Super User Daily: 2026-06-16
The day Fable 5 got pulled, the feed didn't mourn so much as get to work. Heavy, deliberate usage everywhere: 22.7 billion tokens across one person's 30-day remote-control run, a thousand-model Blender job left unattended for ten hours, hundred-hour data jobs handed to Dynamic Workflows. But the cases that actually traveled had nothing to do with writing code: a $3 ad-strategy pipeline, a virtual influencer running on four files, a paid-media OS replacing a $3,500/month agency, a teenager's $300k Starlink-positioning gadget, and an agent negotiating its own refund through a browser. The other through-line is loop and orchestration: people wiring Opus and Codex together as peers, letting Hermes drive both while they play a drum game, and building local-model setups with persistent memory so nothing leaves the box.
@masahirochaen [Claude Code]
https://x.com/masahirochaen/status/2066042326498316612
He ran the numbers on 30 days of remote-controlled Claude Code and the figures are absurd: 360 sessions, 80,000 AI exchanges, 380,000 lines of generated code, 22.7 billion tokens processed, and 122 runs between midnight and 6am. His workflow is to throw tasks before sleep and wake to written code, then nudge things forward from the gym or the train with one-line replies on his phone. He frames the 380k lines as roughly 1,267 person-days, about five years of one engineer's work. The point isn't the code count, it's that time in front of a screen is no longer the same thing as working.
@shion_takk [Claude Code]
https://x.com/shion_takk/status/2066172570156822613
A clean example of letting tokens run: he set Claude Code's Dynamic Workflows on a "1000 geometric models in Blender" batch job and let it run all day. He did the first 200 with Opus 4.8 to validate the flow, then handed the remaining 800 to the agent fully unattended, about 10 hours of runtime including waiting out a 5-hour usage cap. He's blunt that it burns a lot of credits, but when you just need volume and the quality bar is low, it's hugely capable, with Opus 4.8 acting as the orchestrator.
@shion_takk [Claude Code]
https://x.com/shion_takk/status/2065981824183185613
Reports that even after Fable 5 vanished, Claude Code's Dynamic Workflow firepower carries it. From day one of the feature he handed it a very heavy 100-hour data-analysis-and-cleanup job and it completed the whole thing, with some quality defects he says Fable later one-shotted. He gives the exact recipe: pick ultracode via /effort, trigger with /workflow, and you may need to flip Dynamic Workflows on with an env var. A concrete, replicable heavy-data run.
@RoundtableSpace [Claude Code]
https://x.com/RoundtableSpace/status/2066003947761123692
A creative-strategist pipeline that lives entirely inside Claude Code and runs on about $3 of API cost. Apify pulls every active competitor ad from the Meta Ad Library, Gemini watches each video and maps the patterns that repeat across three or more ads, and Claude turns those patterns into 10 ready-to-brief ad concepts in your brand voice. What used to be a media buyer with 40 tabs open and a Google Doc of screenshots becomes a 15-minute pipeline. Concrete non-coding marketing value built end to end.
@ridark_eth [Claude Code]
https://x.com/ridark_eth/status/2066216714644111431
A 19-year-old in China makes $9,000/month designing product sites, shipping each in an afternoon, and the post gives his exact setup. Hand-written brief (5 min), Moonchild builds the design system and every screen from it (20 min), MCP hands the design to Claude as real structure rather than a screenshot (instant), Claude Code reads those exact tokens and builds the live app (20 min), a second Claude session reviews for drift (10 min), about an hour total. The trick is MCP passing actual colors, components and layout so Claude builds from source, and screen five still matches screen one.
@nasicaonchain [Claude Code]
https://x.com/nasicaonchain/status/2066095464605225222
Breaks down how a virtual influencer pulling $10,000/month in brand deals (Victoria's Secret, Razer) runs on just four files. persona.md is her backstory, personality and the topics she never touches; voice.md clones a real voice actress (90 seconds of Fiverr audio, ~$40, cloned in ElevenLabs); flux.md holds 68 physical descriptors plus a fine-tuned LoRA on a rented GPU; brain.md is her memory of brands, followers and conversations. Claude Code reads all four before every interaction so she never breaks character, with a cron job checking her messages. A sharp non-coding playbook.
