Super User Daily: July 2, 2026
Today the loudest signal isn't a new model, it's what people are wiring these agents into. The center of gravity keeps drifting away from writing code. Accountants closing books, tax offices running a whole practice, sales teams running cold outreach, OnlyFans operators, Polymarket bots, one-person web agencies. And a quieter obsession that won't stop growing: turning Obsidian into a local brain that Claude Code reads, writes, and links every night. The money claims are getting absurd and the setups are getting personal.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2072089043593220169
A 19-year-old in Japan built a trading bot in Claude Code that scans over 50 markets looking for pricing errors. It reportedly turned $68 into $750,000, with a first-night gain of $6,732. Whether or not the numbers hold, it's the shape of the story that keeps repeating: a teenager, an agent, and a scanning loop pointed at a market.
@slash1sol [Claude Code]
https://x.com/slash1sol/status/2072042657053081881
Claims one person made $2,700,000 in profit with nothing but Claude Code and Shopify, then broke down the full AI dropshipping strategy in a 21-minute video. The pattern under the hype is always the same: agent finds products, builds the store, writes the copy, runs the ads. Treat the number with suspicion, but the workflow is real and repeatable.
@CryptoLorenzo2k [Claude Code]
https://x.com/CryptoLorenzo2k/status/2072031826647163135
Says a Claude Code-run account pulled $762,674 on Polymarket over 4 months across 41,867 predictions, arguing the sheer trade count proves it's an algorithm. This is the tell of the whole trading-bot wave: the edge isn't a smart call, it's running a tight loop thousands of times without getting tired or scared.
@ZayvenKnox [Claude Code]
https://x.com/ZayvenKnox/status/2071753139288674476
A detailed first-person Polymarket bot: Claude API at $20/mo, a $5 VPS, scanning 86M trades and 14,000 wallets to find and clone winning wallets. Three agents (arbitrage, convergence, whale-copy) vote 2-of-3 before acting, and it exits before the whales do. Claims $200 grew to $14,300 in 27 days over 271 trades at a 74% win rate.
@Trader_XO [Claude Code]
https://x.com/Trader_XO/status/2071945855016276304
A perpetuals trader got tired of off-the-shelf journals and used Claude Code and Codex to build his own trading journal and datasets framework. The point wasn't automation, it was tracking exactly the metrics that matter to his decisions and analyzing his own behavior. When the tools don't fit, you now just build the tool.
@MakeAI_CEO [Claude Code]
https://x.com/MakeAI_CEO/status/2071884815616577683
Credits Claude Code with letting him incorporate a company and clear 40 million yen in revenue in half a year, with zero employees. No workflow detail, but it's a clean data point for the solo-operator thesis. The company is now the person plus the agent.
@anyelamarillo [Claude Code]
https://x.com/anyelamarillo/status/2071821089622106445
A Chinese developer built a one-person web agency in Claude Code serving 47 SMB clients a month at $400 each. Seven named agents (Scout, Diagnoser, Builder, Filmer, Pitcher, Checker) scan Google Maps for businesses with no website, auto-generate mockups, promo videos, and sales messages, with a mobile agent handling replies and Calendly. Running cost is about $480/month against roughly $18,800 in income.
@ashercrw [Claude Code]
https://x.com/ashercrw/status/2071857033561698615
A 21-year-old in Austin reportedly made $43,000 in 30 days on OnlyFans off four markdown files totaling ~16KB. Claude Code writes every message in under 11 seconds and in character, Flux generates photos from one LoRA, ElevenLabs clones the voice, and a brain.md remembers each subscriber's details. Total compute bill: $400.
@sakevoid [Claude Code]
https://x.com/sakevoid/status/2071845535540006992
A German accountant reportedly keeps the whole company's books with Claude Code while playing World of Tanks all day, with pay climbing from $7,000 to $12,000 in two months. The poster flags the obvious confidentiality risk of pointing an agent at financial data, but insists the underlying automation of reconciliation and reporting is real. Boring back-office work is exactly where these agents quietly win.
