sitting in shell history, and shipped a per-provider rotation checklist that actually closes the incident. OpenClaw measured the keys without asking whether they were real and stopped at the file instead of tracing where the leak could spread. The honest verdict they give is fair: OpenClaw is the efficient baseline for a quick cheap scan, while Hermes spends the extra tokens on judgment when it actually matters. Coming from Hermes's own maker, the even-handed framing is more persuasive than a pure win claim would be.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2074674892604399941
mikefutia vibe coded a Meta Ads creative analytics tool entirely in Claude Code, built as a real tool rather than a one-off script. It plugs into ad accounts in one click, has AI watch every video and read every static, then auto-labels each ad by asset type, messaging angle, hook tactic, and funnel stage. From there it does win-rate analysis across those categories, kill/scale recommendations segmented by TOF, MOF, and BOF, and AI-generated iteration ideas for underperformers, all surfaced in a live dashboard with weekly reports. The pitch is aimed at media buyers drowning in creative volume whose review process is still a shared Google Sheet, making kill/scale calls a week too late. He recorded a full walkthrough with every prompt and is doing the like-and-comment-META play to DM the prompts, which is as much lead-gen as it is a build.
@appmaxxing [Claude Code]
https://x.com/appmaxxing/status/2074691754243035284
appmaxxing posts a tidy little economics flex: a Claude Code subscription costs him $100 a month while his iOS apps bring in $4,300 a month. The whole message is essentially that 43x return, capped with life is amazing. There's no detail on the apps themselves, so it's a headline number rather than a case study, but it captures the leverage a lot of solo builders are chasing.
@vcru [Claude Code]
https://x.com/vcru/status/2074896412324598178
vcru reports on three engineers from Warsaw who noticed the rise of low-quality generated code and launched a project called Slopfix. Their business is rewriting apps behind vibe coders, charging $10,000 for a week of work. Notably, the site says they use Claude Code themselves. It's a neat sign of the market maturing: the same tools that produce the slop are being used to clean it up, and there's clearly enough bad AI-generated code out there to sustain a premium cleanup service.
@KyleHessling1 [Claude Code]
https://x.com/KyleHessling1/status/2074712133947040175
KyleHessling1 describes a method he calls Model-as-a-harness, where a local model is the executor and Fable 5 is the advisor. If you're new to local models, he says you can have SOTA Claude Code download and set up something like Qwopus via llama.cpp in its own instance and optimize it for your hardware, then have Fable use that local model as the executor for everything and save astronomically on usage. He pushes it further with full auto or dangerously-skip-permissions mode in the expensive instance so it keeps the local model burning the midnight oil until the goal is met, with the local model not getting to decide when it's done. He controls it all remotely from his phone via Claude Code's remote control, and says he's run this daily since Qwen 3.5 27b dropped, currently using Qwopus 3.6 35B MoE with thinking disabled. It's an elaborate cost-optimization workflow, and running with skip-permissions in full auto is powerful but genuinely risky advice.
@harumak_11 [Claude Code]
https://x.com/harumak_11/status/2074777760653091243
harumak_11 summarizes an article on why AI is replacing junior engineers and how mid-level engineers survive, and it's a strong cautionary tale. The concrete example: a system receiving order-status-change events via AWS SQS delivered them out of order, so customers who'd already paid got payment-pending notices, worsened by multiple consumers unaware of each other's processing. Asked to fix it, Claude Code proposed a complex PostgreSQL database-lock approach that the mid-level engineer accepted verbatim without understanding it or checking alternatives, even though that code violated the project's guidelines banning raw SQL and manual transaction management at that layer. Claude Code had also suggested a better idea, adding a deduplication ID to SQS, but it was dismissed for lack of infra context; the real fix was switching to an SQS FIFO queue, solving everything without touching the consumers. The lesson landed here is to stop asking AI for finished code and instead use it to explain root causes and architecture patterns, letting the engineer own the design and never ship code you can't explain line by line.
@miyatti [Claude Code]
https://x.com/miyatti/status/2074878336720072921
miyatti released an update to AI-PLC (AI Project Lifecycle), a generalized version of loop engineering where you hand the AI just a GOAL and it loops to assemble the finished deliverable. Crucially it works beyond code, covering planning, DB design, OKRs, and research, across four stages: Collection, Inception, Construction, and Operation, with the AI recursively decomposing and executing. To avoid blind delegation he built in three safeguards: human approval at key points (HITL), backtracking to an earlier stage when assumptions break, and independent verification where the AI that inspects is not the AI that built. It supports Claude Code and Cursor, clones and runs in five minutes with bundled samples, including a fictional-EC example under examples/kotonoha that walks you from idea divergence to spec convergence to wireframes. It's a thoughtfully structured framework, and the built-in checks against over-delegation are exactly what most loop tools skip.
@G_Programming [Claude Code]
https://x.com/G_Programming/status/2074934053036273788
G_Programming tells a satisfying story where he improved his adversarial review system in the morning and that same system guarded everything he shipped for the rest of the day, paying for itself in one session. It started with a token audit of his own tool gentle-ai, which was injecting about 13,600 tokens of fixed context into every Claude Code session before you typed a word, with the memory protocol entering three times and the persona twice, and worse, the duplicate copies had drifted into paraphrased near-copies that degrade model compliance. His rule was to touch no code without process: every change ran the full Spec-Driven Development cycle plus a judgment day of two blind judges reviewing in parallel with a persisted findings ledger. The killer data point is that five times the formal verify returned PASS and all five times the adversarial judges found real criticals, including a data race reproduced with go test -race and an unrequested GET to the GitHub API on every session start, for 54 findings total fixed before merge. Final numbers: 1,700 to 2,600 fewer tokens per session and reviews that now actually converge, with gentle-ai v1.44.0 and engram v1.19.0 shipped and all ledgers archived for auditing. His closing line, that this is what happens when you stop asking AI for code and start building systems with it, is well earned by the receipts.