@anyelamarillo [Claude Code]
https://x.com/anyelamarillo/status/2066185381528179173
A 16-year-old built a portable Starlink-beacon positioning device, a free GPS backup, and the post says he made $300,000 from it. He isn't stealing Starlink's network; he reads the radio beacon signals the satellites constantly emit, identifies satellites and computes position from them, working even where GPS fails. He solved the whole thing with Claude, which wrote the program that captures the beacons and does the positioning math. A striking hardware-plus-AI case where the code was the hard part and Claude carried it.
@itsharmanjot [Claude Code]
https://x.com/itsharmanjot/status/2066053005506695670
He replaced a $3,500/month paid-media agency with a single Claude Code plugin running as an autonomous, self-healing "paid media OS." Last week it took $18k of test spend, flagged 3 money-losers by day two, reallocated $9k, and lifted ROAS from 2.1 to 4.8 in 11 days, then auto-tuned the next cycle to lock the gains. It catches things an agency won't tell you: which creatives fatigue two weeks before the drop, which campaigns cannibalize each other, which ad sets burn money on long-dead audiences. A concrete agency-replacement with hard numbers.
@DfxoqEth [Claude Code]
https://x.com/DfxoqEth/status/2066244097078182275
A consultant in Lyon stopped doing client calls and now makes $13,630/month off his Obsidian vault: 2,400 notes on B2B SaaS pricing strategy built over eight months, every framework and objection he ever heard on a sales call. Instead of selling his time he licenses the graph, $290/month per company, 47 companies subscribed. The graph queries Claude on their behalf, so they get the answer his hourly clients used to pay $400 for, and he runs the whole thing solo from one terminal. A clean "turn your knowledge into a product" play.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2065952150497411393
A pure non-coding agentic story: he got a full refund on a Claude Max plan two days after subscribing, and handed the whole negotiation to Claude Code. He'd upgraded to Max for Fable 5, which was suspended for all users the next morning under US export controls, so he figured a spec change right after signup was a refund case. Instead of doing it himself he connected a browser-operation MCP to Claude Code (Opus 4.8) driving his already-logged-in Chrome, opened the support chat and ran the refund flow to completion while he just approved the messages. $220 fully refunded, status flipped to Refunded; an AI applied and an AI approved.
@gregbarbosa [Claude Code]
https://x.com/gregbarbosa/status/2065988953623199772
About as low-friction as a personal build gets: a $13 AliExpress smart ring plus Claude Code was all he needed to make his own health app, and it syncs to Apple Health too. No platform, no SDK wrangling described, just a cheap sensor and an agent. A nice reminder that a lot of "I wish an app did X" is now an afternoon and $13.
@yousukezan [Claude Code]
https://x.com/yousukezan/status/2066165799270928654
A clean study-domain result: a self-described humanities-background (non-CS) engineer built his own learning app with Claude Code and used it to pass the CISSP security certification in three months. The output is a self-built study tool; the result is a genuinely hard cert. Education and self-built tooling, not coding for its own sake.
@kensuu [Claude Code]
https://x.com/kensuu/status/2065952586021388472
A small, very human use case: he wanted a shower timer because in the bath it's hard to track how many minutes you've spent wetting or rinsing your hair, so he had Claude Code throw one together loosely with no detailed instructions. He adds a funny aside, that building such apps without instructions, both Claude and ChatGPT tend to talk down to you. The everyday "I just wanted this to exist" build.
