Patrick Lu

#throwbackthursday Back before I built software and my dream was to make it as a pro soccer player

@DissidentClint The thing is what happens after the laundry is folded? It still needs to get put away. Where’s the robot to do that?

This is incredible. $2500 to never fold laundry again! WUT Many people are creeped out by humanoid robots. The aesthetics of this are a game changer.

@RameshNivedita Actually, a lazy and midwit take

One of the sharpest views I’ve read on AI and it’s impact at work

If you’re still writing code / watching code be written, you’re 6 months behind If you’re still reviewing code, you’re getting lapped If you’re still running a dev server, you’re so out of the game that your competitor has run circles around you Happy building ✌️

@Jason I’m in

We started an AI founder twitter group... reply with "I'm in" if you're a founder and want to be added

@prateeks PDFs work great but web pages don’t? Brother do you know that AI sucks at reading pdf’s? A whole business Reducto exists cuz AI is terrible at deciphering pdf

Founders, when you share your decks with investors, please share formats that are easier for humans and AI to work with. It is a good assumption that the investor you are reaching out will have some AI processing step in their workflow. PDFs work great! The docsend, gamma and other web formats create friction for the consumer of your deck and do not add much in terms of experience.

@ahmad_roumieh DMs closed 😁

If you want an intro to the speedrun team, hit me up! (Dms open)

PSA: don't put secrets you actually care about in Vercel to begin with. Anything prefixed NEXT_PUBLIC_ is in every visitor's browser by design. Everything else is "private" only as long as your host is. And you really shouldn't be running serious backends on Next API anyways

@MatthewBerman We built this with codepress.dev/terminal Full worktree / workflow isolation so you can work with as many agents as you want (we often do 10 at a time)

How are people running multiple agents at the same time? I’ve tried separate worktrees but things still get messy. I want separate environments for every agent thread. How can I accomplish this locally?

Wot.

@brettcalhounn 😂

The best founders to invest in are the ones who don't want our investment.

@kenanhsaleh @a16z @speedrun Would love to chat!

I'm opening my calendar for 1:1 office hours today and next Friday Applications for the next @a16z @speedrun cohort open next week - come AMA before they do. Happy to chat about your idea, fundraising, or application strategy Reply and I'll DM a Calendly link for 15-min slots

@brycent MIA --> Party

SF → B2B NYC → Consumer MIA → 🤔

We 5x'd the quality of AI working in our codebase. Not by switching models, not by better prompts. By building auto documentation that runs every time we ship. Here's why this matters more than people realize. THE PROBLEM WITH GREP When an agent works on a codebase, the default workflow is grep. Need to understand how user sessions work? grep → 50 hits → read each → grep again → repeat. Three things break: 1. Grep returns matches, not understanding. The agent sees WHERE, not WHY. Can't tell canonical from legacy. 2. Every grep-and-read cycle burns context. Agents end up with 30 files loaded and still miss the architecture. 3. Some things aren't greppable. Why does this code use a custom retry loop instead of the library? That answer isn't in the code — it's in someone's head. WHAT WE BUILT Not "better docs." A knowledge graph. codepress_documentation/ ├── INDEX.md └── features/ ├── agent-sessions/ │ ├── README.md │ ├── redis-protocol.md │ └── session-lifecycle.md ├── authentication/ │ └── README.md └── ... One INDEX file is the root: every feature registered with a slug, area, and one-line summary. Each feature gets its own folder with a README: architecture, key files, keywords for discovery. Complex features have sub-files: Redis protocol, session lifecycle, OAuth flows, linked from the README. CLAUDE.md / AGENTS.md files at the right directory levels point agents to the relevant feature docs automatically. When an agent starts a task, it doesn't grep. It navigates the graph. INDEX → feature README → sub-file for details. Pulls exactly what it needs, nothing it doesn't. HOW IT STAYS FRESH Every merge to main triggers a CodePress automation. An agent reads the diff, decides if the change is worth documenting (skips refactors, lock files, typos), classifies new feature vs update, writes or amends the feature doc, updates the root INDEX, updates the relevant CLAUDE.md / AGENTS.md WHY IT WORKS — ACCURACY Docs are denormalized context. One file replaces 30 grep results. Docs explain WHY, not just WHAT. Docs surface historical context that isn't in the code. WHY IT WORKS — SPEED Agent grepping and reading 30 files = 2-3 minutes wall-clock. Navigating to one feature doc = ~10 seconds. Multiply that across every agent run, every day, every engineer. You're running a different pace entirely. THE BIGGER INSIGHT Every system in your company was designed for humans. File structures. Runbooks. Process docs. Dashboards. Codebase layouts. All optimized for how humans think, read, and navigate. That's the wrong abstraction now. The companies that win the next decade won't be the ones with the best models. They'll be the ones who rebuilt every system — code, docs, ops, knowledge — to be legible to agents. Start with docs. Move to runbooks. Then internal tools. Then the whole stack. The AI didn't get smarter. The codebase became legible to AI. Most "AI doesn't work in our codebase" stories are really "our codebase isn't legible to AI" stories. Built this inside CodePress. Happy to walk through how if you're thinking about this.

