AI

Claude Code (not Claude Design) is really good at helping you choose a design and theme. I embded the “grill-me” skill in my AGENTS.md and told it what I wanted. I asked it to try a regular theme and a playful Duolingo-like theme.

This is what it came up with, after asking me a few questions:

I think the winner is clear: The “Duolingo” one is unique and buttons feel super pressable. The hybrid one doesn’t have much going on and is too forgettable. The Retro one is kinda cool, I guess, but it has the AI-piss filter.

I found my previous essay tool that I built over 5 years back.

IMO, the “repeated words” tool is the most useful, especially because it groups words like “organize” and “organized” as the same word.

The readability score would probably be more useful if done by paragraph rather.

A few years back, I had an idea to make a college essay analyzer. It would help you find duplicate words, read the essay out to you, help you find synonyms, etc. This was back in 2019, pre generative AI.

Out of all these tools, I think the repeated word finder would still be useful. It should not only work for exact repeated words, but also similar forms of the same word, like “thing” and “things” or “organize” and “organizing”.

What if we take LLM wrappers to the next level and skip the LLM API entirely? Your app is just two things:

• a website with a text box
• a publicly accessible docs link that the LLMs can read

The docs should explain what the LLM is supposed to do with what the user provides. The “message send” button just opens ChatGPT/Claude with a prefilled prompt.

I think this can be cool for smaller projects: no need to get an API key, worry about rate limiting, set up billing. But of course, projects like these will be more than a prompt that you are packaging in a slightly easier-to-use manner.

I really wish we could “sign in with ChatGPT/Claude” rather than having to pay for our users’ API usage. We kinda can do that if we make an open source project using the Codex/Claude agents SDKs, but it’s difficult for non-technical people to use those.

I’ve been trying to make this mini version of Hyperlocal agent and test friendly. One way is by making all storage and state live in the URL.

I just asked Claude Code to “test Hyperlocal as a non-technical user … make a Boston day trip plan”. The best thing was that it just shared the URL of the map it made, so it was easy to debug.

And of course, building for agents means you’re building for humans. Good DX for agents = good DX for humans.

I recently found out about Pi, and immediately tried to use it to make a “personal assistant” with memory.

It was a lot of fun: I made it monitor GitHub issues and work on tasks I gave it. I also gave it a phone number from agentphone.to so it could call me. I was able to call it, tell it to make edits to itself (for example, add a memory layer with markdown files), and have it call me back when done.

Only issue, Pi does not have a native built in solution for web search like Claude and Codex.