@andreysuperior [Claude Code]
https://x.com/andreysuperior/status/2066142457323045342
He stopped trading and started making more by building the bot instead of running it. He used to sit at the screen six hours a day; now five positions are open and green while he stands in the living room with his phone. He didn't buy the bot, he described what he wanted in plain English and Claude Code wrote the code: $8,000/month on a $50,000 account, reading two messages a day, morning and evening, four minutes total. The classic "build the system, not the trade" framing.
@gagarotai200 [Claude Code]
https://x.com/gagarotai200/status/2066013774398820717
A detailed, beginner-aimed walkthrough of building an AI trading bot with Claude Code, where the point is explicitly to have AI build the trading system rather than predict the market. The stack is serious: HMM-based market regime detection (Crash/Bear/Neutral/Bull/Euphoria), regime-based capital allocation, walk-forward backtesting, risk management, an Alpaca broker connection, automated order placement, a P&L dashboard, and the main loop. He's careful not to claim profits, just that one person can now build the whole research-to-execution stack.
@kirubaakaran [Claude Code]
https://x.com/kirubaakaran/status/2066155848251187363
He used to think a full-scale algo trading platform simply couldn't be built from scratch with AI; he only made small trading-analysis tools. Then he built an entire at-scale platform with Claude Code, thousands of lines of code, the UI/UX, the database structure, all done by the AI. He reads it as proof the barrier has collapsed and a wave of solopreneurs is coming. A concrete end-to-end product from someone who didn't believe it was possible.
@0xrevayz [Claude Code]
https://x.com/0xrevayz/status/2066232320059642158
A cost story with a memory twist: after a $170 Claude bill in 10 days, he bought a Mac Mini M4 ($599 once), installed Ollama, and pointed Claude Code at localhost, no API cost, no data leaving the box, ~10 watts under his desk, for a projected $5,232 saved year one. But local AI is useless without memory, so he wired it to an Obsidian vault through MCP and built offline workflows that remember everything: a Morning Brief, Meeting Cleanup, Research Intake, Weekly Review. Local model plus persistent memory, fully offline.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2066128576005743020
A five-step Figma-plus-Claude-Code workflow for converting a competitor's landing page into your own brand, now at production quality. Capture the competitor LP via the Figma Chrome extension as editable layers (not a screenshot), have Claude Code read the whole structure through Figma MCP, rebuild it into your brand (colors, fonts, copy style, even a banned-word list), generate 16 on-brand product images with GPT Image 2, and let Claude auto-place them and export the page. About $2 of API cost, 15 minutes start to finish, because MCP reading the structure killed the screenshot-and-guess loop.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2066022824591143400
A practical technique for not slamming into Fable 5 usage limits in Claude Code: set the model to Fable 5, crank reasoning to Max, and build a dynamic workflow where Fable acts as orchestrator and only hands the heavy reasoning phases to Opus. His insight is that you hit limits because you make Fable do everything; splitting the roles lets you run much deeper on the same budget. A concrete, replicable budgeting move.
@shupeiman [Claude Code]
https://x.com/shupeiman/status/2066093768164679789
A tight marketing result: he easily built a site with Claude Code and, via an AI-plus-marketing approach, got more than 5,000 newsletter signups off one video. He adds that he didn't outsource the video to staff either, he handed that to Claude Code too. Concrete top-of-funnel numbers tied directly to the build.
@akiyoshisan [Claude Code]
https://x.com/akiyoshisan/status/2065980199771242861
An almost fully automatic AI-video pipeline run from Claude Code with a detailed tool chain. Dreamina (with OmniHuman 1.5) generates the avatar source, MiniMax TTS narrates from a clone of his own voice, Whisper times the subtitles, HyperFrames remove-background makes the 2D character transparent, the HeyGen CLI on WSL searches and downloads BGM and SFX, and HyperFrames assembles and exports the MP4. A concrete multi-tool, non-coding content workflow, all orchestrated by the agent.
@php_martin [Claude Code]
https://x.com/php_martin/status/2066083471697879542
A small but slick design detail: in his website builds, Claude Code automatically calls an image-generation service to produce three style-consistent images for a page. The images aren't pulled from a stock library, they're generated to match the page content, with consistent style and adapted sizes, all automated from description to download in a 33-second clip. A concrete look at the agent owning the asset step, not just the code.