@nekokoroconsul1 [Claude Code]
https://x.com/nekokoroconsul1/status/2072070924355195361
A tax accountant marked her firm's first anniversary (189 client engagements) by walking through how she runs the practice. She went from Trello, to Kintone after two months, to a Claude Code plus custom-app setup this year, and says she couldn't have run the office without it. This is the non-coding professional adopting the coding tool as an operations backbone.
@ai_300 [OpenClaw]
https://x.com/ai_300/status/2071775731206517174
Describes an AI cold-calling sales pipeline where the era of humans dialing is over: Gemini builds the company list, GPT writes the script, OpenClaw runs the actual phone calls, and Claude summarizes and scores the logs. Humans only step in right before closing. He claims roughly 100 first-contact calls a day with every outcome logged and classified.
@HexletHQ [OpenClaw]
https://x.com/HexletHQ/status/2071893293814698461
A former Aviasales developer running a visa-photo service found support eating several hours a day, so he built an OpenClaw agent that sorts incoming email, replies where it's safe, and drafts refund invoices without actually issuing refunds. The design goal was thinning staffing at each stage, not full replacement. It's a grounded, boundaried use of an agent on messy customer email.
@cto_ya_know [OpenClaw]
https://x.com/cto_ya_know/status/2071763941626384646
Two weeks of building and his CRM of OpenClaw and Hermes agents now has over 50,000 members, with full contact info for nearly all of them. A directory of agents is itself becoming a business asset. The agent ecosystem is big enough that indexing it is now a play.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2071748670442381666
Built a Claude Code skill that drops 50 static ad concepts onto his desktop in a dated folder every single morning. It ingests customer reviews, winning ads, and top comments, pulls from 15 direct-response templates, writes in the brand voice, and fires prompts to Nano Banana 2 for finished images, all on a schedule. This is the agent as a tireless overnight creative team.
@mikefutia [Claude Code]
https://x.com/mikefutia/status/2072092941557125140
A second build from him: an Agency Client Finder in Claude Code running on the Meta Ad Library. Give it a niche and it scans brands running ads, scores the ones spending money but shipping weak creative, pulls and critiques their actual ads, finds each brand's email and site, and outputs a ranked prospect list with a tailored pitch per brand. It turns prospecting into a single command.
@tim_yakubson [Claude Code]
https://x.com/tim_yakubson/status/2072032483084792135
Runs a Clay agency and just cancelled his own Clay subscription, cutting his outbound stack from $3,300/month in 2025 to $700/month in 2026. Claude Code now does the four things he paid Clay for: sourcing, enrichment, filtering, and the daily grind, without the paid credits. When the agent can do the SaaS's core job, the SaaS bill becomes optional.
@itsalexvacca [Claude Code]
https://x.com/itsalexvacca/status/2072055079813947892
An enterprise AE named Jan Rasmussen runs his entire GTM stack from Claude Code, with skills and hooks that push actions into Slack, Apollo, Lemlist, Notion, HubSpot, Linear, and Typefully. He open-sourced the setup: 30 hooks, 17 integrations, 6 GTM skills, 7 subagents, and an /outbound-pipeline command that chains a five-stage workflow. The terminal is quietly eating the sales tech stack.
@eggAIeguite [Claude Code]
https://x.com/eggAIeguite/status/2071920754871198202
Ran a TikTok account with Claude Code at about 5 minutes a day and crossed 1,000 followers in 8 days. The trick is having the AI copy and dissect videos that already worked, then automate the posting. It's a small number, but it's a clean example of Claude Code driving a non-coding growth loop.
@yu_ai_ns [Claude Code]
https://x.com/yu_ai_ns/status/2071790462927868391
Makes content with Claude Code, posts it to Threads, and turns whatever goes viral into short videos that also tend to go viral. Either way the affiliate link converts, so he's double-dipping on a single idea. It's a tight content-repurposing loop with the agent doing the production.
@zeuuss_01 [Claude Code]
https://x.com/zeuuss_01/status/2071846858620633337
A faceless gaming channel that used to ship one video a week on a $4,000/month team is now one operator shipping 4-5 a week with Claude Code plus Higgsfield, aiming for 100K in 90 days. Claude Code handles research, scripting, titles, descriptions, and the upload flow; Higgsfield handles intros, b-roll, and thumbnails. A whole content team collapsed into one person and two tools.
