🔥 https://nitter.cf/t.co/fkFGkvvnUT - Selfhosted Web Analytics 🥇 https://nitter.cf/t.co/kzwcIogsZy - AI Benchmarks 🎹 https://nitter.cf/t.co/JmOxsCH5J8 - Learn piano by playing games 15+ years SWE experience. Love 🏓
Joined October 2016
- Tweets5K
- Following203
- Followers367
- Likes10.2K
Pinned Tweet
I have to say this again, and again, but it is VERY IMPORTANT when working with LLMs and making sure you don't run out of tokens.
Do not ask the AI to do something only once (i.e. verify a screenshot), ask the AI to create a system/script to always automatically verify and report those issues in an optimized way.
Also, once a system is implemented, don't stop there. Ask it again, after a few runs, to check the executions and see if it can improve performance/efficiency of the system. And, not only once, this improvement check should be done regularly over time, as your tasks or data do change, so the system has to be "recalibrated".
For example, I just tried using Cludflare crawl/Browser Run to get some info automatically, and Astra xhigh implemented the basic API call to get the data. After this, I suggested to try again, but see if it can reduce the Browser Run time or load, while still retrieving the same data, it did it:
"The light web retrieval test worked: 0.78 browser seconds, down from 4.34 seconds—about 82% less for the same product. Data accuracy checks still passed."
Oh wow, I am paying for some Reddit ads and getting users on a new project of mine, and having @uxwizz on that site helps so much, from almost every visit I got I learned something and changed/improved the website based on that.
Some things that the users did which I didn't think of/forgot to implement:
- clicking on product image to enlarge it
- clicking on an icon and expecting a popover instead of being taken to somewhere on the site
- instead of searching for a product name, they copy-paste entire product title + extra links in the search bar, so search returns nothing, I should make search ignore the extra fluff in search
- keep browsing a category for 4-5 pages without finding what they are looking for (for this case, I asked the AI to implement a subtle pop-over that appears after a few pages promoting another feature where you can simply ask what you want)
Anyway, having session recordings and fixing UX issues makes working with AI so much easy. You just ask the AI to implement stuff, see what users do, then ask to fix where they get stuck.
What is this @ClaudeDevs ?
I can only use Claude on some websites?
I want to give Reddit money, so asked Claude to help me fill the ads form, but it can't, so now I have to just copy-paste stuff like a peasant :(
I lost it lol, I keep asking Claude for advice with something, and it keeps telling me to go to sleep after each reply, lol
Lol, I got my first customer without even noticing, before I even released the app.
I take that as a good sign!
$6.99 MRR, nice!
Not sure if I am doing something wrong, but I still struggle to have Opus do things without my approval with Auto.
It seems to always get stuck in basic commands/flow that Codex could easily do, then it just goes in a loop.
And done, I really wanted to keep using ChatGPT, but they really are focusing on the wrong things.
They keep adding features, Dots and stuff, which are not even available in my region, instead of fixing the core chat functionality which doesn't work 50% of the time.
The models and value are great, but if it doesn't work half of the time, the experience becomes very frustrating.
I have switched to Claude Max 5x for half the price ($99 vs $200), mostly using Opus 5.5 and ocassionally Fable 5.1.
Am I dumb? What even is this question from Opus:
"Should the vendored flipMath.js ever be re-extracted from the patched page-flip package, or can the page-flip dependency and its patch be dropped once the orphan PlacementPageTurn.tsx is removed?"
Anyone knows any snipboard.io (former snag.gy) alternatives?
It seems to be gone.
Websites where I can just CTRL+V images from clipboard to share them.