@teach_fireworks [Claude Code]
https://x.com/teach_fireworks/status/2066001154388201845
He built an open-source Claude Design alternative, fireworks-design, on Dynamic Workflow mode plus "Loop Engineering," using GLM 5.2 as the main coding model. His argument: one generation is drawing a single card from a distribution, so quality is capped by variance; instead he deliberately samples many. The pipeline fans out 8 different aesthetic directions, scores them on 6 dimensions with an LLM-judge panel, makes the winner the skeleton while grafting the best bits of the rest, then runs a critique-and-fix loop until it passes. It welds together best-of-N, self-consistency, self-refine and schema-validated structured output as a real Claude Code dynamic workflow.
@marouane53 [Claude Code]
https://x.com/marouane53/status/2066267503400276445
He noticed OpenRouter's Fusion launch is basically the loop he'd been running by hand, now turned into an API: send one prompt to a panel of models, a judge looks at where they agree, contradict, and what each missed, and a synthesizer writes the final answer from that. They claim Fable-level results at half the price, and the detail he keeps returning to is that about three-quarters of the gain comes from the synthesis step, not from the models thinking differently. A sharp read on why multi-model panels work.
@PawelHuryn [Claude Code]
https://x.com/PawelHuryn/status/2066052757895749891
A concrete recipe for making Opus and Codex work as peers without Fable or OpenRouter Fusion. Create a Codex ideation skill where Claude Code uses Codex as a peer, not a tool; keep one source of truth by referencing @AGENTS.md inside CLAUDE.md; have Claude Code mirror .claude/skills into .agents/skills with a sync hook so editing once updates both; and import .mcp.json into .codex/config.toml so the same MCP servers run in both runtimes. A clean cross-runtime config you can copy.
@noahlh [Claude Code]
https://x.com/noahlh/status/2065997556971376717
He finds the Claude Code harness better right now, especially with workflows/ultracode, so he has Opus 4.8 lead via Claude Code and uses Codex (5.5 on xhigh) as one of the workflow subagents to implement and do adversarial review. It worked really well on two big tasks he finished that day. He notes you can ask Claude to design a skill to invoke Codex, which is handy because Codex stays usable via CLI rather than at API rates. A real working multi-runtime split.
@ivanzhouyq [Claude Code]
https://x.com/ivanzhouyq/status/2066019658470392021
Frames today's agent workflows as fragmented, a few Claude Code sessions, a few Codex sessions, agents with different permissions and memory, sessions stuck on one machine, and pitches Omnigent as the fix. The vision: start a coding session on your laptop, switch the agent from Claude Code to Codex mid-session, keep it running on a server overnight, check status from your phone, all while agents share the same skills and memory, and even share a live session with a teammate. A concrete take on unifying multi-agent work.
@gagarotai200 [Claude Code]
https://x.com/gagarotai200/status/2066080960026959954
A vivid orchestration vignette: while playing Taiko no Tatsujin, a Hermes Agent drives both Codex and Claude Code to run his X account, video editing and system development. The role split is explicit, Codex edits video, Claude handles system development, Hermes runs X operations, each just given its role and then left alone. His point: hand out the roles and go play the drum game; daily work now just happens.
@yibie [Claude Code]
https://x.com/yibie/status/2066106381757985208
Shares Stripe's deep write-up of "Minions," its fully unattended one-shot coding agents, with concrete scale: over a thousand merged PRs per week are generated entirely by Minions with no human-written code, only human review. A run starts from a Slack message in an isolated prewarmed devbox, uses a fork of Block's goose agent loop interleaved with deterministic git/lint/test steps, reads the same agent rule files as Claude Code and Cursor, and connects to a central MCP server (Toolshed) hosting 400+ tools, capped at two CI rounds. The clearest picture yet of unattended agents at real production scale.