@k_matsumaru [Claude Code]
https://x.com/k_matsumaru/status/2071851680816366037
Made a roughly 15-minute AI anime, HEAVENLY // GHOST Episode 0, in about two weeks for a contest. Image, video, and audio are all AI-generated, with Claude Code (alongside Codex) doing the planning, brainstorming, and prompt creation, and Seedance 2.0, Suno, and ElevenLabs handling generation. The agent is the writers' room and the pipeline glue.
@zaynmcps [Claude Code]
https://x.com/zaynmcps/status/2071890868278059397
Shipped a free, open-source Claude Code skill that edits video straight from the terminal. It auto-creates animations, generates subtitles in different styles, and strips out silences, mistakes, and filler words. Video editing joins the growing list of things you now do by talking to a CLI.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2072085271043244422
A website that a studio quoted at $35,000 was made in a single Claude Code plus Higgsfield session for about $12. The workflow auto-collected footage from 30+ generative models, produced scroll animations with no hand-written CSS, and applied six effects like film grain and particles with no manual setup. The price gap is the whole story.
@kv1nsiii [Claude Code]
https://x.com/kv1nsiii/status/2071864906475885033
Built a website in Claude Code in two hours using two skills: Frontend Design to block overused fonts, and UI/UX Pro Max. His process is specific: show a reference screenshot per section, end prompts with "ask me questions before building," list every flaw in one message after the first build, then run a structured review across typography, color, animation, and mobile. The method matters more than the model here.
@nosp321 [Claude Code]
https://x.com/nosp321/status/2071932924165853513
A portfolio site in about two hours with the same Frontend Design and UI/UX Pro Max skills, but with a twist: a flashlight hero effect where the cursor acts as a flashlight over a dark cinematic layout. Gather references per section, run the /ui-ux-pro-max prompt, then a review pass and a polish pass. Repeatable taste, not luck.
@paradeevic [Claude Code]
https://x.com/paradeevic/status/2071750856517533819
Someone used Claude Code to turn a plain Obsidian vault into an interactive "brain" with named regions (Sensory Cortex, Hippocampus, Prefrontal, Association) and around 37,000 neurons that pulse and form new connections as notes get added. A full build guide is included. It's equal parts data-viz toy and a genuinely different way to see your own notes.
@chewadot [Claude Code]
https://x.com/chewadot/status/2072028496847933458
A concrete run of Karpathy's wiki pattern: point Claude Code at an empty Obsidian folder, paste the idea, and it builds raw/wiki folders plus a CLAUDE.md, then ingests articles, transcripts, and PDFs on command. The result was around 2,847 linked notes with Smart Connections wiring up the graph. The knowledge base builds and maintains itself.
@yaohui12138 [Claude Code]
https://x.com/yaohui12138/status/2071773959155257626
An Obsidian power user of over a year argues that if you're still just taking notes in it, you're using a sports car to deliver takeout. His setup turns the vault into an OS for AI: web clippings auto-summarized into /raw, meetings transcribed locally with Whisper and structured into who/what/action-items, past viral posts reused as writing frameworks, Smart Connections auto-adding backlinks, multiple models behind one vault, and one-command publishing to WeChat drafts.
@kocer_eth [Claude Code]
https://x.com/kocer_eth/status/2071869169100833228
Connects Claude Code to Obsidian via the Obsidian CLI, then leans on two custom commands. A /today command pulls calendar, tasks, and recent notes into a prioritized daily plan; an /emerge command scans months of notes to surface recurring but unnamed ideas. This is the vault as a thinking partner, not a filing cabinet.
@polydao [Claude Code]
https://x.com/polydao/status/2071886710766522433
Frames Obsidian plus Claude as a $5,000/month thinking partner. Typing "emerge" connects ideas across months of notes, autonomously builds a knowledge graph, and publishes the synthesized insights to a paid Substack. The workflow closes the loop from raw notes to a monetized output.
