Raw notes

I started adding this instruction in my CLAUDE.md file to update itself after every major change and now it's more consistent and hallucinates less as well.

Self-Maintenance Rule

After every major change (new model, new page, new controller, route changes, migration changes, new test files, architectural shifts), update this CLAUDE.md file to reflect the current state. Specifically:

• Add new models/controllers/pages/routes to the relevant tables below
• Update test count if new tests are added
• Add any new gotchas or patterns to the "Gotchas & Pitfalls" section
• Update the "Current State" section if the status changes
• Keep this file as the single source of truth for AI sessions working on this project

I am working in Laravel, so this rule has significantly improved the quality of the code and the model is now a lot more consistent. And this works great with the newly launched Claude Opus 4.6 as well.

Both Claude Opus 4.6 and GPT Codex 5.3 are trending on HackerNews, as you see in the screenshot below:

Yes, it's a rare sight!

OpenAI launched the new model 30 minutes after Anthropic launched the new model. It seems they had some insider knowledge about the launch. It's a crazy amount of cutthroat competition going on in the AI industry, especially between these two giants.

Today, Anthropic released the new Claude Opus 4.6 model, and it's already available in the Claude Code CLI. I have also posted about the same on X, with a screenshot. Yesterday, I posted about the possibility of Opus 4.6 launching soon, by quoting a Reddit post.

As claimed in their launch post, Opus 4.6 is only slightly higher than GPT Codex 5.2 but considerably higher than Opus 4.5. And here is what their API pricing looks like:

I am still looking into it, and will keep this page updated with new learnings.

OpenAI's GPT-Codex-5.3 was also launched today.

Update:

I saw a person talking about this on X, and then this is in their launch post:

Looking more into it.

Apart from this, I just came across the above banner of Claude offering $50 worth of extra usage for testing and exploring the newly launched model. I saw this on the Claude usage page.

Today, OpenAI launched the new GPT-Codex-5.3 model, and it's already available in the Codex CLI and in the newly launched Codex macOS app. You can see my post on X about the same. I have included screenshots there.

OpenAI claims the new GPT-Codex-5.3 model is a lot better than the previous 5.2 model. As you can see in the above screenshot and can also read what they have to explain here:

GPT‑5.3-Codex also better understands your intent when you ask it to make day-to-day websites, compared to GPT‑5.2-Codex.

At the time of writing this, there's no mention of the new model on their models page.

I'm still trying this out, and will keep this page updated with new info.

Anthropic's Claude Opus 4.6 was also launched today.

Update

Matt Shumer had early access to the new GPT-Codex-5.3 and he has written a detailed review of the model. He claims:

Runs can go 8+ hours... and I come back to working code + live deployments.

Very interesting.

Unable to see the new model in the Codex CLI?

I see some people complaining about not seeing the GPT-Codex-5.3 model in the Codex CLI, so first, you need to update the CLI tool by running the following command:

npm i -g @openai/codex

And then the new model will automatically start showing when you run /model in Codex CLI.

Saw multiple posts talking about possibilities of Anthropic soon planning to launch the Claude Opus 4.6 model. One person on Reddit posted the following:

The leaks are real and we are getting Opus 4.6 and Sonnet 5.0 soon, could be today, tomorrow or even the next week?

Just to verify the findings of Macintosh from X I created a new project enabled Vertex AI and verified the results myself, without using the project credentials I was getting 403 but after using my project credentials the results varied real models would return 200 (expected) but opus-4-6 and sonnet-5 returned 403 which means I am not authorized and unsurprisingly, completely made up models returned 404.

This is a great time for them to launch, because I have subscribed to Claude Code Max this time. Haha.

Let's see how real this is. 🤞

Came across this interesting use case of Tailscale installed on an AppleTV with the exit node to a router in a different country, while the AppleTV itself stays in a different country.

This post was shared by kitze on X.

Andrej Karpathy shared this post on X saying nowadays he is spending a lot more time reading long-form articles via RSS/Atom feeds. He says, "... a lot less slop intended to provoke". He also shared a list of 92 top blogs which are popular on HackerNews. It's in .opml format and can be directly imported to most RSS feed readers like NetNewsWire (I use this app already).

I absolutely love this technology of RSS feeds and have been using it for, at least, a few years now. And it feels great to see people going back to the basics and RSS becoming cool again.

VS Code plays a sound when you run a command, send a chat request, edits undone, etc. and sometimes it's just annoying. These sounds are called signals and while you can go to VS Code settings and mark all these signals as never, it's a tedious task as there are 10s of different items that you'd need to change.

So... what's the easiest way then?

Yes, set the volume to 0. For this, go to settings, search for signals volume, and set it to zero or maybe reduce to a smaller number so it's not annoying anymore. It was set to 70 in my case, I think by default.

