Matt Ruiz

@PHLSixersNation Sixers in 6

How we feeling about tonight?

Here's what I do: 1) I never speak directly into the LLM interface and create endless chat wrappers that use the LLM API. 2) Create three tables for every chat message sent: request (your message), response (AI message) and exchange (junction table) and store them in your own database. 3) Create scheduled functions/cron jobs that i) check recent chat and compare with existing knowledge points ii) checks recent knowledge comparisons and score performance. These single knowledge data points can then be categorized (manually or through other three-table AI calls) and used in many different parts of your project/system/etc. I've found much better categorization/fact extraction/etc when the chats are for specific tasks. For example, an app wide AI chat is no bueno for me (at least I haven't dug into that). But, a chat specific to working with a LeadContact is something powerful. The LeadContact chat can access, say, the knowledge data points extracted/refined/etc with the #leadgen and #marketing category/tags are considered in the context. If you're working on a cold email builder then you can have a cold email specific chat which has it's own nightly scheduled function/cron which focuses on knowledge extraction for this specific use case. A side benefit of the three tables is that on the exchange table you store the input/output tokens and the model - this provides a lookback into costs for the chat.

This may be a dumb question but I’ll ask it here anyways: I can’t find a good way for my various AI chats to automatically sync its conversation history into a structured knowledge base. So that as I update various chats from time to time and refine context, my knowledge base automatically grows with this new info.

@marcelpociot A lot of non coding people praising it - rarely saw an experienced dev giving it shine.

So MCP servers aren't really the silver bullet we all thought they'd be, right? Or am I the only one who barely uses any? 😅

Switching to plan heavy has helped a lot with costs. I find myself asking AI to run “npm run build” much less (which is needed to start my express backend). I was doing a lot of “don’t code yet just talk” while in Agent mode and even then it would run into many compile issues.

@Shpigford Maybe they aren’t using planning mode? You burn though usage quick if you just let Claude yolo. But IMO it’ll take you 5X longer, than if you just spend an hour planning with Claude first, write multiple docs, THEN let it yolo. Spend way less credits, get better results faster.

this stumps me. i've had a claude max plan since day 1 and code all day every day and have never hit the limit. but i see folks mentioning it every day.

@jamonholmgren @charliehelps I've been thinking of getting into game dev for this reason.

With over a million trees and 200-odd vehicles all trying to check line-of-sight to each other, this is *very* difficult. It was really kicking my butt. Collision checking is expensive and it was dropping framerate considerably. I asked @charliehelps in a github issue, and it accurately detailed the problem, the issues with my approach, and suggested a very good alternate method. I then dropped the entire issue into @cursor_ai and had it implement it (Charlie could have, but I'm most comfortable with Cursor still). Nearly one-shot it, and it's running at 100+ fps, with accurate line-of-sight calculations.

There should be a hardcoded value in @claudeai models that says: "Never load data inside of useEffects". It seems that no matter what rules/docs/etc I have this pattern shows up.

@jamonholmgren That’s what I’ve been doing. Sonnet with plan mode is pretty reliable before the coding part - at least w/ RN and Express/Supabase.

For reducing token usage, is Sonnet 4.5 a good choice for normal coding, and then switch to Opus 4.5 when Sonnet gets it wrong?

Been revisiting this a bunch

@stevederico @AnthropicAI Good question! This is the "Pricing Explained" screen so it's more informational as opposed to 'buy now'. Maybe that's why Claude went that route? Here's the response after prompting: "Instead of a slideshow, can you add a 3-sectioned tab bar above the pack options"

@MatthewZRuiz @AnthropicAI why is the recommended plan cut off? isn't that the button that drives the most revenue?

Another well-designed mobile screen by @AnthropicAI

Been using Claude, ChatGPT, and Cursor a lot lately for React Native coding. Claude is the best React Native UX Designer/Developer out of the three.

@ionleu user

userID vs. userId ???

@MelkeyDev Definitely react native, expo, typescript, and firebase. Here’s a real-world playlist for learning how React Native works/etc youtu.be/6t1SyQGhebs?si…

What is the best way to learn mobile development? Would it be Expo + React Native, or something like Swift for iOS and Kotlin for Android? I prefer iOS if its between the two operating systems

@daboigbae I code mostly using AI now. In the least, Cursor's tab suggestions are usually spot on - at least with React Native things.

How many times a day do you use AI?

@DonCasso_ It’s still early!

Haha am I too late to the AI gold rush or do you think there is still space to be a differentiator in the space? With the way things are moving Urgency has to be part of your playbook

@Parsafazoli Nice! How often were you posting each day?

Posted tiktoks everyday for 40 days. Waitlist for my app: 60->1,107. If you're thinking about creating content, this is your sign to start.

@LaLightsOn Have always been a fan of Firebase - especially since it syncs up very nicely with React Native. Haven’t tried supabase

My journey was: 1) Building - just start making things change color/shape and move the things around on the screen 2) Docs - once you get a feel for the basics, read the documentation for your UI library/framework/etc of choice. For example, if you were learning React Native, give all of these tabs a read through at least once: reactnative.dev/docs/0.74/gett…. This will help expose you to a lot of concepts - you'll be confused by most but at least you have some sort of reference when you see the terms out in the wild.

Great proof of concept for using @n8n_io and @apify for web scraping

@fuad_c_ Not yet but once there is a 'following' I'll switch to that mode. Mostly solving personal/work problems that I've notice span across longer periods of time. In the least, my work/family life is a bit more efficient. Anything even slightly better is a few extra dollars/month.