Sunny

← All weeks

What I’ve been reading (W11, 2025)

Coding with AI and the limits of vibe coding

  1. How to use LLMs/AI for writing code https://simonwillison.net/2025/Mar/11/using-llms-for-code
    • Set realistic expectations: LLMs are advanced autocomplete tools that are helpful but imperfect.
    • Check training cut-off dates: Ensure their knowledge aligns with current library versions.
    • Provide clear context: Detailed prompts help guide them more effectively.
    • Ask for options: Explore multiple solutions that LLMs can provide.
    • Be specific: Give precise instructions to generate the desired code.
    • Always test: Run and verify the code for accuracy.
    • Iterate: Refine your prompts based on the responses you receive.
    • Use code-running tools: Test generated code quickly and efficiently.
    • Experiment: Learning through LLMs strengthens your coding skills.
    • Know when to step in: Adjust or modify the code manually.
    • Enjoy the speed: LLMs can accelerate the prototyping process.
    • Leverage your expertise: Your knowledge enhances the effectiveness of LLMs.
  2. Vibe coding makes buggy software—it struggles with messy situations like handling multiple coupons at checkout or what happens if someone deletes a shared folder. Users get frustrated when the software makes wrong calls on refunds, security, or privacy issues. But it’ll get better, just like how AI-generated images went from weird blobs to photo-realistic art in just a few years. Soon, vibe coding could handle tricky logic easily—like instantly deciding refund rules or sorting out privacy settings—making coding feel like having an entire engineering team ready to go. And even if it stays rough, casual and experimental software might dominate, similar to how Instagram photos or TikTok videos aren’t professionally produced but still wildly popular https://andrewchen.substack.com/p/predictionsthoughts-on-vibe-coding
  3. Good design exposes the structure of how things work and lets users figure out how to apply it to new situations like numbered gears on a bicycle or simple time/power buttons on a microwave. On the other hand, bad design hides this structure with unnecessary labels, making it harder for users to understand and form a mental model of how the system works, like when products overcomplicate things with labeled modes instead of trusting users to figure it out. https://www.geoffreylitt.com/2025/03/03/the-nightmare-bicycle.html
  4. Tips from the Gmail founder to build better products: The right approach to new products is focusing on just three core features and getting those very right, then forgetting everything else. The iPod, for example, was great because it was: 1) small enough to fit in your pocket, 2) had enough storage for hours of music, and 3) synced easily with your Mac. Similarly, Gmail focused on speed, storage, and a simple interface at launch. If your product needs “everything” to be good, it’s probably not innovative. Focus on creating something simple and compelling. The iPad’s lack of traditional features like file managers may seem like flaws, but it’s part of its appeal—it’s a fast, simple device that’s easy to use. The value of the iPad is in how people will use it and efficiently, whether for browsing, photo sharing, or collaboration https://www.founderstribune.org/p/if-your-product-is-great-it-doesn-t-need-to-be-good-by-paul-buchheit
  5. If you send many messages or emails, make them easy to read and act on. Use clear subject lines, add labels like “Action needed” or “Just FYI,” and keep paragraphs short so people can quickly scan and respond. Stick to one topic per email, summarise when following up, and avoid unnecessary “Thanks” replies. Before asking for help, try solving it yourself and mention what you’ve already tried. When emailing busy people, skip the long intro and get straight to the point—concise emails get faster replies https://www.43folders.com/2005/09/19/writing-sensible-email-messages