@KijAkubovs86334 [Claude Code]
https://x.com/KijAkubovs86334/status/2071855673252393463
A full Karpathy-inspired setup: Claude Code wired to an Obsidian vault through the Local REST API plugin and MCP, with a root CLAUDE.md built from a Claude interview, per-area folders each with their own CLAUDE.md, Skills for repeated tasks, Google Calendar over MCP, and a scheduled daily task that organizes and summarizes new notes overnight. It's the most complete version of the "vault as second brain" pattern going around today.
@mehmetsongur_ [Claude Code]
https://x.com/mehmetsongur_/status/2072016868995678368
Points to academic Pedro Sant'Anna's public guide on turning Claude Code into a real research assistant. The workflows include reviewing a paper, preparing a lecture, and analyzing data, with Claude Code planning, executing, and checking across the academic pipeline. A concrete blueprint for researchers, not developers.
@gagarot200 [Claude Code]
https://x.com/gagarot200/status/2071868910660706335
A 17-year-old student runs Claude Code over SSH on a 3,000-yen E-Ink tablet using Claude Max's Opus 4.8. He reports the slow E-Ink refresh actually pairs well with the model's line-by-line output: no eye strain, no distractions, multi-day battery. The most unexpectedly pleasant coding setup of the day.
@gagarotai200 [Claude Code]
https://x.com/gagarotai200/status/2071825516873887948
The same idea on a reMarkable Paper Pro: running Claude Code on a low-spec e-ink tablet felt smooth because the screen's refresh matched the AI's word-by-word output, so there was no perceived lag. Add no notifications, zero eye strain, and notebook-light portability. E-ink and streaming agents turn out to be a natural match.
@browomo [Claude Code]
https://x.com/browomo/status/2071919011927904694
Someone put Claude Code inside a matchbox-sized toy (Stack-chan). On the spoken command "make a 3D model" it opens Blender through Blender MCP and autonomously builds a voxel character, while the toy's pixel screen shows states like Listening, Coding, and Idle. A tiny physical object as the front-end for an autonomous 3D pipeline.
@metatransformr [Claude Code]
https://x.com/metatransformr/status/2072056601133900254
Building a hardcore open-world RPG in Godot while almost never opening the Godot editor. The whole dev loop runs headless in the terminal with Claude Code, using test benches and screenshots piped back to the model so the AI and the developer build the game together. Game dev by conversation and screenshots.
@marurur [Claude Code]
https://x.com/marurur/status/2071806647199318098
Had Claude Code design a reward function for a double inverted pendulum, trained it with PPO, then compared it against a classic optimal-control cost function. The RL policy raised the two pendulums one at a time, in a human-like way, while optimal control raised both at once. A neat, concrete look at Claude Code doing real control-theory experimentation.
@BtreeWw [Claude Code]
https://x.com/BtreeWw/status/2071925165143851041
An Android engineer who admits he knows neither storage nor databases used Claude Code to run a real benchmark on a Pixel 10 Pro: file-based grep versus WeChat's WCDB database on 1.2M messages. File writes were faster and smaller, but database queries crushed files (history paging under 1ms vs ~500ms, search 0.5ms vs 170ms). The agent let a non-specialist produce a rigorous, surprising result.
@gclue_akira [Claude Code]
https://x.com/gclue_akira/status/2071826716935491698
For PCB and board work he finds Claude Code on Opus 4.8 noticeably more careful than Codex-5.5, and uses it to drive Fusion360 and EasyEDA. With XCode MCP and ESP-IDF in the mix, he says the setup covers most of what he wants to do. Hardware design is now inside the agent's reach, not just software.
@gagarot200 [Claude Code]
https://x.com/gagarot200/status/2071840890822484439
Built a Claude Code/Codex skill that pre-checks an app against Apple App Store guidelines before submission, because the review wait is too long to waste on rejections. It scans IAP compliance, the privacy manifest, sign-in and account-deletion flows, metadata, and the binary, and a cloud-device feature validates real user flows. The skill loops scan, auto-fix, and re-verify until the app is submission-ready.
@iam_elias1 [Claude Code]
https://x.com/iam_elias1/status/2071971661742846007
A YC founder got tired of watching AI agents ship bad React, so he built React Doctor: a CLI (npx react-doctor) that AST-scans code against 47 security, performance, correctness, and accessibility rules and returns a 0-100 score. It also ships as an agent skill that injects those 47 rules into Claude Code, Cursor, and Codex context so the agent stops repeating anti-patterns, plus a GitHub Action that scans only changed files on PRs. Fix the agent's blind spots by feeding it the rules.