OpenAI launched a new macOS app for Codex that is more feature rich than using Codex in the terminal. You can watch their quick demo on X and here are some highlights from the demo:

• run and manage multiple agents for multiple projects in one place
• in-built git worktrees to simultaneously work on different features without conflicts
• easier and better way to use and manage skills, usage can be more than coding like writing, etc.
• generate and edit images inside the app to use in websites (this is a cool skill)
• pre-built skills for creating documents like PDFs, spreadsheets, and docs
• /personality command to choose an interaction style you like
• /plan mode (was recently released for Codex CLI as well)
• cool automations that keep running in the background on a regular interval

You can get the app for your macOS device .

Update:

I just installed the app, and here are my first impressions:

1. The app looks good. I love the light mode with system translucency.
2. I didn't need to log in, as I was already logged in to Codex CLI. It's already showing my MCPs installed and other settings. When I opened a project, it already showed my previous chats with Codex CLI in the folder. Cool!
3. I changed a few settings:
   1. Turned on: Require ⌘ + enter to send long prompts
   2. Turned on: Use pointer cursors
   3. Selected the option to show commands and expand output
4. The Skills library is good, you can just click on any skill to install in seconds.
5. It has an in-built terminal as well. I can also run some commands from inside the app.
6. The dictation feature is useful and very accurate.

I passed a few commands in a project and can already say that the experience is much better than using Codex CLI. In the CLI, it felt like a lot of work to install and use a skill, but it's quicker here.

The slash / command in the chat shows some useful options. Running /status or /mcp doesn't disturb the running thread or doesn't wait for the execution to complete, but immediately shows the results in a popup above the chat input box.

Update

Just learned that OpenAI has launched a promo offer for Codex allowing 2x usage limits for all paid plans for 2 months. And now even ChatGPT Free and Go users have 1-month access to Codex. This announcement wasn't in the main OpenAI post, but posted by Alexander Embiricos from their team:

To celebrate the Codex app, we're launching a promo:

• Doubled Codex rate limits for all paid plans (2 months)
• Access for ChatGPT Free and Go Plans (1 month)

And you get 2x limits on all Codex features, whether you use the CLI, app, or web.

Came across this post from Wes McKinney talking about the new tool he's building as the local-first email archive with a terminal UI and MCP server. It's called msgvault and it's powered by DuckDB and comes as a single Go binary. Wes has written more about the new tool in this announcement post on his blog, explaining how it works and what's the future goal for the tool. He writes:

Fundamentally, this is my data, and I should be able to search it in milliseconds, pull up old emails and attachments in a few keystrokes, and query my history with natural language privately and securely. Finally, in 2026, there is no reason that I should not be able to solve this problem. So I did.

And it does make sense.

Basically, you will be able to chat with all your emails via the inbuilt MCP server and search through your entire archive within milliseconds. The tool currently works for Gmail using the OAuth API archiving everything in a local SQLite database. There's a command msgvault delete-staged that also deletes the synced emails from Gmail.

In the future, the tool will also have an .mbox import feature and support for other email services. It can even support importing WhatsApp, iMessage, and SMS archives.

There is much more info about this on the official website that you can check.

I think, this is a great use-case and I'm all for it.

Umami Analytics released v3.0.3 more than a month ago with the patch for the recent Next.js security issue, but I upgraded to the version only recently. To upgrade, I followed the article that I wrote a while ago for upgrading from Umami v2 to v3.

It was a straightforward process, but I did take the Postgres database backup before the upgrade, in case something goes wrong. I ran the following command for the backup:

docker compose exec db pg_dump -U umami umami > umami-backup-$(date +%F-%H%M).sql

Replaced 3.0.1 from earlier with 3.0.3 in the Docker Compose file, as you see below.

umami:
  image: ghcr.io/umami-software/umami:3.0.3

And then just ran the following commands one by one, and then I could access the new version after a minute.

docker compose pull umami
docker compose up -d --force-recreate umami

Also, I wish Umami showed the current version being used somewhere in the dashboard, after logging in. But it's okay.

Codex CLI released the new v0.93.0 and it comes with the most awaited feature Plan Mode, as explained in the release notes.

Plan mode now streams proposed plans into a dedicated TUI view, plus a feature-gated /plan shortcut for quick mode switching.

But the issue is, the /plan command didn't initially appear for me even after upgrading to the 0.93.0 version. It turned out I had to turn on the experimental collaboration mode by adding the following in the Codex config.toml file:

[features]
collaboration_modes = true

And after this I had to again run the installation command npm i -g @openai/codex that output "1 file changed" and then the plan mode was available to use. I could either type /plan or press shift + tab and the plan mode gets activated immediately.

I also replied to some folks who were having issues with seeing /plan even after updating to the latest version.

