Since reading David’s article, “the solution might be cancelling my AI subscription”, I cannot stop thinking about how different our experiences are.

Six months ago, right before subscribing to Anthropic’s Claude AI, none of the following custom-built apps existed: a useful, personally fitting bookmark manager; a purpose-built RSS reader that works hand in hand with my bookmark manager; and a simplified task manager that also works in conjunction with the other two. As an experienced user of many RSS readers, bookmark managers, and task managers, knowing my friction points with these apps, I have always dreamed of what would be perfect-for-me versions of them. So I built those, one by one.

Thanks to Claude Code and my vision for what would make the perfect versions of each of these apps, I could build them without being constrained by not knowing Next.js, TypeScript, CSS, etc. I don’t plan to make commercial versions of these nor open-source them. There are a few, and I can tweak them as I see fit and as my needs evolve; this is where I’ll keep focusing. AI empowered me and will continue to do.

You see, the solution for me is to keep iterating and keep my subscription. Two different stories, two different outcomes. Maybe there is something that I didn’t catch in David’s experience.

I don’t remember the last time I built a full automation workflow from the ground up in n8n, thanks to Claude AI and MCP support. I started manually before knowing n8n had MCP support integrated, which makes me feel more competent in understanding what’s going on. It reminds me of when I bought an iPad touch so that I could learn to build apps with Objective-C and Xcode; I like to understand what’s going on under the hood, in the digital world, at least.

I just completed a new workflow: automatically saving new bookmarks stored on Micro.blog to my custom-built bookmark manager web app running on Vercel. Since Micro.blog doesn’t support webhook calls, I had to resort to a scheduled n8n workflow that pulls any newly saved bookmarks via Micro.blog APIs and saves each one using a new API route on my bookmark manager web app on Vercel. It’s much more efficient than asking Claude AI to do this using a skill (which was working perfectly, by the way) to save into a Craft Collection block entry.

Report: Apple Plans to Make On-Device AI a Key WWDC Focus:

The arrangement represents a noticeable departure from Apple’s original Apple Intelligence announcement, in which the company said all cloud-bound queries would be handled exclusively by its own Private Cloud Compute infrastructure running on Apple silicon. Apple is likely to retain the Private Cloud Compute branding despite the change, people familiar with the partnership told The Information.

There is much left unanswered in this article. How much of Private Cloud Compute is in use today and for what purpose? How much of this capacity will be running the new Apple Intelligence? Will Apple expand PCC into Google’s datacenters? If so, what is Apple’s own infrastructure going to serve? And how much of an improvement can we expect to run local models on our devices compared to the original Apple Intelligence model?

It took me the drive, Codex and Claude Code, four days, and voilà! Things 3 is now a thing of the past; I’m using a perfect-for-me custom-built task manager web application running on Vercel/Neon. Things 4 might go subscription-only; I don’t care.

After largely completing the transition from Things.app to a version built with vibe coding on Next.js and hosted on Vercel, I decided to switch from Airtable to a Postgres backend. The free-tier on Airtable only allows 1000 public API calls, which wasn’t sufficient. To avoid hitting this cap constantly, I, with the help of Codex, migrated to a Postgres database called Neon, available through Vercel’s marketplace. I already use Neon for my bookmarking web app, so the change was quite smooth. I didn’t need to move any data since I am still in the late development phase. Now returning to other fine-tuning tasks.

Well, well, well, it seems to be happening much quicker than originally thought. I’m about 70% done with this already! I still like Claude Code more, but Codex is more efficient at testing the UI in its browser than Claude Code. And the in-browser cursor, which is 100% independant than the Mac cursor, is super cool to see in action.

So far so good with OpenAI Codex. I prefer Claude Code’s look and feel, but Codex seems more like something Microsoft would build: plain, without soul. My data model is completed and implemented in AirTable. The basic web UI is running, but is lacking many basic operations like create, update, and delete for tasks, projects, etc. It took me less than two hours to deplete my token allotment. It’s a slower rate than with Claude Code… but still, it’s quickly gone.

Yesterday, I was criticized for adding an AI-powered feature to my Microblog frontend. It was suggested that, since AI can generate blog post titles, I might appear lazy. I explained that I’m not a native English speaker and sometimes need help with writing. Is that really wrong? For those using spell checkers and grammar checkers, are they lazy, too? At what point do the tools we choose to develop and utilize indicate our laziness?