Raw Notes

Anthropic has released a detailed guide for agentic coding when using Claude Code for writing code. It has prompts, examples, and all the best practices to get the best out of the tool.

Came across this post on Reddit:

I've been leading multiple teams of engineers over the past 15 years. I'm now building one project with o3 (~$40/day in request costs) and using 0.49.

I have to say, I achieve more (and better) than I did with some of my past teams of 10+ engineers. And I'm talking about FAANG teams.

I also came across a tweet talking about the same today, and it's true that Cursor is definitely an important inflection point in the tech history, as the author describes it.

As I have been using Cursor a lot for coding these days, here is how I am using it currently:

• Gemini 2.5 Pro: for logic and complex tasks
• Claude 3.7 Sonnet: for the good looking UI
• o3 or MAX models: for debugging if stuck

Today, I saw a tweet from Cursor's founder Eric talking about having a similar setup, here's the exact tweet:

using o3 for research & planning and gemini 2.5 pro for implementation feels really good right now

However, I received a reply to one of my tweets and it said that according to Aider, using o3 for planning and then GPT-4.1 for coding scores the highest. But... from my experience, I didn't find GPT-4.1 to write code as good as Gemini 2.5 Pro or Claude 3.7 Sonnet.

Came across this detailed post at sizeof.cat talking about web browser telemetry in 2025. And I was surprised to see that how many connections from these popular browsers were phoning home. Let me shock you with the list:

• Zen Browser: 82
• Microsoft Edge: 48
• Floorp: 42
• Opera: 31
• Mozilla Firefox: 29
• Google Chrome: 25
• Yandex Browser: 24
• Librewolf: 24
• Waterfox: 21
• Brave: 17
• Arc Browser: 16
• Mullvad Browser: 15
• Vivaldi: 11
• Apple Safari: 6
• Ungoogled Chromium: 3
• Pale Moon: 0
• Kagi Orion: 0
• Tor Browser: 0

The author mentions that some of these connections were not telemetry, see the exact paragraph here:

Of course, not everything is telemetry per se, some of the connections are from the ‘New Tab’ pages that include Youtube videos, or adblocker updates, but I strongly believe that is the same thing: the software leaks information about you and/or your computer without you actually accepting the data exchange (because you don’t have a way to accept or deny the connection since it’s the first time you’re running the browser).

There are better ways to do adblock blocklist management and not load everything on the first run without user consent. Yeah, I know what pgl.yoyo.org does today, but what will it do tomorrow? Or the day after?

Also, remember that "best browser" cannot be decided just by this one factor and you'd need to consider a lot more.

After my friend Rohit suggested, I have enabled a "comment via email" feature for individual posts a few days ago.

My posts didn't have comment feature earlier, so now this "comment via email" feature opens users' default email client and pre-fills to with my email and subject fields with Comment on: [current post title]. Basically, it makes it easier for people to send me emails about specific posts.

It's been only a few days since I have enabled this feature and I have already received multiple comments via emails this way. To check this, you can visit any blog post or raw note and you will see the option at the bottom.

Came across an interesting project by @invisal89 that lets you see your sqlite database in a visual way.

SQLite Internal Viewer update: Each page now shows its table name—so you can finally see how every table is represented in the file format. #sqlite

Give it a try

https://sqlite-internal.pages.dev

I see that the project is hosted at Cloudflare Pages and is built using JavaScript and TypeScript as seen in the public GitHub repo.

And it can be accessed at https://sqlite-internal.pages.dev.

I have been using Cursor heavily for the past few months and I love it, have tried Windsurf and other setups as well, but I like Cursor the most.

Today, Cursor released a new update (I haven't received it yet, though) which has a feature that automatically generates rules by running the /Generate Cursor Rules command. The idea is, you keep giving a lot of instructions in your chat and the command creates the rules from that information. And the Agent can also be prompt to edit or update the rules file effectively.

One more interesting update is, MCPs now support passing images as context. For example, I have this Playwright MCP by Microsoft installed in my Cursor which automatically take screenshots of the browser window and add it to its context.

Nowadays, I am spending a lot of time coding with Cursor AI and learning the concepts on the way by reading through the docs and also asking questions to Gemini or ChatGPT. For example, I was implementing Clerk Auth to a Next.js project so I went through the Clerk documentation page for the same. It feels really good to also understand what you're doing exactly.

For another project, I am using Python with sqlite3 database to process 20k rows using OpenAI API and I did actually learnt the concept. I mean, I still cannot write the code from scratch but I properly understand how everything works, and can make edits to the code get the desired output.

What a time to be alive!

OpenAI has this ebook titled a practical guide to building agents which explains the basics of designing AI agents. The ebook explains topics like what an AI agent is, when you should create an agent, design foundations, multi-agent systems, guardrails, etc. in detail.