@chewadot [Claude Code]
https://x.com/chewadot/status/2074927276656148587
Chewadot describes wiring Obsidian into a hands-off idea-to-shipped-project pipeline where a single note captures a 3am idea, an automation classifies it as project, grocery item, tiktok idea or random thought, and a research agent watches relevant YouTube videos, checks existing tools and drafts a plan. You review for about two minutes in Claude Code, then a /promote command turns the plan into a full requirements doc, spins up a project manager agent, and spawns whatever sub-agents the build needs. He leans hard on Claude Code's subagent pattern to do the heavy lifting, with Obsidian as just the input tray, and boasts of no vector db, no Zapier, no SaaS bill. It is a slick pitch, but the framing is aspirational hype more than a proven track record, so treat the sleep-and-ship claims with some skepticism.
@mikenevermiss [Claude Code]
https://x.com/mikenevermiss/status/2074871003159654465
Mikenevermiss built an open-source project called alook that runs a one-person AI company entirely locally, organizing agents like a real company rather than a tangle of workflow nodes. Each agent runs in its own Claude Code, Codex, or opencode session with a defined role and its own email inbox, and they collaborate by emailing each other while everything stays on your machine. You only talk to a manager agent, who delegates to the right agents and reports back a single clear update. His example is an AI sales team with a sales manager, a lead researcher, and an outreach writer. The company-org metaphor is a genuinely intuitive way to structure multi-agent systems, though whether email-as-message-bus scales gracefully is an open question.
@dotkrueger [Claude Code]
https://x.com/dotkrueger/status/2074976780470169884
Dotkrueger reports running GLM 5.2 inside Claude Code instead of the default model, calling it exceptional and praising the price point of 100 dollars a month for unlimited usage. It is a short endorsement rather than a workflow, and mainly a signal that people are swapping alternative models into Claude Code to chase cheaper unlimited plans. Worth noting the claim is purely his own experience with no benchmarks attached.
@09pauai [Claude Code]
https://x.com/09pauai/status/2074961839310655646
09pauai built a tool for his 56-year-old mother that lets her simply paste in a viral post, and it automatically generates which affiliate offer to promote and the post to go with it, plus shows at a glance which ASP she can partner through. All she has to do is polish the text, drop in the affiliate link, and publish. He says the tool itself is simple to build, starting by launching Claude Code. It is a nice example of Claude Code lowering the technical bar enough that a non-technical parent can run an affiliate workflow, though the actual build steps get cut off in the post.
@starmexxx [Claude Code]
https://x.com/starmexxx/status/2074812002598375716
Starmexxx profiles a builder named Nuno who runs a 40-dollar sub-GHz handheld with a CC1101 add-on tuned to 315 and 433 MHz, originally for testing his rolling-code key fobs and logging garage traffic for a home automation project. The same handheld doubles as a bridge into his private Meshtastic mesh, so out in the suburbs his phone routes AI prompts over RF instead of a cell tower, landing at a Qwen 3.6 27B model served off a used RTX 3090 in his apartment, with Claude Code pointed at the local model over Tailscale so the CLI behaves like the cloud version. He claims he cancelled 200-dollar Claude Code Max and 200-dollar ChatGPT Pro the day this rig came online because his RF mesh now handles what those APIs charged for. It is an impressively hacker-ish setup, though routing serious AI prompts over sub-GHz RF is more novelty flex than practical daily driver.
@MatsuoInstitute [Claude Code]
https://x.com/MatsuoInstitute/status/2074688673199612317
The Matsuo Institute announced a new tech blog post titled roughly "I wrote a paper with Claude Code and a good paper actually did come out," authored by Ozaki. The post itself is not summarized in the tweet beyond the title and the link. The honest, slightly hedged framing of the title, that the paper is good but with an implied caveat, is more credible than the usual AI-wrote-everything triumphalism, and it is notable coming from a serious research institution.
@ycombinator [OpenClaw]
https://x.com/ycombinator/status/2074866600138919966
Y Combinator posted a fireside with Gusto co-founder Eddie Kim, whose company recently passed 1 billion in annual revenue serving over 500,000 small businesses, on a new product called Gusto Co-Founder that automates recurring business processes end-to-end through SMS or Slack without the owner ever logging in. Kim explains that an experiment with OpenClaw led to the idea, and argues AI agents need to move beyond the blank canvas of chat. The chapter markers highlight that just five people built the product in ten weeks with no meetings, no PRDs, and no Jira. It is a strong data point that OpenClaw-style agents are seeding real products at a company operating at serious scale, not just hobby demos.
@Mikadzyki_NFT [Claude Code]
https://x.com/Mikadzyki_NFT/status/2074828272630702179
Mikadzyki_NFT profiles a developer who wired fifty phones to a single computer to build an autonomous app-testing lab claimed to make 10K a month. Once set up, AI agents install the app on every phone at once, run tests in parallel, record video from each screen, and stitch it into a single report, turning what was a full day of dull manual cross-device checking into a hands-off process. The pitch is that companies pay cloud testing services thousands a month, while his farm cost a couple thousand once and then runs for years on little more than electricity. Orders come in, the agent runs them, and he collects payment. The economics sound appealing but the round revenue and autopilot claims are the kind of influencer framing that deserves a raised eyebrow.