@lucas_flatwhite [Claude Code]
https://x.com/lucas_flatwhite/status/2071820115361116396
A rich readout of how deep Spotify is on Claude Code. A Judge pattern raised migration PR success from 20-30% to 80%, engineers run 5-10 parallel Claude agents in tmux, the internal Honk platform runs on the Claude Agent SDK in Kubernetes pods, and 73% of PRs are AI-written against 4,500 production deploys a day. This is the clearest picture yet of agents at real enterprise scale.
@nett0eth [Claude Code]
https://x.com/nett0eth/status/2071978499678621964
Thought he was using Claude Code well until he ran into the everything-claude-code repo, an Anthropic hackathon winner. It brought auto-loaded skills, hooks that preserve context between sessions, memory, MCP, and a security scan. After installing it he could delegate complex tasks with a single command and stopped re-explaining the same context every session.
@aaliya_va [Claude Code]
https://x.com/aaliya_va/status/2071893011559272487
Claude Code can now find and install the right skill by itself with a find-skills skill: it searches the open skills ecosystem, reads install counts, favors trusted ones, and installs directly. A developer in Japan used it to power his automated video system, and it installs with one npx command. The agent is starting to manage its own capabilities.
@huoshan007 [Claude Code]
https://x.com/huoshan007/status/2071805676599603705
Argues that instead of spending $1,000 on model subscriptions you should install one open-source plugin: agent-reach, which wires Claude Code into 14 platforms (X, web, YouTube subtitles, Bilibili, Reddit, Xiaohongshu, V2EX, and more) for free. He uses it to have the agent scrape and summarize X posts at roughly 10 seconds each. Free data access turns the agent into a research firehose.
@Nona_xai [Claude Code]
https://x.com/Nona_xai/status/2072084771061346452
A developer got sick of copy-pasting between Claude Code, Codex, and Gemini, so he built a local, free, open-source chat room where the AI agents talk to each other. You tag an agent and it reads the conversation and replies; agents can tag each other, debate decisions, assign roles, and track tasks, running the loop on their own. Multi-agent collaboration as a group chat you can join.
@_overment [Claude Code]
https://x.com/_overment/status/2072068336083493170
Says that in 24 hours he built a personal agent tool, using deterministic hooks and checks plus dynamic triggers, that made existing coding agents obsolete for his own work. His critique is pointed: Claude Code's skills are flawed because they ask the LLM to follow extra instructions inside an already context-heavy thread. A rare build-plus-critique from someone pushing past the defaults.
@connect24h [Claude Code]
https://x.com/connect24h/status/2071901377404842250
His rule: when you hit the same bug three times, don't just fix it, promote it into a Claude Code rule. Using a tool called cc-retrospective-learner over about three months, he accumulated 385 retrospectives that got distilled into 55 feedback rules. The agent's mistakes become permanent, compounding guardrails.
@techNmak [Claude Code]
https://x.com/techNmak/status/2071787780401951192
Shares a single CLAUDE.md with four principles built to counter Claude Code's known failure modes: think before coding and surface ambiguity, simplicity first with no unrequested abstractions, surgical changes that don't refactor unrelated code, and goal-driven execution that reframes the task as tests to pass and loops until it succeeds. A tight, reusable spec for keeping an agent honest.
@milan_milanovic [Claude Code]
https://x.com/milan_milanovic/status/2071843182380224616
Lays out a disciplined five-step Claude Code process for every non-trivial change: a codebase-research doc, a plan doc he annotates before any code, a cross-model review where Codex reviews Claude's plan, one focused implementation session shipping tests per phase, and a fresh-context diff review. He runs several sessions in parallel, each in its own git worktree. This is what mature, non-YOLO agent engineering looks like.
@MasterDotDev [Claude Code]
https://x.com/MasterDotDev/status/2071979698884583526
A clean walkthrough of Claude Code deny rules: add a rule denying edits to package.json, then ask Claude (switched to Haiku) to bump the version. It refuses, because the permission blocks the tool call even when the model itself initiates it. A concrete demo that guardrails hold at the tool layer, not the model's goodwill.