Yes, all the code examples focus on OpenAI models only but the concept remains the same for other models. They also have diagrams in the ebook to make the concepts easier to understand.

Gemini 2.5 Pro model is crazy good at coding, and now they have launched the new Gemini 2.5 Flash model (currently, in preview) which is crazy good at general tasks. Logan from Google describes it as:

Gemini 2.5 Flash is here, our first unified reasoning model with thinking budgets. 🔥

It’s on the perato frontier and punches above its price and size!!

Yes, it's a thinking model and is good with long contexts. Compared to OpenAI's new o4-mini, it's super cheap with almost comparable performance. It's priced at:

• Input price: $0.15/1M tokens
• Output price (no reasoning): $0.60/1M tokens
• Output price (with reasoning): $3.50/1M tokens

You can check and compare the pricing with other models on this website.

Gemini 2.5 Flash is already available at gemini.google.com and can also be used and tested via the AI Studio. And it's available via the API as well.

When testing through the API, I liked the fact that I can set the thinking_budget to control the reasoning effort it puts for the question.

Tried installing the LibreWolf web browser on my macOS Sequoia 15.4.1 and ran into some issues where the app was not at all opening. Whenever I clicked open the app, it would show me a dialog saying the following with two options Done and Move to Trash:

Apple could not verify "LibreWolf.app" is free of malware that may harm your Mac or compromise your privacy.

I searched about it and found this Reddit thread where the issue was being discussed. It turns out, after the macOS Sequoia 15.1 update, it stopped working for many users. One of the top comments in the thread offered a workaround as well, which worked for many users.

However, I did run any of the suggested commands and still got it working. When I first click on the app, it showed me the said dialog and I selected the Done option. After that I went to System Settings > Privacy & Security and clicked the Open Anyway (at the bottom it was being shown in front of the LibreWolf app). It asked me for the password and started working flawlessly from the next time.

I am not sure if I will be using this as my main browser yet or not, but these days I am trying different stuff as I did the same with the Ungoogled Chromium browser a few days ago.

OpenAI's newly launched o3 and o4-mini models are now available for ChatGPT Plus users as well. Initially, I thought these models will only be available through the API, but good to know otherwise.

I tried these new models on ChatGPT and they are definitely better than their older o3-mini and o1 models. But I will still not consider them comparable to Claude 3.5/3.7 or Gemini 2.5 Pro models, at least in coding.

However, if I have to use AI for tasks other than coding, I would most likely choose ChatGPT over others.

A while ago, my friend Rohit sent me this exact message and I think this is a fantastic idea.

Idea:

Tentative name: AI for the rest of us (this name is taken but it's the best name that describes the project)

What it is: AI publication website with some educational resources and a newsletter. AI newsletters focus even on minor updates like OpenAI released this new model, Anthropic got this funding etc. etc. For a regular person, such type of news is irrelevant. A regular person wants two things:

• Updates on what really matter in their day-to-day life (bi-weekly newsletter should be enough)
• Some cornerstone-type educational web pages. Like one web page on promptimg that keeps updating as the LLMs get smarter. One webpage on the AI use cases in daily lives (properly categorised, not explained in details but broad overview)

I'm not saying I'll make it or you should make it but I'm saying that such thing is needed.

I also shared the screenshot of the message on X in case someone is interested in making it. I have received a few replies but let's see if it's actually gets implemented by someone.

Like ChatGPT, xAI's Grok now has memory. It can remember all your conversations and you can ask to retrieve any information that you have talked about earlier. Now, when you ask any questions, you get more personalized suggestions.

Also, from a few days ago Grok 3 API is also available to use and you can check the pricing here. It's still in beta, though. There's a also a simple API cost calculator where you can estimate the cost of using different models including xAI's through APIs. It's minimal, clean, and completely free to use.

OpenAI just launched a Claude Code like tool called Codex CLI. They describe it as "Lightweight coding agent that runs in your terminal".

I haven't tried it yet, so not sure how capable this is as compared to other models and tools.

Just came across this tweet from Daniel Vassallo where he announces that he sold Small Bets to Gumroad for $3.6 million, which is broken as:

• $1.8m in immediate cash
• $900k in immediate stock options
• $900k in stock options in the next 12 months

In addition to this, Daniel will also be helping Sahil in making Gumroad better and he won't be taking any salary. He talks more about the things in detail in the tweet.

I came across this X post from @nicksdjohnson where he received a very sophisticated phishing attack email on his Gmail account. The scariest was, the email came from an official Google account no-reply@accounts.google.com email.