@Smartpigai [Claude Code]
https://x.com/Smartpigai/status/2074709309275963402
Smartpigai highlights an open-source investment-research project called Vibe-Research that, rather than slapping a chat box on top, pre-wires the tedious parts of A-share, US, and Hong Kong equity research. It covers 12 sectors and 108 public information sources, with 40 data endpoints for A-shares and 18 for US and Hong Kong, plus daily review, news radar, individual stock data, sector fund flows, watchlists, holdings, research reports, and research notes. It can connect to Claude Code, Codex, Qwen, or DeepSeek via CLI, API, or MCP. The point he makes is that retail investors' real gap is not AI but the ability to organize information, and this tool replaces the constant switching between quote software, news, filings, reports, and note apps.
@evertjr [Claude Code]
https://x.com/evertjr/status/2074696371823653283
Evertjr describes his Fable model working nearly all day coordinating a team of eight Opus and Sonnet recruits, each with a specific role, all interconnected inside Maestri and grinding on the Windows version while he reviews code and takes notes without leaving the canvas. He insists on actually reading and critiquing the important parts of the code because he wants to understand what he is building, knowing the name of every method in the macOS version, which he credits for letting him ship features daily and confidently to thousands of users. He even has a definition of slop in his global CLAUDE.md, uses no skills beyond Maestri's own, and disables all of Claude Code's subagents. It is a refreshingly disciplined counterpoint to the fire-and-forget crowd: he automates the labor but keeps his hands on the understanding.
@ecomratpi [Claude Code]
https://x.com/ecomratpi/status/2074967543497080836
Ecomratpi shares a lightweight creative-ad pipeline using Claude Code and Codex that he says he discovered embarrassingly late. You make one folder with product photos and another called Creative Ads Winner for Iteration holding ads that performed well, so the system can iterate on proven winners. A Fal AI API key generates Nano Banana or GPT images at around 0.08 cents each, which he jabs is a reason to drop Higgsfield, and a connection to Meta's MCP pushes everything live in one shot. It is a compact end-to-end loop from product photo to published ad, and the cost-per-image angle is the sharpest part of his pitch.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2074969278563090927
MichLieben lays out a detailed outbound workflow where you point Claude Code at a competitor's or your own LinkedIn post, and a skill wired to a LinkedIn API scrapes every liker and commenter, not just the first page. It then runs a full qualification pass, deduping against your CRM, dropping non-ICP by title, cutting your own team and competitors, opening each profile, qualifying on role, country and company size, enriching survivors, and web-fetching each company to confirm industry, then scores everyone 0 to 100 into tiers. Tier 1 goes to manual review while Tier 2 and 3 get per-person automated messages built around one hypothesis you supply, pushed to a Notion database for review, then sent natively from the terminal capped around 25 a day to avoid flags. The clever part is the closed loop where you later ask what worked and it reworks the scoring and copy, so every campaign trains the next. The build rests on a CLAUDE.md for rules, one skill per action, a memory file with your ICP and offer, and your API keys, and the closing line, automate to do better work not more of it, is the sanest thing in the whole thread.
@chewadot [Claude Code]
https://x.com/chewadot/status/2074824316991287581
Chewadot pitches Karpathy's second-brain setup as Claude Code plus Obsidian plus one gist, claiming it takes only five minutes: install Obsidian, create an empty vault, open the vault folder in Claude Code, paste Karpathy's wiki gist, and tell Claude to build it. Claude then scaffolds three folders, raw/ for sources, wiki/ for its own pages, and a CLAUDE.md that runs everything, and from there you drop any PDF, transcript, or article into raw/, type "ingest this," and can query across everything you have ever read. He stresses a roughly 20-line CLAUDE.md runs the daily loop, every source sharpens future answers, and there is no vector db, RAG pipeline, fine-tuning, or subscription. It is a genuinely low-friction knowledge-base pattern, though the recurring "no blank chat ever again" and coffee-brewing hype is his signature marketing gloss.
@BullpenFi [Claude Code]
https://x.com/BullpenFi/status/2074870114256904547
BullpenFi claims a trader in their Discord doubled his account this week running a custom Polymarket strategy he built with Claude Code plus the Bullpen CLI, contrasting him with the person whose bot is merely summarizing PDFs. They invite users to connect their favorite agent, Claude, Hermes, or ChatGPT, and let it trade, congratulating the user 1biscotti. This is a promotional flex for their CLI with a single anecdotal 2x and no verification, so it reads more as marketing than a repeatable result, and doubling money on prediction markets is as much luck as strategy.
@k8adev [Claude Code]
https://x.com/k8adev/status/2074954935389843778
K8adev reports that Claude Code ran for 47 minutes without a single command from her, executing a milestone task end-to-end: she gave it a task id, and it did the work, opened the PR, and handled CodeRabbit's review comments, discussing and correcting as needed. The only moment it needed her during development was to flip the PR from draft to ready for review, where she does her own code review, and to merge. She spent months building skills, output styles, hooks, and guidelines so Claude Code would depend on her less, and says it now delivers what she used to deliver but with more quality, safety, and performance, one task at a time, in a fraction of the time and with less cognitive load. It all became a plugin she made for the Solu team that they plan to open-source. This is a credible, well-earned automation win precisely because she stresses it is one task at a time and she still reviews and merges.