@Neko_ai_p [Claude Code]
https://x.com/Neko_ai_p/status/2065986582948331704
A genuine non-engineer (started this year from "what's a terminal?") reports that an HR app he built step by step with Claude Code got adopted by a friend's company and is going into operation. He's honest that he doesn't yet know how it'll go in real use, but it's a real shipped output reaching a real workplace from someone with zero background. The clearest kind of barrier-collapse story.
@swarm_japan [Claude Code]
https://x.com/swarm_japan/status/2066170043420102933
Summarizes a 26-minute video where a Google Cloud engineer builds an app from scratch in Claude Code by handing it five roles: PM, UI/UX, engineer, security, analysis. The interesting part is the structure: Planning Mode where Claude proposes the design first and the human reviews and approves before anything runs; it runs on Google Cloud with no API key needed and data staying inside the internal project; and a non-technical PM can go from wireframe to a GitHub PR. Control of the whole project sits with Claude while the human acts as the gate.
@kokisennyu [Claude Code]
https://x.com/kokisennyu/status/2066111516928360830
A neat hybrid-compute workflow done entirely inside Claude Code: develop locally, offload only the heavy processing to Colab's GPU, then pull results back to local. He notes it pairs well with AI development, OCR verification, image processing, LLM experiments and RAG embedding generation. A concrete way to borrow cloud GPU without leaving the agent.
@oops073111 [Claude Code]
https://x.com/oops073111/status/2066011091227029724
A practical automation trick for private business systems with no API docs: log in by hand once, save the backend's cookie and bearer token, then let the agent manage them to fire POST requests and operate the system on its own. He says it works especially well for B/S-architecture systems, and notes the agent can reach a private system's endpoints without knowing the API spec. He uses it, mostly through Claude Code, for financial stats, KOL incentive distribution and CLI tasks.
@sitinme [Claude Code]
https://x.com/sitinme/status/2066128834143920405
A three-phase engineering workflow built around requirements, not code. Instead of jumping to implementation, the engineer first makes Claude Code act as a product interviewer that keeps probing the project goal, who the users are, the core scenarios, the must-have features, the boundary conditions, and produces a clear requirements spec before any code. His thesis: most AI coding failures aren't the model failing to code, they're the model starting from incomplete or wrong context, so clarifying the problem first cuts downstream uncertainty.
@alphabatcher [Claude Code]
https://x.com/alphabatcher/status/2066243646417150447
A concrete 4-agent Claude Code setup built on the idea that the harness, not just the model, is what makes it work. A writer agent writes code, a tester agent tests against the spec, a reviewer agent attacks the diff, and a coach command writes the brief and calls the play, via writer.md, tester.md, reviewer.md and ship.md run with /ship. His point: letting one agent do every job in a single blurry session gives you code, tests and review all from the same context, which is the mistake.
@majidmanzarpour [Claude Code]
https://x.com/majidmanzarpour/status/2066160513969537470
He open-sourced a three.js "game director" skill system for Codex and Claude Code that helps agents make more polished playable browser games. It guides gameplay loops, graphics, HUD/UI, debugging and QA, with optional 3D/image/audio asset generation through Tripo AI, ElevenLabs and Nano Banana via API keys. A concrete shareable skill that pushes agents toward real game output, not just code snippets.
@shinshin86 [Claude Code]
https://x.com/shinshin86/status/2066286269928534235
He built a page where you copy a prompt into Codex or Claude Code and the agent scaffolds a working original AI-VTuber chat app, which you then grow by talking to your own agent. Built on AITuber OnAir, the app supports LLM/TTS/comment integration configurable from a settings screen after launch, and lets you pick a PNGTuber, VRM or Live2D avatar. A concrete shipped template that turns a prompt into a real, extendable app.