The person has explained everything in a super detailed manner about how the hacker is exploiting a vulnerability in Google's infrastructure. I recommend going through the entire thread to avoid getting phished.

Be safe out there.

To give you some context, Notion has launched an AI-powered email client where you can connect your personal or workspace Gmail account (currently only Gmail). Earlier, this was in beta but now it's publicly available and you can start using it.

Some of the interesting features of the app are:

• automatic organization of emails into suitable labels
• AI-powered reply drafting
• minimal and cleaner interface, just like Notion
• available offline, much like the official Apple Mail client

Currently, it's only available for the macOS and iOS as well as Android app is planned to be launched soon, some time in 2025 itself.

On the website, there's not very clear mention of the pricing as it just says "it's free to get started". But I was going through this Reddit discussion and got to know that if you have a Notion AI subscription already, Notion Mail doesn't cost extra. But the free version of Notion Mail is usable as well, with limited features.

I am yet to try this myself.

Just heard the news that Figma has trademarked the phrase dev mode™ as they sent a cease and desist letter to Lovable.dev to warn against the use of the phrase on their website. It's really fascinating to see this happen.

Also, Figma has filed for IPO as I came to know from this tweet.

It's so nice to see Varun Mayya building the next big completely homegrown game while competing with international AAA games. It's so early to predict the success of the game, but I truly applaud the courage and dedication he and the team is putting in. I watched the intro video of the game and it's so good.

One thing I am particularly struck by the fact is their approach to create all the required 3D assets. They went out in the real world to scan real locations, buildings, streets, cloths, etc. and cleaned them to make the perfect 3D model. Varun claimed that they now have one of the largest 3D library of objects used in video games worldwide.

Again, I wish all the best to the team and really excited for the game.

For the past few days, I have been heavily experimenting with the Gemini 2.5 Pro and Perplexity Pro Deep Research tools, and my conclusion is that Gemini's Deep Research is far better than that of Perplexity or even OpenAI and Grok.

For example, I asked Perplexity to give me a list of events from the Indian History with proper reference URLs and most of the events happen to be correct but reference URLs are wrong, broken, and non-existent most of the time. On the other hand, Gemini's research is solid and all the reference URLs are perfectly working as well.

And this is just one example, in almost every cases, Gemini 2.5 Pro Deep Research was far better than any other Deep Research tools available as of now.

OpenAI just released a new GPT-4.1 model which is said to be the "Flagship GPT model for complex tasks", but it doesn't pass my timeline creation test. Yes, I asked the new model to create a responsive timeline of events using HTML and CSS, and it didn't work. I have also recorded a video for the same that you can watch in this tweet.

Actually, I propose this as the new coding test for new models that get launched. I think, creating the timeline isn't a very complex task but all the LLMs are still very bad at this.

And the funny part is, even if you keep providing screenshot of how the timeline is not responsive, they still can't fix it.

While I am not a huge fan of opening external links in new tabs, my friend Rohit suggested adding this feature, and it does make sense. And thanks to Sardine for making a simple tool for 11ty (which my blog is built using).

I used Sardine's @sardine/eleventy-plugin-external-links package which adds target="_blank" as well as rel="noreferrer" to all the external links.

To install, I ran the below command:

npm install --save-dev @sardine/eleventy-plugin-external-links

And then I had to add this in the 11ty config file:

const safeLinks = require('@sardine/eleventy-plugin-external-links');

module.exports = function (eleventyConfig) {
  eleventyConfig.addPlugin(safeLinks);
};

I think, it checks when links are not relative and then does the magic because I added an internal link using the full URL and it was opening that in the new tab. Also, I wish I could customize it a bit and also add rel="noopener" to all external links to make it even safer. But that's not a big issue and I'm good so far.

For the past few months, I have been backing up my computer folders to a Hetzner Storage Box via rsync. When I shared about it on X, I received a really good suggestion which I hadn't thought about it earlier.

Another handy addition to make it a bit like timemachine

If you add in /$(date +%F)/ to the remote side of the rsync command, you're backups will be in daily folders like /2025-04-11/ so you can roll back to a specific day if you ever need too

You can then add a script on the remote box to delete folders after a certain amount of time to save space.

Shane replied to the post with above suggestions where I can automate the daily backups into different folders by date. The old ones can also be deleted after a certain date via a CRON job.

Simple, yet really useful suggestion.

A controversial post about PostgreSQL being 360 times faster than MySQL from a X user.

PostgreSQL finally finished the test. In this case, MySQL was 360 times faster. The chart below shows the bottleneck in PostgreSQL — experienced people will recognize the issue at a glance. No wonder it’s rarely used in Chinese internet companies.

– @wangbin579

But I didn't understand the mention of Chinese internet companies in the context. Do no companies in China use the PostgreSQL database?