Recap from last week
I was working on building a reader for Agora (i.e. a place you can read books - there are two to start with).
Previous prototypes relied on people reading the book outside the app (physical book, kindle etc), and some were fine with this. Most however, were confused that an app relating to books didn’t have the book in the app. Over 50% of people I spoke to said this was the reason for not activating after sign up.
For some reason, I’ve been anti-reading on my phone but I’m coming around to it. I think it was the ease of distraction, but as long as the words go in, it’s good for you. Plus having a reader in the app allows for a much more integrated experience in the long-term.
There are some great places on the internet that allow you to download classic books for free (standardebooks.org was my tipple), and then it’s mainly a matter of styling and tracking a reader’s progress. Not that simple, but in theory it is.
One other thing I learned: Claude Code really, really drains your battery.
Tools and tricks
Brainstorm and design your plan like your life depended on it: For any new feature build, the most important part is planning before writing any code. I’ve been using the brainstorming AI sub-agent to help me with this and have found its level of questioning pretty impressive. I’ll often spend hours going through this process and have found the coding that follows so much easier.
5 x ‘yes’ response is a cause for concern: A new mental model I have when using AI: if you just hit enter / say yes to it 5 times in a row, you need to go back and engage with it.
Nothing worse than when your AI is off to the races completing tasks you didn’t even know you wanted. You must remain in charge of it. So if I’m just hitting yes to what it’s saying, I stop and ask it questions. More often than not, you’ll help it find an error in its thinking and fix it up before your codebase becomes a monolith.
Question anything an AI builds. I was trying to work out exactly how a pixel offset function worked and when I was going through it, I realised that the code written was based on pixels on the screen rather than a percent of pixels (pixel offset is how your device can send you back to exactly where you left off a previous session). Without this understanding, the reader I was building wouldn’t have worked across devices.
A nice code review prompt I’ve found for Claude: Please dispatch two subagents to carefully review phase 5. Tell them that they're competing with another agent. Make sure they look at both architecture and implementation. Tell them that whomever finds more issues gets promoted.
The competition seems to get the best out of them both.
What have I read / watched
Another good mental model for https://sundaylettersfromsam.substack.com/p/is-ai-a-partner-a-tool-or-a-servant
This confirmed what I was thinking about 5 x yes: https://simonwillison.net/2025/Oct/22/living-dangerously-with-claude
A good video on a real life developer using AI: https://www.youtube.com/watch?v=dHIppEqwi0g
Next up
Building the social discussion page. By the end of this week, you should be able to leave comments on a chapter and see others’ comments.
Nick