@VincentLogic [Claude Code]
https://x.com/VincentLogic/status/2066138192663712033
Clarifies that his logo-to-animation tool isn't a standalone app but an Agent skill: load it into Claude Code or Codex, say "turn this logo into an animation," and it runs. The flow: the AI analyzes the pixel image and identifies the shape structure, rebuilds it with vector paths while comparing error against the original, picks a fitting animation style, and outputs clean dependency-free SVG you can embed directly in a page. He notes Disney's 12 principles (squash/stretch, anticipation, ease-in/out) make the result feel natural rather than mechanical.
@designoutloud_ [Claude Code]
https://x.com/designoutloud_/status/2066277806024339695
A small, satisfying craft build: he made a working font with AI where every letter is constructed from just two shapes, a square and a circle. He used ChatGPT for the design logic and Claude Code to make the font actually functional, building on another creator's workflow as a technical starting point, then tested it in Photoshop and confirmed it works. A concrete output in a domain (type design) most wouldn't try with an agent.
@kohya_tech [Claude Code]
https://x.com/kohya_tech/status/2065982822654107694
A concrete open-source contribution with measured gains: he made a PR to sd-scripts that turns the Qwen-Image VAE into a 2D-only variant for Anima, cutting memory to about a third and processing time to roughly three-fifths. The division of labor is notable, Codex did the implementation and Claude Code did the review. AI on both sides of a real performance optimization.
@obscaries [Claude Code]
https://x.com/obscaries/status/2066173061980713348
Introduces Claude-Red, a collection of offensive-security skill packs that supercharge Claude Code for cybersecurity work, spanning web security, exploit development and red-teaming, so the AI reasons more like a security researcher. A concrete domain-specific skill pack that points Claude Code at security work rather than general coding.
@m7md_dev1 [OpenClaw]
https://x.com/m7md_dev1/status/2066201982914732469
While learning Claude he stumbled on a "cowork" feature that controls your machine, and it set off a lightbulb: since he's been trying to automate job hunting, why not give OpenClaw a spare machine to control and have it apply to jobs on LinkedIn for him, driven by workflows. He tags it as a "failed experiment" for now, but it's a concrete intended non-coding automation, an agent applying to jobs autonomously on a dedicated machine.
@potatobiz1 [Claude Code]
https://x.com/potatobiz1/status/2066033879153349080
A nice cross-device build note: he made an app in a few hours, doing the hands-on part while lying down on his phone with the Claude Code mobile app, then finishing the harder half on a PC. He stresses how good it feels to build whatever app you want and use it the same day, no App Store review in the way, and that being able to just ask the AI from any device made setup painless. The everyday joy of same-day personal software.
@tetumemo [Claude Code]
https://x.com/tetumemo/status/2066091488099127525
A clean content-automation loop: he turned 17 past newsletter articles into a podcast with two AI hosts via NotebookLM, which read and break the content down even for long pieces. The flow runs from article-finished, triggers a Skill, executes NotebookLM podcast creation from Claude Code, then downloads the file, and he found the whole thing remarkably smooth. A concrete non-coding content pipeline built on the agent's skill system.
@ClaudeCode_UT [Claude Code]
https://x.com/ClaudeCode_UT/status/2066109644981080226
A knowledge-compounding pipeline that wires Claude Code, NotebookLM and Obsidian together to auto-accumulate sources. Claude Code searches out 10 related YouTube videos, hands them to NotebookLM (processed on Google's servers, so it costs zero of your own tokens), which generates mind maps, flashcards and infographics, and the structured, pre-linked Markdown auto-saves into your Obsidian vault. He claims 10 sources accumulated in 6 minutes after a 30-minute setup, arguing the system that automates collection-to-storage beats the effort of taking more notes.
🗣 User Voice
User Voice
Token cost vs. depth is the central tension: people deliberately let workflows burn (@shion_takk's 10-hour Blender run) while others fight the limits, with @ClaudeCode_UT showing you can run far deeper on the same Fable budget just by making it orchestrate and only handing heavy reasoning to Opus.