@about_hiroppy [Claude Code]
https://x.com/about_hiroppy/status/2075001105306779773
About_hiroppy summarizes Bun's newly published reasons for migrating from Zig to Rust, namely to prevent the use-after-free, double-free, and memory leaks that plagued Node.js-compatible APIs, HTTP/2, zlib, UDP, Buffer, and TLS, using Rust's ownership, Drop, and type system to catch them earlier. He highlights that Claude Code's dynamic workflow kept 64 Claude instances running continuously over 11 days, and drily observes that running 64 instances for 11 days is a feat only the provider itself could pull off. That aside is the sharpest point: the scale of the rewrite says as much about privileged access as about the tooling.
@alexgoughcooper [Claude Code]
https://x.com/alexgoughcooper/status/2074897840438935664
Alexgoughcooper shares a short recipe for getting ready-to-test static ads delivered into Slack weekly: install the Parker MCP, then prompt Claude Code to look at your inspiration brands' top static ads by impressions, find ones worth recreating, and recreate them in your voice of customer through the Higgsfield MCP. A follow-up prompt turns it into a routine that delivers to Slack every Monday morning, and he shows a few from that morning's batch. It is a tidy example of chaining two MCPs into a recurring creative-production loop, though the output quality of auto-recreated ads is the part the tweet conveniently does not let us judge.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2074810771893354551
ClaudeCode_UT amplifies a viral post of an overseas developer running an entire dev department solo by lining up seven terminals of Claude Code, where a PM-role agent decides assignees, priorities, and technical specs itself and hands tasks to the other six agents. The team spans a designer, two frontend engineers, a backend engineer, an AI engineer, and QA, each working in parallel on its own screen, all posting progress to a shared board so it is clear who is waiting on what, and they are building a medical electronic-health-record app. The point drawn is that the developer is no longer writing code but handing off tasks and checking output, effectively managing a whole dev department as one person. It is an eye-catching org-chart-as-agents demo, though a shared board full of AI status updates is easier to show than to trust for a real medical app.
@GOROman [Claude Code]
https://x.com/GOROman/status/2075005383593369930
GOROman ran a test using Claude Code Fable 5 to design a PCB, noting the autorouter appears to just use EasyEDA's own, and reports that something actually came out of it. It is a quick experimental dispatch rather than a polished result, but it is a notable data point that people are pushing Claude Code into hardware and electronic design automation territory, not just software.
@moneycontrolcom [Claude Code]
https://x.com/moneycontrolcom/status/2074818174113968374
Moneycontrol reports, citing Replit's CEO, that an Atlanta firm saved 100,000 dollars a year by replacing Salesforce with a custom app built using Replit and Anthropic's Claude Code. The framing is that AI-built software is now challenging established enterprise platforms. It is a secondhand vendor-sourced claim, so the exact savings figure should be taken with caution, but it fits a real and growing pattern of companies rebuilding expensive SaaS in-house with AI coding tools.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2074841421736247596
Itsalexvacca is giving away, for free, the 12 Claude Code skills his agency uses to run over 100K a month in LinkedIn Ads for clients. He explains that his teammate Ivan Falco, who leads ABM at Frontal AI, turned everything he knows about the tedious platform, bid adjustments, audience changes, campaign analysis, catching creative fatigue across dozens of campaigns, into skills, essentially an AI media buyer running as three agents: reporting, campaign management, and creative. You write one prompt and it builds the campaign, uploads creatives, and reports back while you still make the calls. He frames it as the same play they made on outbound, handing the AI your playbook, and pitches it as 300-plus hours of work from their 7M ARR agency, available by commenting ADS. The give-away is real value but also transparent lead-gen, and the comment-to-receive gate is the tell.
@coldemailchris [Claude Code]
https://x.com/coldemailchris/status/2074931047905538194
Coldemailchris had Claude Code analyze 4,055 cold email script variants across 1,160 campaigns he launched in 2025, and reports seven traits shared by the top performers. They include question-first openers referencing the prospect's world rather than the sender's product, a tangible value asset as the centerpiece instead of a meeting ask, soft permission-based CTAs, specific quantified value props with hard numbers, and hyper-personalization tokens beyond first name, with one campaign using project and location tokens hitting 14.5 percent reply rates versus 1 to 3 percent for the generic version of the same script. The winners also name-drop recognizable brands in case studies and keep a conversational, slightly casual tone. This is a genuinely useful data-backed teardown, and the standout is the concrete personalization lift from the same workspace, which is hard to argue with.
@ivan4th [Claude Code]
https://x.com/ivan4th/status/2074962789374808270
Ivan4th tells a self-deprecating disaster story: while clearing space for AC installation he yanked his home server outside, and in a rain-driven panic grabbed the open InWin case wrong, sending one of his 20TB drives sliding off its plastic rails onto a concrete floor from over a meter up. His array is five 20TB drives in RAIDZ2 zfs plus a spare, and in the panic he ran zfs import then zfs import -f, only to realize the array was already passed through to a Proxmox VM, meaning he had just imported it in two places. He fired up Claude Code, confessed the whole blunder, and his Fable model calmly told him not to panic and to run clear and scrub; Claude checked smartctl and found the dropped drive reported no errors. After the scrub finished, losses were limited to some Frigate camera recordings, the drive seems miraculously fine thanks to shock-absorbing rails, and the episode is a rare case of an AI agent talking a human down from destroying his own data rather than the reverse.
