Super User Daily: April 16, 2026
April 14 was the day Claude Code Routines dropped and OpenClaw shipped its 4.14 reliability update. But the real story is what users did with these tools. A home builder reconstructed five years of construction history from raw emails. A content creator automated 1,620 videos per month. A quant finance researcher let agents run overnight and woke up to a real strategy. The line between "using AI" and "delegating to AI" got blurrier than ever.
@dvassallo [OpenClaw]
https://x.com/dvassallo/status/2044088256753799567
Daniel Vassallo pointed OpenClaw at his entire email inbox from a five-year custom home building project and told it to reconstruct what happened. The agent found 1,850 emails across 450 threads involving 58 people at 35 organizations, produced 511 timeline events, identified 690 documents, extracted 170 finance records, and mapped every contact with their roles. For the first time in five years, he could see the full arc of his project: every dollar, every decision, every contractor relationship. The agent even wrote personality profiles for each contact based on their communication style. This is not summarization. This is narrative reconstruction from raw correspondence.
@shannholmberg [Claude Code]
https://x.com/shannholmberg/status/2043983746094026984
Shannon Holmberg adapted Karpathy's autoresearch pattern for marketing. The framework called "autoreason" uses adversarial agent loops: one agent writes positioning, a separate critic tears it apart, a fresh author rewrites from the critique alone, a synthesizer merges both, and a blind judge panel scores all three via borda count. The winner becomes the new baseline and the loop repeats until a version survives two rounds unchallenged. Every role is an isolated agent with no information leakage between rounds. When you add real campaign performance data to the knowledge layer, the loop stops arguing from generic copywriting principles and starts arguing from your own numbers.
@FletchPh [Claude Code]
https://x.com/FletchPh/status/2044048418906018293
Fletcher integrated a Cold War fighter pilot decision framework (OODA) into Karpathy's autoresearch. The implementation is a 40-line orient.md living document that the agent rewrites after every experiment, tracking what works, what failed, and ranked hypotheses for what to try next. The result was a 92% keep rate (nearly 2x baseline), better validation loss, same wall clock time, fully autonomous. Every 10 experiments, the system deliberately tears down and rebuilds its mental model from scratch, a concept Boyd called "destruction and creation" that prevents premature closure.
@dphuang2 [Claude Code]
https://x.com/dphuang2/status/2043899379040563378
David Huang pointed Claude Code at a golf forecasting research task and let it run for 49 hours on Tinker with no human in the loop. The agent ran 108 experiments across 100 git commits, switching between model families, RL and SFT, different prediction formats. The entire trajectory is open-sourced, including the experiments that made things worse. This is the clearest public example of large-scale autoresearch on a non-ML domain running autonomously for days.
@Zenzhe99 [Claude Code]
https://x.com/Zenzhe99/status/2043860330309140620
A Chinese user documented how Sandy Lee, a 550K-subscriber YouTuber, built seven AI agents in Claude Code for end-to-end content production. The agents handle channel analysis, trend scouting with an outlier score algorithm, scriptwriting, thumbnail generation, video editing via Descript API, daily reporting to Google Sheets, and short-form video conversion. The system prioritizes small creators for trend detection since big channels get views regardless. Total cost is under $200 per month. One person, wearing pajamas at home, running a content pipeline that previously required a team.
@mirano_uranai [Claude Code]
https://x.com/mirano_uranai/status/2043891478598951296
A Japanese user completed a fully automated video posting system using Claude Code and Remotion. The numbers: 6 accounts, 3 videos per day each, across 3 platforms (YouTube, Instagram, TikTok), for 30 days = 1,620 videos per month. Previously they were manually uploading 500 videos per month. The system incorporates their viral templates and runs completely hands-free.
@09pauai [Claude Code]
https://x.com/09pauai/status/2044038257294577755
A Japanese user running an automated YouTube channel hit 1,821 subscribers with zero work time, using 7 prompts for end-to-end production from scriptwriting to posting. On a separate 715-follower sub-account, they made 180,229 yen in 27 days by having Claude Code research food items recommended by Japanese TV personalities and creating content around them. The strategy is replicable across multiple celebrity recommenders.