Memory is the most-requested missing piece: @0xrevayz says a local model is useless without it and wires Claude to an Obsidian vault through MCP, while @ClaudeCode_UT and @tetumemo both build NotebookLM-to-Obsidian pipelines so knowledge accumulates instead of evaporating.
Multi-runtime orchestration is becoming the default ask, not an edge case: @PawelHuryn, @noahlh and @ivanzhouyq all want Opus and Codex working as peers with one source of truth, shared skills, and sessions that survive across machines.
Design taste still has to be injected: @ridark_eth's whole $9k/month pipeline hinges on MCP passing a real design system to Claude as structure, so screen five still matches screen one.
And losing Fable stings in concrete ways: @nptacek noticed sessions that kept Fable's context outperform baseline, and @shion_takk and @teach_fireworks both built their heaviest workflows on the assumption the top model would keep showing up.
Token cost vs. depth is the central tension: people deliberately let workflows burn (@shion_takk's 10-hour Blender run) while others fight the limits, with @ClaudeCode_UT showing you can run far deeper on the same Fable budget just by making it orchestrate and only handing heavy reasoning to Opus.
Memory is the most-requested missing piece: @0xrevayz says a local model is useless without it and wires Claude to an Obsidian vault through MCP, while @ClaudeCode_UT and @tetumemo both build NotebookLM-to-Obsidian pipelines so knowledge accumulates instead of evaporating.
Multi-runtime orchestration is becoming the default ask, not an edge case: @PawelHuryn, @noahlh and @ivanzhouyq all want Opus and Codex working as peers with one source of truth, shared skills, and sessions that survive across machines.
Design taste still has to be injected: @ridark_eth's whole $9k/month pipeline hinges on MCP passing a real design system to Claude as structure, so screen five still matches screen one.
And losing Fable stings in concrete ways: @nptacek noticed sessions that kept Fable's context outperform baseline, and @shion_takk and @teach_fireworks both built their heaviest workflows on the assumption the top model would keep showing up.
📡 Eco Products Radar
Eco Products Radar
Codex - by far the most-paired tool today, run as a peer to Claude Code for ideation, implementation and adversarial review
Fable 5 - the top model pulled mid-week under export controls; its absence drove most of today's "what now" workflows
OpenClaw / Hermes - the persistent self-hosted agents people run alongside session tools, with Hermes increasingly the orchestrator
Cursor - still the common reference point for session-based coding agents
MCP - the connector layer behind nearly every serious workflow, passing Figma structure, browser control, and Obsidian memory
Gemini - the go-to vision/video reader inside multi-tool pipelines (ad analysis, content)
Opus 4.8 - the fallback heavy-reasoning model after Fable, used as orchestrator or escalation tier
DeepSeek / GLM 5.2 - cheaper coding backbones people swap behind the Claude Code harness
Obsidian - the recurring memory substrate, wired in via MCP for persistent knowledge
NotebookLM - the content engine of choice for turning sources into podcasts, mind maps and flashcards at zero token cost
Codex - by far the most-paired tool today, run as a peer to Claude Code for ideation, implementation and adversarial review
Fable 5 - the top model pulled mid-week under export controls; its absence drove most of today's "what now" workflows
OpenClaw / Hermes - the persistent self-hosted agents people run alongside session tools, with Hermes increasingly the orchestrator
Cursor - still the common reference point for session-based coding agents
MCP - the connector layer behind nearly every serious workflow, passing Figma structure, browser control, and Obsidian memory
Gemini - the go-to vision/video reader inside multi-tool pipelines (ad analysis, content)
Opus 4.8 - the fallback heavy-reasoning model after Fable, used as orchestrator or escalation tier
DeepSeek / GLM 5.2 - cheaper coding backbones people swap behind the Claude Code harness
Obsidian - the recurring memory substrate, wired in via MCP for persistent knowledge
NotebookLM - the content engine of choice for turning sources into podcasts, mind maps and flashcards at zero token cost
Comments