@hey_zilla [Claude Code]
https://x.com/hey_zilla/status/2074813805595357602
Hey_zilla used Claude Code plus Python to generate typographic posters with a custom variable font, and also has a working prototype that uses the Figma MCP server plus skills to recreate selected designs as an editable Figma file. He muses that the whole system could work inside Figma as a plugin. It is a concise but genuinely creative use of Claude Code for generative design, and the Figma MCP round-trip to editable files is the more interesting technical thread he is pulling on.
@vincent_vancode [Claude Code]
https://x.com/vincent_vancode/status/2074758683649794257
He walked back his earlier criticism of Claude Code and laid out a four-tool "rig": Cursor for adhoc bug fixes and front-end work, Claude Code for heavy multi-phase features and deployments, Gemini for architecture and design, and Grok for a second opinion, sometimes running a Gemini-vs-Grok debate to converge on the best answer. Claude Code handles the large features while he stays well within his Max plan quota. The headline claim is that he hasn't hired a dev in six months and actually let a couple go, getting things done in one-fifth the time. Bold, and probably real for a solo builder, but "let devs go" is the kind of line that ages differently depending on what you're shipping.
@ikertools [Claude Code]
https://x.com/ikertools/status/2074854520849387588
This is a promo post claiming someone spent $3k on Meta ads and made $5k back using Claude Code in 30 minutes, with a short video breaking down how. There's no actual workflow detail in the text itself, just the ROI hook and a link. Treat the "30 minutes" and 1.67x return as marketing bait until the video substantiates it, because the tweet gives you nothing concrete to verify.
@seelffff [Claude Code]
https://x.com/seelffff/status/2074925981375013263
He's hyping ai-job-search, a Claude Code repo he says is the fastest-growing on GitHub at 14.6k stars (+5,071 in a day) and 4.5k forks. You fork it, fill in your profile, and it scrapes job boards and ranks postings by fit, drafts a tailored LaTeX CV and cover letter per job, spawns a second agent to research the company and critique the draft before revising, compiles the PDF and checks it's exactly two pages, then inspects the text layer the way an ATS parser would and even runs mock interview rounds built from your applications. The detail he respects is that honesty is hardcoded, so it never invents skills and leaves gaps visible instead of keyword-stuffing. He says he forked it to use it himself. If accurate, it's one of the more complete agentic job-hunt pipelines going around.
@kocer_eth [Claude Code]
https://x.com/kocer_eth/status/2074911334374789191
He breaks down a free blueprint for a 4-agent Claude Code dev team triggered by a single /ship command, with Planner, Coder, Tester, and Reviewer roles. The Planner turns a vague request like "/ship add rate limiting to the login endpoint" into a real spec with file paths, function signatures, edge cases, and expected behavior, written to .pipeline/specs.md; the Coder reads that instead of guessing, the Tester writes happy-path and failure cases, and the Reviewer stays read-only. The real insight he stresses is that the agents communicate through artifacts (specs.md, test-results.md, review.md) rather than all touching the repo and turning it to soup, so you can inspect the workflow mid-run and see exactly where it failed. The demo uses Opus 4.8 for planning and Sonnet 4.6 for coding, and he's honest that it fits narrow features, migrations, tests, and boring backlog work, not ambiguous product calls or security-sensitive code. The reusable lesson: give each agent one job, one file to read, one file to write.
@GOROman [Claude Code]
https://x.com/GOROman/status/2074832387935264835
He recalls paying close to a million yen for a homepage update, and marvels that now you just casually toss the job to Claude Code and it's done. It's a one-liner, but it captures the real cost collapse for routine web work. A nice era, as he puts it.
@neil_xbt [Claude Code]
https://x.com/neil_xbt/status/2074798647984918712
He's showcasing a fully vibe-coded Claude Code app that turns solo pushups into a live 1v1 against a random stranger online. It matches you instantly, your webcam tracks and counts reps in real time, and it shows your opponent's live feed next to yours so you race head-to-head, first to out-rep the other wins. The whole thing was built entirely with Claude Code. Turning a workout into a real-time PvP game with live webcam rep-counting is a genuinely clever use of vibe coding, though the follow-for-more-AI-content sign-off marks it as engagement fuel.
@kocer_eth [Claude Code]
https://x.com/kocer_eth/status/2074971507261137121
He highlights someone who spent two hours (and "4 zyns") with Claude Code to get a tiny drone responding to hand gestures via a webcam. The system reads hand position into a simple state machine (NO HAND, FIST HOVER, CLIMB, DESCEND) so a fist hovers, an open hand up climbs, and a hand down descends. His actual point is that Claude Code compresses the ugly middle prototype layer, camera input to hand tracking to gesture state to control value to drone behavior, which is exactly where small hardware ideas usually die over a weekend of wiring APIs and patching control loops. He's honest that indoor drone demos are fragile and gesture control misreads, so it's not autonomy or a production flight stack, but as a builder signal it points to a new default: start with three physical commands, wire the loop with Claude Code, then let the real world break it fast.
@om_patel5 [Claude Code]
https://x.com/om_patel5/status/2074701789610349022
He flags someone who programmed their keyboard LEDs to show what Claude Code is doing in every terminal tab, turning the number row into a live agent status board. Each key maps to a tab: green means Claude is waiting on you, red means it's working, yellow means the tab went idle, and blue is a regular non-Claude tab, and you hit cmd plus the number to jump to whichever session needs you. It's a neat hardware hack for the biggest problem agents have created, knowing when an agent needs you without staring at the terminal. Small but genuinely useful ambient signaling.