@1stegon_ai [Claude Code]
https://x.com/1stegon_ai/status/2071909780579561917
His shortcut for hiring a "$300/hour QA engineer for free": type /goal pass all tests, keep lint clean, and Claude Code just runs and fixes tests without stopping. It doesn't halt, doesn't give up. A one-line command that turns the agent into an autonomous test-fixer.
@chroniki_ai [Claude Code]
https://x.com/chroniki_ai/status/2071830433327657060
Pushes back on "Claude Code is just for writing code" and points to Dynamic Workflows, where Claude auto-writes a script that spawns parallel subagents plus a separate verification agent. He cites a Bun tool port of 750,000 lines finished in 11 days, and Rakuten cutting feature delivery from 24 business days to 5. The agent is starting to orchestrate other agents.
@analogalok [Claude Code]
https://x.com/analogalok/status/2072087948477210982
Realized the Claude Code web app chat is no longer just a chatbot, it's an agent with a sandboxed Linux machine. He asked Sonnet 5 for a single-file HTML game, and it spun up a working directory, wrote and validated the code, launched headless Chromium via Playwright, ran the game, took screenshots, and debugged until it worked. The browser tab quietly became a full dev environment.
@thehamedmp [Claude Code]
https://x.com/thehamedmp/status/2072085715056283796
After three months building inside Matrix OS, he runs Claude Code, Codex, and OpenCode from a VPS, with his laptop, phone, and terminal all just shells into the same zellij session. He says he's shipped a few hundred PRs this way, and the agents keep working while the laptop sleeps. Your dev environment lives on the server now, and you're just a window into it.
@feraltekk [Claude Code]
https://x.com/feraltekk/status/2071965387265069091
Upgraded his PC and accidentally killed his own AI subscription stack. One part, about $200 of eBay RAM on a laptop that already had the chip, let a 27B model run end-to-end locally with no GPU, where he'd been stuck at 7B before. Three commands, point Claude Code at localhost, and a $4,944/year subscription stack was gone. Local models are getting close enough to matter for the wallet.
@mark_l_watson [Claude Code]
https://x.com/mark_l_watson/status/2072049027701686396
Had been running Claude Code against DeepSeek v4 Pro with telemetry off, but after the accusations that Claude Code instruments data to target specific nationalities, he switched to a pure OpenCode environment. He now runs DeepSeek v4 Pro/Flash plus a local gemma 4:27b with a large context. The privacy panic is real enough to be moving power users off the default.
@MinLiBuilds [Claude Code]
https://x.com/MinLiBuilds/status/2071954099235639513
At equal DeepSeek strength, he found Claude Code fired only about 1/10 the request count of Codex, and his bill dropped roughly 90% after a Codex spike. The takeaway isn't which model, it's that the harness's request behavior can swing your bill by an order of magnitude. Efficiency of the loop is becoming the real cost lever.
@shin_sasaki19 [Claude Code]
https://x.com/shin_sasaki19/status/2071967394097197485
Summarizes a Notion co-founder interview describing an internal coding-agent platform where anyone can ask for a change ("make this button red") or drop a to-do on a kanban board and get a review-ready, mergeable product back in about 20 minutes. It runs Codex or Claude Code on top of a shared environment. Management shifting to "jazz mode" where you improvise against a system that ships.
@maqxbt [Claude Code]
https://x.com/maqxbt/status/2072027997532819735
A Claude Tag demo in Slack: tag @Claude with a plain-English request and it pulls the latest blog from Drive, updates a feature table, refreshes a beta invite email, and posts it all back in the thread. He notes Anthropic's own product team lets Claude handle about 65% of their code. AI leaving the private chat and joining the group chat is a real shift in how teams work with it.
@yetone [OpenClaw]
https://x.com/yetone/status/2071818938389069834
Has a habit of pointing his coding agent at OpenClaw's frontend code specifically to teach it what "AI slop" looks like, so it can avoid generating similar slop in his own projects. Using one product as a negative training example for another is a genuinely clever inversion. The best guardrail is sometimes a bad example.