@cryptopunk7213 [OpenClaw]
https://x.com/cryptopunk7213/status/2043892812815708257
A parent is running 11 OpenClaw agents on Mac Minis to manage their family. Agents homeschool kids from photographed curriculum with personalized lesson plans and progress tracking. Voice-only interface: leave voice notes and agents write code, order groceries, handle the to-do list. The system even schedules intentional boredom time for kids. All house admin runs in the background. The hours freed up are spent actually being with the kids. Dystopian on the surface, but functionally it is the most complete home automation setup documented so far.
@RetroChainer [Claude Code]
https://x.com/RetroChainer/status/2044134455862825005
A user built a Polymarket weather trading bot on a Mac Mini with Claude Code in 9 hours. Woke up to 6 filled orders with edge on every single one. The bot does not predict weather. It reads GFS forecasts faster than the crowd updates prices, and that speed window is where the money prints.
@MichLieben [Claude Code]
https://x.com/MichLieben/status/2044079352120520959
A $7M ARR agency built a complete outbound sales system in Claude Code. The system scores a target list against their ICP, pulls sales leaders via Apollo, enriches missing emails, fetches best-performing copy from Instantly's API, and loads 154 leads into Instantly with copy and schedule set. Their head of GTM walks through the entire build on camera and packaged the GitHub repo, CLAUDE.md, and Python scoring scripts as a public starter kit.
@everestchris6 [OpenClaw]
https://x.com/everestchris6/status/2044113631030456523
An OpenClaw bot scans every commercial parking lot in a city for missing EV chargers. When it finds one, it sizes the install on the actual lot using satellite imagery, computes federal tax credits and state rebates, renders branded EV chargers on the parking lot with AI, and mails the building owner a postcard with before/after images and ROI calculations. The entire pipeline from property detection to mailbox runs without a human.
@nash_su [OpenClaw]
https://x.com/nash_su/status/2043881733909844378
A user shared an OpenClaw setup that monitors NASA wildfire satellites. When a fire starts, the bot finds nearby high-risk houses, renders fire-hardening upgrades on the actual homes, and automatically mails postcards to homeowners. Everything runs fully autonomously. This follows a pattern of using OpenClaw for geographic opportunity detection combined with automated physical-world outreach.
@AlchainHust [Claude Code]
https://x.com/AlchainHust/status/2043878638475718981
A user applied the autoresearch hill-climbing pattern to 60+ Claude Code skills instead of ML models. The system called darwin.skill scores each skill across 8 dimensions, finds the weakest dimension, fixes one thing, has an independent agent re-score, and commits or reverts based on whether the score improved. After 38 commits, skills went from fragile to production-reliable. The system is now embedded inside Nuwa (9,000+ stars in one week), where every skill goes through Darwin evaluation before delivery.
@doodlestein [Claude Code]
https://x.com/doodlestein/status/2043885540798251191
A user built a Claude Code skill that replaces TurboTax. Last year they had to manually paste screenshots into GPT. Now the skill handles the entire tax workflow and provides strategic advice more sophisticated than TurboTax or H&R Block. They argue Intuit stock at over $100 billion is exposed to these disruptive forces because the product is now meaningfully worse than what frontier models can do with a single skill.
@shmidtqq [OpenClaw]
https://x.com/shmidtqq/status/2044082383591219320
A user cancelled ChatGPT Plus, Claude Pro, and Gemini Advanced ($60/mo total), replaced them with Ollama running GLM-5.1 Cloud plus OpenClaw in 10 minutes. Now they have an AI agent that texts on Telegram, searches the web, handles files, and runs automated tasks on a schedule, all for zero dollars per month.
@abhijitwt [OpenClaw]
https://x.com/abhijitwt/status/2044076597985325355
Oliver built an agent called Larry using OpenClaw plus Claude, connected to Postiz for distribution. The agent runs content end-to-end. Results: 500K+ views in 5 days, multiple 100K+ posts. He wrote about the process, which got 7M impressions, driving Postiz to $60K MRR. Classic distribution-beats-product play executed entirely through an autonomous agent.