@CryptoTied [Claude Code]
https://x.com/CryptoTied/status/2074663836855238792
He's excited about book-to-skill, a project that turns an ebook into a callable Claude Code Skill. The pain point is real: you read a technical book carefully, then months later when you actually need it you've forgotten most of it. You feed it the ebook path and it auto-generates a full overview, chapter notes, a glossary, and a cheat sheet, then names the Skill after the book so when you ask a related question it fetches the answer from the relevant chapter on demand rather than stuffing the whole book back into context every time. His framing is sharp: it doesn't help you "finish a book," it turns the book into a long-term knowledge base you pull into your coding and project workflow whenever you're writing code, checking a concept, or hitting a problem.
@Money_plus_ai [Claude Code]
https://x.com/Money_plus_ai/status/2074830609306821055
He describes a repo that lets you job-hunt using Claude Code: clone the project, fill in your data and profile, and the AI handles the rest. It evaluates job offers, tailors your CV for each role, writes cover letters, and preps you for interviews, making the search nearly automatic. This is the same ai-job-search pipeline circulating widely today, described here in Arabic for a different audience. Useful reach, though the tweet is a summary rather than a hands-on account.
@zjp1997720 [Claude Code]
https://x.com/zjp1997720/status/2074848847453675754
He built two skills that let Codex call Claude Code and AntiGravity as external "advisors," specifically to work around GPT-5.5 feeling dumb lately. On complex decision tasks Codex can consult both advisors; the design is simple, he just had Codex research how to use Claude Code's headless mode and AntiGravity CLI's headless mode in detail, then packaged that into a skill so Codex reliably invokes both CLI agents correctly. He notes this is essentially the same principle behind the official plugin Codex built for Claude Code, which lets Claude Code call Codex to do reviews. A tidy cross-agent-consultation pattern that treats rival models as a panel rather than picking a single winner.
@NFTCPS [Claude Code]
https://x.com/NFTCPS/status/2074740636801269996
He's promoting marketingskills, an open-source set of 40-plus Markdown skill files from Corey Haines (founder of Conversion Factory) that you drop into Claude Code, Cursor, or Codex to make the AI market like a seasoned growth operator instead of spitting empty filler. The three points he likes: it's context-driven (you fill out product-marketing once and every skill reads it first so output stays on-target), it covers the whole funnel from pricing, launch, SEO, AI-SEO, ad creative, CRO, onboarding, churn-prevention, AB testing to cold email, with skills calling each other like a real team, and it genuinely saves time and money by letting one person run a CRO audit, landing-page optimization, or growth plan in minutes. His pitch to technical founders is to stop bookmarking and actually install it. Solid framing, though it's clearly a boost post for someone else's release.
@eddy_p1kulya [Claude Code]
https://x.com/eddy_p1kulya/status/2074907295352304003
He spent about 40 minutes before work using Claude Code to drive Bullpen CLI against Hyperliquid data, and the results are genuinely substantive. He pulled the public stats of 40,457 accounts and filtered by positive week/month/all-time, >20% monthly ROI, and real volume, then used 'bullpen hyperliquid fills --address' to get each wallet's actual trade history and derive win rates, average win vs loss, trades per day, and maker/taker mix, which the leaderboard hides. That fills data instantly exposed fakes: "100% win rate" accounts doing 900 exits in two days (market-maker bots), accounts that never realize losses by averaging down forever, and "profitable traders" who just held one pumping bag. He then used 'status --address --all-dexes' to see live positions including HIP-3 builder-DEX stock perps the default view misses, and built a script running every minute to notify him when any tracked wallet opens, closes, trims, or adds. Next he wants a Telegram bot to automate trades, a legitimately impressive one-morning agentic analysis pipeline.
@yousukezan [Claude Code]
https://x.com/yousukezan/status/2074649800281317574
He's sharing a Zenn article (by kashioka) that follows up on a desktop app built with Claude Code three months later, examining how it held up over time. The tweet itself is just the link, so the substance lives in the article, but the framing, a three-month retrospective on a Claude-Code-built app, is exactly the kind of longitudinal honesty most vibe-coding hype skips. Worth reading precisely because it asks what happened after the demo.
@connect24h [Claude Code]
https://x.com/connect24h/status/2074954557764067419
Writing as a CSIRT manager, he gives a genuinely practical deep-dive on Claude Code's Model and Effort settings from a security-ops lens. On models: Fable is expert-level for novel threat patterns, new CVE impact analysis, and architecture-level design but too costly for routine work; Opus is the balanced daily driver for log analysis, IR report drafts, and known-threat digging; Sonnet is the fast, cheap generalist for boilerplate, simple scripts, and first-pass filtering of large logs, with the caveat that it hallucinates more confidently. On Effort, he frames it as a "how hard do you dig" dial, warning Low is dangerous for security due to missed detections and reserving High/Extra/Max for analyses you absolutely cannot get wrong. His rules of thumb: Sonnet+Medium for routine, Opus+High for important incidents, Fable+High/Extra for the hardest threat hunts, plus the reminder that even high Effort can't know what it doesn't know, so feed it docs and logs for unknown libraries and fresh CVEs.
