Building @ https://nitter.gq/t.co/IKM1RhSsA9 | Ex-Microsoft | Sharing build logs and reality checks | Gamer & Movie Buff 🎮 | Foodie & traveller
Joined February 2010
- Tweets1.6K
- Following151
- Followers138
- Likes457
Planning to start a new experiment until the weekend: start one claude code session per task with 5 line handoff note at the end.
let's see if this reduces drift + consumes less tokens.
will post back results on monday.
I've never used sub agents ? what's the use case if anybody use it on a regular basis ?
If claude adds a cache to your serverless function, check where it lives. in-memory means every instance gets its own copy, so once it scales out you're basically uncached.
if you use claude code put .env in the deny rules in settings.json. otherwise it can just cat the file to check the config
Claude tip of the day: feed Claude the git log since the last release and have it draft the app store’s "What's New" notes in plain language.
Takes hardly two minutes and beats having to scratch your head over what to fill and staring at a blank box :)
this is really cool if true !
This is a masterclass in how to handle layoffs.
Today, Canva CEO’s announcement is next-level classy. Thoughtful, honest and generous.
CEOs worldwide should take note!
Canva is laying off 22% of its workforce, but look at how they’re handling it.
6 months of salary, 12 months of healthcare, career support, removal of the equity vesting cliff, a $10,000 fund for employees who want to start their own business, and even letting employees keep their MacBooks.
You can’t always avoid layoffs, but you can choose how you treat people when they leave.
Vineeth retweeted
October was a big month for Claude Code 🍁
We shipped mods, made the desktop app better at running lots of sessions side by side, gave you more insight into where your usage goes, and plenty more. Here's a recap!
The hardest part of moving from bigtech (ex: Microsoft/FAANG) to an indie dev is the size of the problem you deal with on a daily basis.
When you start out, in your mind, you would think you are gonna create the next google or instagram now that you have the time, but in reality your post-job/no-regular-salary budget allows you to create only a simple ios/macos app !
Agree / disagree ?
If you are someone with a Grok premium sub and creating videos for your ads , here's a tip that might help:
Grok imagine agent mode is really good for building and stitching videos. In the regular mode, the limit is 15 seconds with one image and 10 seconds when you've more than one reference. Instead of creating 10 and extendign to 30 (which tends to lose the audio, lip sync and causes face drift over time), just use agent mode and ask it to create 3 10 secs videos and stitch and share the final 30 seconds video, works so much better.
Just add "create new clip exactly from where the previous clip stopped, and return a seamless 30 seconds video".
I've tried this for upto 45 seconds videos too, it's great at understanding spatial differences and maintain references. Even accent (I tried with both US and Indian accent, and both work really well).
Most of the leaks from vibe coded apps come down to the usual suspects:
1. database rules left open,
2. API keys shipped in the client, and/or
3. storage buckets set to public.
Before launch, make it a habit (best is to add in one of your .md files) to audit exactly those three and show proof, not just that it "looks good".
It takes a minute, but saves you a lot of misery post release.
Claude tip of the day: Instead of using a agent memory product, use a 30-line CLAUDE.md with your rules + a decisions.md file that the agent appends to after every fix. It covers most of the "it made the same mistake again" pain.
Only add additional tooling when the text files stop working, not before.
The interesting part here isn't that it was done in 2 hours. It's the pipeline he used: a strong model for the hard extraction and a cheaper one for bulk filtering.
Most people overlook this simple thing, they run every step on the biggest model and complain about rate limits catching up.
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org/
- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
Hot take: if every LLM call in your app hits a flagship model, you don't have an AI cost problem, you have a routing problem. Classification, extraction and "is this spam" don't need the smartest or the most advanced model.
save the big one for the step that actually needs that level of reasoning.
unpopular opinion: most AI apps don't need an agent framework.
A while loop, three well-described tools, and a step limit get you surprisingly good results. Add the framework only when you are sure what problem it solves for you.