@BradWilcoxIFS @lymanstoneky This correlation probably only holds true in US. e.g. many women in China / Taiwan / Korea don’t take husbands surname. Also many European countries. Sounds like FUD

"Spouses who don’t share a surname divorce at about a 50% higher rate than those who do share a surname, and their divorces come about 30% earlier in their marriages." Striking new divorce research from @lymanstoneky reinforces my earlier research on marital quality & naming:

@signulll In the AI era, task management via something like linear is a complete waste of time x.com/lightninglu10/…

can ppl tell me how they manage their linears? ours is getting out of hand cuz it has so much stuff & a lot of not relevant anymore or out dated. i wish an ai agent would proactively consolidate, & clean up. would love that feature. like a "linear master" agent that is continuously running.

@antinertia @ycombinator Need the alpha, DM’ed 😁

can someone from @ycombinator DM me? it doesn’t feel normal that the success of their startup launches is so random, some do millions of views, others flop i’d be happy to teach their founders how to nail a launch and get millions of views, with a simple workshop or OH session, for free

@ivanburazin Share the alpha

If you're actually solving a valuable problem, you'll get legit enterprise logos within months. AI penetration into big companies is happening so fast that you can get case studies with Google, Anthropic, OpenAI, big banks, and Fortune 100s. - Harvey got law firms - Cursor got devs - We got YC companies & Fortune 500s If your product provides value, validation is faster than ever.

@trq212 @trq212 problem I’ve been running into is copying long text I don’t think it works to copy and scroll, so just can’t copy without making terminal larger which only works when I’m on my big monitor

@trq212 Fixes all the problems I used to have with the flickering and weird rendering

Have you tried the new Claude Code renderer? What has your experience been like? If you haven't, you can enable it with: CLAUDE_CODE_NO_FLICKER=1 claude

tried this it sucks try codepress.dev instead

@raphaelschaad 100%. AI is just another developer that can crank out high quality code (when given the right environment). Can then turn that code into outcomes. At the end of the day, everything in the world is code.

I don't think we use AI for too much. I do think we overuse AI too deep in the stack. The mistake is routing work through LLMs when deterministic compute would do it faster and more reliably. Take email triage. One approach is to send every incoming email through a gigantic model, consult markdown skills, update memory, and let the model decide what to do every single time. A better approach is to use AI to study the inbox, detect patterns, and generate/manage rules. Then let deterministic filters do the repetitive work instantly. Archive it. Label it. Move it. No model call needed. This is true for agents, and honestly for humans too: pilots and surgeons handle repetitive tasks best when there's a clear process. The most interesting systems are not AI all the way down. They use AI where intelligence is needed, then hand off to deterministic compute where possible. AI decides the rules. Deterministic systems run them.