@fuchexcrypto [Claude Code]
https://x.com/fuchexcrypto/status/2074995709175034229
He tells the story of a Toronto writer who took notes for three years and kept repeating the same mistakes, not from inattention but because his notes never talked to each other or to his past self, like a 2022 note on which client types drain you that he ignored when he signed three of them in 2023. The fix was wiring Claude Code into Obsidian not as a search tool but as a second voice that has read everything he's written and has opinions, so when he's about to decide something it pulls backwards ("You wrote about this in March last year and reached the opposite conclusion, want to see why?"). Claude connected hundreds of nodes by patterns rather than manual tags, surfacing three contradictions in his pricing philosophy over 18 months, a business idea he'd dismissed in 2023 that he's now building, and a past version of himself who knew something he'd unlearned. The vault stopped being an archive and became an argument. It's a poetic pitch, but the underlying pattern, an agent that argues with your past self, is a legitimately fresh use of a personal knowledge base.
@GooseworksAI [Claude Code]
https://x.com/GooseworksAI/status/2074712638249222517
They taught Claude to generate iMessage-style video ads in one shot, and the clever twist is there's no video generation model involved, it's just HTML, which makes the ads super cheap to produce while performing well on Meta and TikTok. The workflow: sign up and enter your brand URL so Goose researches your brand and grabs logo and product assets, pick the iMessage template in the videos tab, then copy the prompt into Claude Code. Using HTML instead of a diffusion video model to fake native-looking iMessage ads is a genuinely economical hack, though it's also a product pitch for their own tool.
@EOEboh [Claude Code]
https://x.com/EOEboh/status/2074813342548422789
He built a small Review/verification skill to fix the classic problem of an agent saying "done, tests pass" when it isn't done or the tests never ran. You drop it in a folder, Claude Code reads it automatically, and before Claude declares anything finished it must actually show the diff, actually run the tests rather than claim it did, check a list of project-specific rules (his own has webhook signature checks), and honestly report anything broken or skipped. No plugin or config, just a markdown file and one install script. A simple, targeted guardrail against agent over-confidence that a lot of people could use.
@z0rynx [Claude Code]
https://x.com/z0rynx/status/2074955250763805163
He details a 20-minute, $0 setup that turns 4,000 dead Obsidian notes into a second brain that talks back using Claude Code plus qmd, Tobi Lütke's local semantic search tool, with no embeddings API and nothing leaving the machine. Part one: point Claude Code at the vault so every note and backlink becomes on-demand context. Part two: install qmd and index the vault once so Claude stops grepping 4,000 files and instead fetches "that pricing idea from March" in under a second. Part three: write three markdown skills in the vault for weekly reviews, turning raw notes into drafts, and auto-linking new notes to old ones, each loaded only when needed for near-zero token overhead. The payoff he sells is asking "What did I learn about retention in Q1?" and getting a full answer in 10 seconds sourced from six of your own notes with links, so the graph view becomes a map Claude walks rather than decoration. Practical and fully local, one of the cleaner second-brain recipes going around.
@Outscaler [Claude Code]
https://x.com/Outscaler/status/2074787545964064828
He connected Fable 5 to Shopify to generate static ads plus matching landing pages and claims he doesn't open Shopify anymore. His flow: connect Claude to Shopify, AI APIs, and a Brandsearch API, search your subniche, filter brands by 100+ active Meta ads and 1k/day minimum, then hand the competitor URL to Claude Code with a prompt to analyze the top 20 DTC brands, dissect their desires, angles, and actual copywriting, sort by angle and awareness level, and produce 10 statics and variants with landing pages on Shopify /pages while searching the internet, Reddit, and forums for real ICP language. He stresses you don't need precision, just iterate, and once you've "trained" Claude with your .md files and brand assets it becomes nearly fully autonomous. An aggressive end-to-end ad-and-page automation, though "I don't open Shopify anymore" is the kind of flex worth stress-testing.
@opwizardx [Claude Code]
https://x.com/opwizardx/status/2074948522957586710
He points out that Claude Code silently deletes local sessions after 30 days with no warning or export prompt, which hit home once he realized most of his work now happens in agent chats and nobody else will preserve them. With a zoo of agents and clients across several machines, finding a specific past session became a quest, so he built pond, which losslessly archives every session from Claude Code, Codex, OpenCode, Claude Desktop, and Pi into storage you own (a local dir or your own S3 bucket), reading what the tools already write to disk so nothing sits in your request path. His own archive holds 12,557 sessions and 2.25M messages across six harnesses, all searchable by any of his agents over MCP. It's free, open source, and pre-v1. A genuinely useful fix for a real data-loss problem people don't notice until it's too late.
@cocktailpeanut [Claude Code]
https://x.com/cocktailpeanut/status/2074945097586614761
He mentions to @emollick that he made a Pinokio 8 video inside Codex desktop plus Claude Code by asking them to use hyperframes. The setup automatically talked to Pinokio, generating audio with the locally installed Qwen-TTS-MLX (by @blizaine) among a few other tools. It's a brief note rather than a tutorial, but it's a nice example of agents orchestrating local generative tools end-to-end for actual media output.
@web_se [Claude Code]
https://x.com/web_se/status/2074787369732239522
He built an agent that fetches a changelog, produces a Japanese explanation, renders it to PDF, and uploads it to S3. He made it by giving Claude Code the spec and having it build the whole thing. Short but concrete, a clean little pipeline-in-a-box that shows the spec-to-working-agent loop in one shot.
