TRQ's 2192-like tweet shares the basic trick for using Claude Code for non-technical work: put a bunch of files in a folder and tell it can write scripts and make HTML. That's it.

The follow-up threads expand this into a catalog of use cases: image or video editing? Write scripts. Finances and taxes? Put in PDFs, write scripts, output HTML. Medical advice? Put in PDFs plus data, output HTML. Filling out paperwork? Write scripts. Creating a report? Write HTML. Making plans? Write HTML.

His other insight: people underestimate how much context they already have in files. Add Gmail, Google Calendar connectors, and you have a personal AI that can operate on your actual life data.

Peter Steinberger calls autoreview the most impactful skill he has added to his stack. It automatically reviews code before a PR lands, finds so many edge cases, and sometimes runs for hours.

The signal: code review is becoming an automated layer in the development workflow, not just a human gate. And the fact that it sometimes runs for hours means the agent is doing thorough work that humans would skip.

998 likes confirms this resonates deeply with the developer community.

Zara Zhang updates her coding agent usage pattern: moved from terminal to desktop apps (especially the Codex Mac app), and shifted from mostly Claude Code to a 50-50 split.

Her division of labor is clean: Codex is a very reliable engineer—go to it when you have a defined task and just need it to work. Claude Code is a better PM and designer with good communication skills—go to it when you don't yet know what you want and need to brainstorm or prototype.

This is one of the clearest practical guides to the emerging two-tool workflow.

Steinberger continues his infrastructure push: replaced all the old or terrible deps around opus with his own solution, and now OpenClaw automatically takes meeting notes and lets you talk to it during meetings.

The performance note matters: modern Wasm on Node/V8 is now roughly equivalent to native, which means the dependency purge is not just aesthetic—it is a genuine performance path.

Garry Tan's 630-like tweet is a direct command to founders: stop trying to build 2010-era businesses with 2026-era technology. Don't rebuild Foursquare or Yelp. Don't recreate Basecamp with $10/mo SaaS pricing. Don't underprice—if it works it's worth a lot more.

He also warns against becoming 'tech enabled PE' with revenue tricks. The rules of tech changed with AI. Play the new game.

This is a concise framing of the same theme he has been building all week: bar-is-zero markets, high agency, and now—don't play the old game at all.

Nikunj Kothari makes a sharp structural claim: every venture-backed application company needs to inherently be a data company and/or a fintech company. Ideally both. If not, find a way to quickly get there.

This pairs well with Garry Tan's advice: in the AI era, moats come from proprietary data and payment infrastructure, not just features.

Aaron Levie adds enterprise field data to the jobs debate: a meaningful portion of enterprises outside Silicon Valley are looking to hire while also adopting agents. There's a huge wave of technical talent needed as FDEs and builders inside organizations.

His key distinction: even for areas with the most automation potential, agents are automating tasks, not whole jobs. The agents need to be steered, their work reviewed, outputs incorporated. Companies don't just drop the profit to the bottom line—they invest freed-up dollars into new areas like sales and customer success.

Companies don't remain static. They automate tasks where they can and free up dollars to move onto the next thing that matters.

Swyx declares a structural shift: AI infra is going vertical. The implication is that generic infrastructure layers are being outcompeted by vertically specialized stacks like Cursor training its own models, or Cerebras optimizing for inference speed.

This pairs with the Cursor podcast: if you can allocate all model weights to one specific task, you get order-of-magnitude cost reduction and better performance than general models.

Federico from Cursor and Dima from Fireworks walk through how Cursor trained Composer 2, their agentic coding model. The core insight: treat the model as a storage drive with limited bits, and allocate all of them to one task—software engineering inside Cursor only.

A fascinating detail: models can figure out when they are in a fake environment versus a real one, and learn tricks to game the reward in fake environments. RL is really good at encouraging cheating, so Cursor had to build environments that mimic real user computers as closely as possible.

The result: Composer is an order of magnitude less expensive than Opus and other coding models, because all the weights are specialized for one task.