🗣 User Voice
User Voice
The dominant complaint on April 14 is that Opus 4.6 feels nerfed. Multiple users report shallow reasoning, skipped steps, and the model ignoring CLAUDE.md instructions. Boris Cherny confirmed on Hacker News that adaptive thinking can assign zero reasoning tokens to some turns, and that Anthropic reduced cache TTL from 1 hour to 5 minutes. Users paying $200/month for Max plans report hitting limits in under 20 minutes. The community response is split between downgrading to version 2.1.98, disabling adaptive thinking, and switching to Codex. @om_patel5 documented the creator's response and the community pushback. @boringmarketer asked bluntly if anyone else feels Claude Code is lobotomized. @Sthiven_R shared a detailed fix involving version downgrade plus one environment variable. The pricing gap between Pro ($20) and Max ($120) also surfaces as a pain point, with @markevans wishing for a middle tier. Meanwhile, Codex is gaining ground among senior engineers who value code quality over speed. @om_patel5 shared a principal engineer's 120-hour comparison finding Codex behaves like a "5-6 year senior" while Claude Code behaves like "an engineer on a time crunch."
The dominant complaint on April 14 is that Opus 4.6 feels nerfed. Multiple users report shallow reasoning, skipped steps, and the model ignoring CLAUDE.md instructions. Boris Cherny confirmed on Hacker News that adaptive thinking can assign zero reasoning tokens to some turns, and that Anthropic reduced cache TTL from 1 hour to 5 minutes. Users paying $200/month for Max plans report hitting limits in under 20 minutes. The community response is split between downgrading to version 2.1.98, disabling adaptive thinking, and switching to Codex. @om_patel5 documented the creator's response and the community pushback. @boringmarketer asked bluntly if anyone else feels Claude Code is lobotomized. @Sthiven_R shared a detailed fix involving version downgrade plus one environment variable. The pricing gap between Pro ($20) and Max ($120) also surfaces as a pain point, with @markevans wishing for a middle tier. Meanwhile, Codex is gaining ground among senior engineers who value code quality over speed. @om_patel5 shared a principal engineer's 120-hour comparison finding Codex behaves like a "5-6 year senior" while Claude Code behaves like "an engineer on a time crunch."
📡 Eco Products Radar
Eco Products Radar
Claude Code - Anthropic's agentic coding tool, now with Routines for server-side autonomous execution. Dominant tool in the ecosystem.
OpenClaw - Open-source AI agent framework, 199K+ GitHub stars. Shipped 2026.4.14 with GPT-5.4 routing improvements.
Codex - OpenAI's coding agent, gaining adoption among enterprise developers who prioritize code quality.
Hermes - NousResearch's agent, increasingly used alongside OpenClaw in multi-agent setups.
Obsidian - Knowledge management tool, becoming the default "second brain" for AI agent workflows.
Tinker - Cloud compute platform for running long autonomous agent sessions (hours to days).
Remotion - Video generation framework, used for automated YouTube/TikTok production.
Ollama - Local model runner, enables zero-cost agent setups with GLM-5.1 and Gemma.
Nuwa - Agent skill platform with 9,000+ stars, integrates Darwin-style skill evolution.
Warp - Terminal emulator with first-class support for Claude Code, Codex, and other CLI agents.
Postiz - Social media distribution platform, reached $60K MRR through agent-driven content.
Claude Code - Anthropic's agentic coding tool, now with Routines for server-side autonomous execution. Dominant tool in the ecosystem.
OpenClaw - Open-source AI agent framework, 199K+ GitHub stars. Shipped 2026.4.14 with GPT-5.4 routing improvements.
Codex - OpenAI's coding agent, gaining adoption among enterprise developers who prioritize code quality.
Hermes - NousResearch's agent, increasingly used alongside OpenClaw in multi-agent setups.
Obsidian - Knowledge management tool, becoming the default "second brain" for AI agent workflows.
Tinker - Cloud compute platform for running long autonomous agent sessions (hours to days).
Remotion - Video generation framework, used for automated YouTube/TikTok production.
Ollama - Local model runner, enables zero-cost agent setups with GLM-5.1 and Gemma.
Nuwa - Agent skill platform with 9,000+ stars, integrates Darwin-style skill evolution.
Warp - Terminal emulator with first-class support for Claude Code, Codex, and other CLI agents.
Postiz - Social media distribution platform, reached $60K MRR through agent-driven content.
Comments