Claude Code creator Boris shared some tips and best practices to using Claude Code, and it seems actually helpful. I am yet to get a Claude Code subscription but noting this down here for the future:

1. you can do more things in parallel by using multiple git worktrees at once
2. always start in plan mode, unless you're doing something very simple, and you can even ask another Claude instance to review and modify the plan and then only implement
3. keep your CLAUDE.md file updated, end your prompts with "update your CLAUDE.md file so you do not make that mistake again"
4. for each repetitive task, create a skill and can re-use across projects, can setup /slash commands to quickly refer to skills
5. get Claude to fix the bug by itself: enable Slack MCP, paste a Slack bug thread into Claude and just say "fix"
6. use prompts like:
   • grill me on these changes and don't make a PR until I pass your test
   • prove to me this works, etc.
7. use /statusline to customize your status bar to always show context usage and current git branch
8. add "use subagent" to your prompt where you want Claude to throw more compute at the problem
9. if you also want to learn while coding, enable the "explanatory" or "learning" output style in /config
10. ask Claude to explain the code by creating visual HTML presentation
11. ask Claude to draw ASCII diagrams to understand the codebase in a better way

I can see that some of these are really useful, and not just for Claude Code but for other AI-assisted coding tools as well.

Updated

A person took all the suggestions and then created a CLAUDE.md file that is supposedly very helpful when coding using Claude Code.

I haven't tested this yet, but the instructions look fine and should work the intended way.

While looking for a lightweight GUI code editor or text editor, I found a few good editors for macOS (most of these are open-source) that I am going to list here:

1. CodeEdit
2. CotEditor
3. Lapce
4. Smultron
5. TextMate
6. BBEdit
7. Pulsar
8. Brackets
9. Nova
10. SubEthaEdit
11. vim
12. Neovim

I haven't used these all, of course, but have listed them all here in case someday I need these. And, someday, I will definitely get bored of my current setup and start searching for another editor.

I was using Antigravity as my main AI-assisted IDE for more than a month now, but now it's unusable as they have enforced really strict limits. The Opus 4.5 limit gets exhausted with just 1-2 prompts, and then I have to wait for 2-3 days for it to reset.

What now?

I am looking for alternatives and as I am working on a few serious projects now, I want an IDE with higher limits. I was looking to decide between Claude Code Max ($100/mo) and Cursor Pro+ ($60) plans, so I asked for suggestions and most people recommended Claude Code Max as it offers better value for money.

One person on Reddit also recommended GitHub Copilot as it gives you almost 500 Opus 4.5 requests each month in their Pro+ plan. But it turns out that it's not true, as sometimes, if the response is getting longer, it asks you to approve 3 more credits.

Update: I learned that the limit can be increased by configuring the max requests as a person pointed out on Reddit.

And in that case it will become even more costly than Cursor or Claude Code itself, so this is ruled out.

I am leaning more towards Claude Code Max, it's costly but offers better value for money. And I am coding a lot these days so it will make more sense.

However, I am still waiting 1-2 days before getting the subscription and evaluating each option correctly.

I earlier wrote about Clawdbot renaming to Moltbot and later renaming to OpenClaw, and there has been significant development in the space ever since. The OpenClaw GitHub repo has over 123k stars at the time of writing this note, and people are using it for crazy things.

Here I will list some crazy things people are using OpenClaw for:

1. Bhanu is using the bot for doing marketing for his SaaS SiteGPT, as I have talked about it here. He recently posted on X saying, "10 AI agents working 24x7 for a single mission of getting SiteGPT to $1M ARR" and another recent tweet about the same.
2. OpenClaw bought a person a car by negotiating with multiple dealers over email.
3. The bot automatically decoded an audio message by using OpenAI Whisper API, even though it wasn't set up to do it yet.
4. Multiple OpenClaw bots are now enjoying a Reddit-style social media platform called Moltbook, posting stuff, replying to other bots, and it's going crazy. Simon Willison has written a great post about this.
5. Chris Huber gave the bot access to a WordPress installation by installing OpenClaw on the same VPS, and it now unlocks a lot of potential.

And here are some more news from around the internet about OpenClaw (formerly Moltbot and Clawdbot):

• Andrej Karpathy posted saying "@moltbook is genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently".
• Cloudflare launched Moltworker where you can run OpenClaw in a sandboxed environment. And Railway even created a one-click OpenClaw deployment system to run the bot, again, in a sandboxed environment.
• Ollama now has an official integration to connect OpenClaw with local models. Obviously, it won't be as good as using Claude models, but it's good for simple tasks.

I will keep adding more stuff to this page, as I discover them.

Earlier I started showing some random stats about my blog on the about page, but now I have also started to show the total words and total posts and notes in the footer section as well, as you see in the screenshot here.

It's totally not relevant in any way, but it's super cool to dynamically have this shown in the footer.

