One of the frustrating aspects of LLMs is their lack of consistency unless you develop specific skills, which can take time to implement effectively. For example, I wanted to generate documentation for my most recent n8n automation workflow, but Claude was unable to do it, and I can’t remember the prompt that finally made it possible. I should have saved it somewhere for easy retrieval. I’m wasting precious credits. 🤦🏻♂️
Simon Willison on AI-assisted programming:
The more time I spend on AI-assisted programming the less afraid I am for my job, because it turns out building software - especially at the rate it’s now possible to build - still requires enormous skill, experience and depth of understanding.
My Defaults as of 2026-01-10
Changes from the last edition are in bold. ✉️ Mail Client: Fastmail 📨 Mail Server: Fastmail 📝 Notes: Craft + Apple Notes ✅ To-Do: Things 3 📷 iPhone Photo Shooting: Camera.app 📚 Photo Management: Photos.app + Photomator 🗓️ Calendar: Calendar.app 🗄️ Cloud file storage: iCloud 📰 RSS: Reeder connected to Inoreader 📇 Contacts: Contacts 🕸️ Browser: Mobile Safari + ARC Browser on Mac + ChatGPT Atlas 🧠 AI: ChatGPT + Claude AI 🔎 Search: Kagi Search 💬 Chat: iMessage (WhatsApp when abroad) 🔖 Bookmarks: AnyBox 👓 Read It Later: Inoreader 📜 Word Processing: Ulysses, Craft 📊 Spreadsheets: Numbers 🛝 Presentations: Keynote 🛒 Shopping Lists: Reminders 🧑🍳 Meal Planning: None 💰 Budgeting & Personal Finance: Numbers 🗞️ News: La Presse (Apple News for English news) 🎶 Music: Apple Music 🎧 Podcasts: Apple Podcasts 🔐 Password Management: iCloud Keychain & Apple Passwords 👨🏻💻 Blog hosting: Ghost, Micro.blog, Scribbles.page 🌐 Web Services: Cloudflare, Chillidog Hosting, DigitalOcean
Apparently, people are barely using Stack Overflow to ask questions, thanks to LLMs and AI. I expect a similar trend among people in a community like this one on Micro.blog. Some questions would be super easy to answer by asking ChatGPT or the like. I do understand that many people still want this human touch, though.
Something is about to happen. Again. 🤗🫣
On Tahoe Icons
Just finished reading “It’s hard to justify Tahoe icons”, which many UI design pundits and non-UI experts, as well as simple, passionate Mac users, have been referring to a lot recently. I appreciate the documentation effort. It’s really well done. Very convincing. But…
Designers age and are gradually replaced by a younger generation. Whether you like it or not, they bring new beliefs (justified or not), design principles (better or not), and values (questionable or not). Recently, there’s a trend where software appears much less crafted than it once was. Everything seems thrown together, flat. And still…
I like those macOS Tahoe menus with icons (🫣), but yes, there is an absolute lack of consistency.
I can confidently say I’ve learned at least one significant new concept every day over the past two to three weeks. It is not only satisfying but also quite stimulating. This will keep me from aging, that’s for sure.
Claude Code skills are probably the most intriguing aspect of Claude Code and Claude AI. I’m not so sure yet how to take advantage of them. My understanding from this excellent video is that you have to be an expert at something to create those skills.md file.
An Important Lesson
When I started my studies in computer science over 40 years ago, we learned to read functional specifications and then translate them into machine instructions (COBOL, FORTRAN, Pascal, etc.). It was the training of a programmer. I knew that one day I could become the person who writes functional specifications. I didn’t become a programmer, nor did I work in the development world.
Due to my recent experience with Claude AI, Claude Code, and Vercel to create custom applications, I realize that I have become the one who writes functional specifications, but for processing by artificial intelligence. What does this tell me about the profession of a software developer? The need to write specifications remains essential, if not more so, even with powerful tools like AI. I think it’s a valuable lesson.
On OpenRouter.ai
I just finished reading about the service openrouter.ai. I was curious to understand the purpose of this service as well as its business model. I saw several instances of this service being used in n8n workflows. The problem I see with this service is that it makes the consistency of the quality of responses from the requested LLMs even more unpredictable. Each request could be handled by LLMs with different characteristics and performance from one time to another. I’ll pass on this, but I still learned something tonight.