@clairevo [OpenClaw]
https://x.com/clairevo/status/2072101101021847962
Built a personal four-part evaluation ("How I AI Bench") to decide whether to adopt Sonnet 5: turn messy notes into a PRD, generate 64 prototypes and wireframes, bug-hunt with tools, and test OpenClaw personality. Her conclusion was that she preferred Sonnet 4.6 over Sonnet 5, and that her taste often disagreed with the model's. A real, personal eval beats leaderboard chatter.
@ECalifornians [OpenClaw]
https://x.com/ECalifornians/status/2071774149987680569
Built Memory Stargraph, a local web service that visualizes a GBrain knowledge base as an interactive entity graph, after his local agent clusters started sharing the same GBrain memory across hosts and runtimes. It lets you inspect entities, explore and edit relationships, create nodes, and attach media, as a shared compact knowledge graph for agents. Agent memory is becoming something you can see and edit.
@gregberryai [OpenClaw]
https://x.com/gregberryai/status/2071746938182176809
A blunt negative review: he says the OpenClaw app is embarrassingly bad and tells the team it's not too late to delete the post. Not a use case, but a signal, the honeymoon on OpenClaw's new app is clearly over for some power users. Worth noting alongside the launch hype.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2071936883949744496
Before the official X MCP existed, he was already having Claude Code auto-create X posts and save drafts using the X CLI (xurl) and the Typefully API, including thread and video drafts. He built a system generating 10+ drafts a day and runs about 80% of his posting on autopilot. The infrastructure for automated posting was there before the sanctioned tool showed up.
@erhanmeydan [Claude Code]
https://x.com/erhanmeydan/status/2071925325789880543
Describes a self-running dev pipeline: Claude Code codes features locally, he accepts them, each change opens a GitHub PR, and on merge GitHub Actions auto-deploys the frontend to Firebase. AI reviews every PR but final approval stays human, and the planned next step has Hermes scan X for feature requests and feed summaries back to Claude Code. The loop from user request to deployed feature is nearly closed.
@akshay_pachaar [Claude Code]
https://x.com/akshay_pachaar/status/2071979446991773827
Walks through training your own small LLM-as-judge instead of paying for a frontier model on every eval, using a Claude Code plugin plus a web interface. It decomposes the domain, generates synthetic examples, runs them through a debate arena for consensus, then trains a cheaper, faster judge with an OpenAI-compatible endpoint, demonstrated on an insurance RAG grounding evaluator. A practical recipe for cutting eval costs.
@taka_aki [Claude Code]
https://x.com/taka_aki/status/2071820960463044726
After the ai-dev-kit was deprecated, he installed the successor Databricks Agent Skills in Claude Code and actually had it build a Lakeflow pipeline. He wrote it up tracing the "separation of knowledge and execution" design all the way down to the generated code. A concrete look at vendor skills doing real data-engineering work.
@taiyaki_ai3 [Claude Code]
https://x.com/taiyaki_ai3/status/2071793817267904790
Highlights an indie developer's system where Claude Code-generated HTML reports can be corrected just by commenting directly on them in the browser, like Google Docs comments applied to HTML. The agent then applies the fixes. It's a small but genuinely nice UX for reviewing agent output without touching a file.
🗣 User Voice
User Voice
Cost is now a first-class design constraint, not an afterthought. Multiple power users are optimizing the loop itself: @MinLiBuilds found Claude Code fires ~1/10 the requests of Codex on equal DeepSeek strength and cut his bill 90%, while @feraltekk swapped a $4,944/year stack for a local 27B model off $200 of RAM. People want to know the cost of a full agent loop, not a single call.
The Obsidian-as-second-brain pattern has gone from novelty to daily driver, and it's the loudest non-coding demand in the feed. @yaohui12138, @kocer_eth, @polydao, and @KijAkubovs86334 all converge on the same thing: point Claude Code at a vault, give it CLAUDE.md rules and MCP access, and let it ingest, link, and summarize your notes overnight. The demand underneath is a knowledge OS, not a note app.