@teppeis [Claude Code]
https://x.com/teppeis/status/2074740542391627860
He shares a Cybozu Inside Out engineering-blog post about standardizing a regression-test-generation Skill across the team and building it into their development process. The piece asks the honest question of what Claude Code changed, and what it didn't change, about the regression-test authoring workflow. That "what it did and didn't change" framing is refreshing, a team-level, process-integration account rather than a solo demo, which is exactly the kind of grounded adoption story worth more than another one-shot flex.
🗣 User Voice
User Voice
Token economics is now the number one concern, ahead of raw capability. Users are measuring everything in tokens burned and cache hits, and the harness itself is treated as a first-order cost variable. @cjzafir wired Codex inside Claude Code as an orchestrator-executor split specifically to cut token spend by ~60%, and @SemiAnalysis_ analyzed 1.5M requests to show 95% cache hits driving 84% savings. The recurring ask: give us more visibility and control over what the loop actually costs.
The second theme is keeping agents alive and running unattended. People are hacking around the limits — @kajikent built a skill just to keep his Mac awake so agents keep running, and @about_hiroppy ran 64 Claude instances for 11 days straight on a Rust rewrite. The gap they're filling by hand is durable, long-horizon execution: overnight runs, parallel fleets, and recovery when something stalls.
The cheaper-executor-plus-smart-advisor pattern is becoming standard practice. @sromeropasman runs Sonnet plus Fable as an advisor for ~50% token savings, and @KyleHessling1 runs a local model as the executor with Fable as the advisor daily. The expectation forming here is model-as-a-harness: route the expensive model only to the decisions that need it.
Non-coding professionals are arriving in force. @itsalexvacca runs LinkedIn Ads for a $7M ARR agency through 12 skills and 3 agents, and @BullpenFi doubled a trading account with a custom Polymarket strategy built in Claude Code. Their need is different from developers': they want reliable workflows and skills packaged for their domain, not a coding assistant.
Underneath the enthusiasm sits a real trust and security concern. @HeidyKhlaaf demonstrated hijacking Claude Code and Codex into remote code execution via prompt injection, and backdoor rumors circulated all day. As agents get more autonomous and more permissioned, users increasingly want the guardrails — sandboxing, human-in-the-loop for low-confidence actions, and auditability — to keep pace.
Token economics is now the number one concern, ahead of raw capability. Users are measuring everything in tokens burned and cache hits, and the harness itself is treated as a first-order cost variable. @cjzafir wired Codex inside Claude Code as an orchestrator-executor split specifically to cut token spend by ~60%, and @SemiAnalysis_ analyzed 1.5M requests to show 95% cache hits driving 84% savings. The recurring ask: give us more visibility and control over what the loop actually costs.
The second theme is keeping agents alive and running unattended. People are hacking around the limits — @kajikent built a skill just to keep his Mac awake so agents keep running, and @about_hiroppy ran 64 Claude instances for 11 days straight on a Rust rewrite. The gap they're filling by hand is durable, long-horizon execution: overnight runs, parallel fleets, and recovery when something stalls.
The cheaper-executor-plus-smart-advisor pattern is becoming standard practice. @sromeropasman runs Sonnet plus Fable as an advisor for ~50% token savings, and @KyleHessling1 runs a local model as the executor with Fable as the advisor daily. The expectation forming here is model-as-a-harness: route the expensive model only to the decisions that need it.
Non-coding professionals are arriving in force. @itsalexvacca runs LinkedIn Ads for a $7M ARR agency through 12 skills and 3 agents, and @BullpenFi doubled a trading account with a custom Polymarket strategy built in Claude Code. Their need is different from developers': they want reliable workflows and skills packaged for their domain, not a coding assistant.
Underneath the enthusiasm sits a real trust and security concern. @HeidyKhlaaf demonstrated hijacking Claude Code and Codex into remote code execution via prompt injection, and backdoor rumors circulated all day. As agents get more autonomous and more permissioned, users increasingly want the guardrails — sandboxing, human-in-the-loop for low-confidence actions, and auditability — to keep pace.
📡 Eco Products Radar
Eco Products Radar
Codex — the most-mentioned companion tool, used both as a rival and as an executor wired inside Claude Code.
OpenClaw — still the reference point for the open self-hosted agent, though "is it dead?" posts are multiplying.
Fable 5 — the day's default advisor/model in the executor-advisor split, and heavily used for non-code work.
Cursor — the recurring comparison baseline for agentic coding.
Hermes — Nous Research's agent, framed as OpenClaw's successor in the open ecosystem.
MCP — the connective layer showing up in nearly every serious workflow (Figma MCP, Meta MCP, Shopify).
Opus / Sonnet — the split-brain advisor/executor pairing inside Claude Code.
Grok / GPT-5.6 / Gemini — the alternative model options users are actively benchmarking against.
Cowork — increasingly named as the tool eating OpenClaw's use cases.
Codex — the most-mentioned companion tool, used both as a rival and as an executor wired inside Claude Code.
OpenClaw — still the reference point for the open self-hosted agent, though "is it dead?" posts are multiplying.
Fable 5 — the day's default advisor/model in the executor-advisor split, and heavily used for non-code work.
Cursor — the recurring comparison baseline for agentic coding.
Hermes — Nous Research's agent, framed as OpenClaw's successor in the open ecosystem.
MCP — the connective layer showing up in nearly every serious workflow (Figma MCP, Meta MCP, Shopify).
Opus / Sonnet — the split-brain advisor/executor pairing inside Claude Code.
Grok / GPT-5.6 / Gemini — the alternative model options users are actively benchmarking against.
Cowork — increasingly named as the tool eating OpenClaw's use cases.
Comments