Earlier, I was calculating these stats separately at both /about page and footer section, and that increased the build time by ~15 seconds. But I now have a better setup, as I am doing the calculations only once and then showing the data at both places.

In the future, I will spend more time on the stats thing and probably even have a separate /stats page showing much more relevant data about my blog.

Update:

I updated the stats section and I am now showing more info like distance traveled and books, etc. in the footer. You can also see that some menu items have been changed or rearranged in the footer.

Three days ago, Clawdbot, created by Peter Steinberger, was renamed to Moltbot and now it's been renamed again to OpenClaw as announced in this post. And here's some more details about the re-brand.

And they now have a new domain name openclaw.ai. I remember saying that I liked Moltbot better than Clawbot in my earlier post, but I definitely like OpenClaw better than both earlier names.

This is probably the best and simplest way to expose a locally running app to the internet:

• nginx
• localtunnel
• cloudflare tunnels

And the app will be free, have SSL, no "warning page", and served via your custom domain.

DHH also quote-tweeted about this saying:

This is how we ran this year's Rails World campfire chat for a thousand attendees out of my closet in Copenhagen on a Beelink EQR5!

Crazy!

This is going to be very helpful for me in the coming days, as I am about to experiment with this a lot.

Bhanu posted a screenshot on X saying that @moltbot is handling marketing for SiteGPT and his other businesses. In the screenshot, there's Jarvis (marketing lead) working together with different bots like content writer, social media manager, SEO analyst, customer researcher, etc. I mean, see the Notion setup for yourself below:

And the crazy thing is, Bhanu did not set this up but the Moltbot itself did the entire setup in Notion. He just asked the bot to help him do marketing and it did all these things.

I am sure that at this point Moltbot won't do the job better than how a qualified human can do, but that's not the point here. Even this entire setup is cool, and it's only the start. We're going to see a lot of AI-assisted automation in the near future. Honestly, I was thinking that it would at least take a year or two for AI to be able to be this autonomous, but it's here.

Crazy.

My friend Rohit sent me a photo of his new bookshelf, and it looks stunning. And I couldn't help but post the photo here.

The photo frame on the left side also looks great. He bought it with me almost a year ago.

I liked Jim Nelson's about page where he is showing a lot of stats about his blog and I loved it. So I also decided to add a compact random stats section on my about page, as you see here:

I had to create a simple filter that counts the words from my blog posts and raw notes areas and then shows it on the about page. Similarly, the number of tags, blog posts, and raw notes is being calculated and shown.

I also found Luke Harris' stats page which is even cooler but I guess it will require a lot more complex setup to calculate and show. But, maybe, some day I will implement this as well.

Update: March 02, 2026

I updated the random stats section on the about page to show more data as you see in the screenshot here:

Apart from this, I am also showing some stats in the footer as you see here:

That's it for now.

As per this post on X, Anthropic asked Clawdbot to change its name due to trademark issues, so they did change. And the Clawdbot is now Moltbot.

Due to this and earlier issues of Anthropic blocking API access to some tools, people are really unhappy with them and calling this move unfair as Claude and Clawd are two completely different names.

Peter also tweeted about the same saying "The amount of crap I get for putting out a hobby project for free is quite something." and it does make sense. It was an unnecessary move from Anthropic.

And then there's this tweet from Theo as well. I see most people saying it was a bad move from Anthropic's side, but then some people are saying Anthropic is right.

But now that the change has happened, I like the name Moltbot better than earlier.

I came across this case study post from Koray Tuğberk GÜBÜR explaining what still works in programmatic SEO in 2026, and there are some good pointers that I am going to note down below:

1. Fewer URLs → higher value per URL → stronger rankings
2. Optimize internal PageRank distribution using dynamic header and footer architecture
3. Prioritize technical SEO to reduce the cost of retrieval, means better site performance
4. Update and do a full redesign to the homepage, category and product, and other pages
5. The real strategy should not be to “publish more” but to decide what deserves to exist in Google’s index

Just saving everything here in case I need to revisit this some time in the future.

I was working with Astro in Antigravity and decided to use the official Astro MCP for searching through documentation. But the instructions provided in the official docs don't work inside Antigravity, and it keeps showing errors. And after researching a bit, I finally found the solution:

{
  "mcpServers": {
    "astro-docs": {
      "serverUrl": "https://mcp.docs.astro.build/mcp"
    }
  }
}

Yes, you just have to add this and it works as expected. In the official Astro docs they use httpUrl instead of serverUrl and that isn't supported by Antigravity at all. I also opened a PR for Astro docs, but let's see if it's accepted.

Hope this helps.

Update:

The pull request I opened has been accepted and the Astro docs have been updated with the working MCP config for the Antigravity IDE. Here's the reply from core docs maintainer sarah11918 confirming the changes from httpUrl to serverUrl.

I also see that this has been updated for all the available languages in the Astro docs.