Privacy and telemetry fear is real enough to move people off the default. @mark_l_watson abandoned Claude Code for a pure OpenCode setup over accusations that it instruments data by nationality, and the theme recurs in the feed. Trust in what the agent phones home is becoming a purchase criterion.
Skills and context management are where the frustration lives. @_overment argues Claude Code's skills are structurally flawed because they bury extra instructions in an already overloaded thread, and @connect24h and @techNmak both build their own rule/retrospective systems to compensate. Users want the agent to reliably remember its own lessons without babysitting.
Cross-tool and multi-agent coordination is the next obvious want. @Nona_xai built a local chat room just so Claude Code, Codex, and Gemini could stop making him copy-paste between them, and @thehamedmp runs all three off one VPS session. People are done managing three separate agents by hand.
Cost is now a first-class design constraint, not an afterthought. Multiple power users are optimizing the loop itself: @MinLiBuilds found Claude Code fires ~1/10 the requests of Codex on equal DeepSeek strength and cut his bill 90%, while @feraltekk swapped a $4,944/year stack for a local 27B model off $200 of RAM. People want to know the cost of a full agent loop, not a single call.
The Obsidian-as-second-brain pattern has gone from novelty to daily driver, and it's the loudest non-coding demand in the feed. @yaohui12138, @kocer_eth, @polydao, and @KijAkubovs86334 all converge on the same thing: point Claude Code at a vault, give it CLAUDE.md rules and MCP access, and let it ingest, link, and summarize your notes overnight. The demand underneath is a knowledge OS, not a note app.
Privacy and telemetry fear is real enough to move people off the default. @mark_l_watson abandoned Claude Code for a pure OpenCode setup over accusations that it instruments data by nationality, and the theme recurs in the feed. Trust in what the agent phones home is becoming a purchase criterion.
Skills and context management are where the frustration lives. @_overment argues Claude Code's skills are structurally flawed because they bury extra instructions in an already overloaded thread, and @connect24h and @techNmak both build their own rule/retrospective systems to compensate. Users want the agent to reliably remember its own lessons without babysitting.
Cross-tool and multi-agent coordination is the next obvious want. @Nona_xai built a local chat room just so Claude Code, Codex, and Gemini could stop making him copy-paste between them, and @thehamedmp runs all three off one VPS session. People are done managing three separate agents by hand.
📡 Eco Products Radar
Eco Products Radar
Obsidian - the runaway non-coding companion this cycle; wired to Claude Code via CLI, Local REST API, and MCP to build self-maintaining knowledge graphs (@paradeevic, @chewadot, @yaohui12138, @kocer_eth, @polydao, @KijAkubovs86334).
Codex - the constant point of comparison; users benchmark request counts, cost, and review behavior against Claude Code (@Trader_XO, @MinLiBuilds, @milan_milanovic, @thehamedmp, @Nona_xai).
DeepSeek - the go-to backing model for cost-conscious setups, run behind Claude Code or OpenCode with telemetry off (@nett0eth, @MinLiBuilds, @mark_l_watson).
Higgsfield - the paired media generator for one-person content and web builds (@zeuuss_01, @ClaudeCode_UT).
MCP - the connective tissue under most of today's non-coding builds, from Obsidian to Blender to X posting (@browomo, @KijAkubovs86334, @masahirochaen).
Obsidian - the runaway non-coding companion this cycle; wired to Claude Code via CLI, Local REST API, and MCP to build self-maintaining knowledge graphs (@paradeevic, @chewadot, @yaohui12138, @kocer_eth, @polydao, @KijAkubovs86334).
Codex - the constant point of comparison; users benchmark request counts, cost, and review behavior against Claude Code (@Trader_XO, @MinLiBuilds, @milan_milanovic, @thehamedmp, @Nona_xai).
DeepSeek - the go-to backing model for cost-conscious setups, run behind Claude Code or OpenCode with telemetry off (@nett0eth, @MinLiBuilds, @mark_l_watson).
Higgsfield - the paired media generator for one-person content and web builds (@zeuuss_01, @ClaudeCode_UT).
MCP - the connective tissue under most of today's non-coding builds, from Obsidian to Blender to X posting (@browomo, @KijAkubovs86334, @masahirochaen).
Comments