This is a weekly diary of my own AI use, mostly through Claude Code. The prompts and chat logs are automatically scraped from my devices but I edit and review before publishing.
These are not all of my weekly activities (I filter out any with client info or anything personal) but hopefully you'll get a sense of how I'm using the technology.
Prompts are lightly edited for clarity and security, and truncated if they're very long. Reflections are written by me at the end of each week and automatically pushed to this page.
Prompt:
combine your knowlesge of google search console, wordpress mcp, linkedin (recent posts on the dash and stats in fhe systems assets folder) and kit and give me an end of august report across socials blog and audience.
What I was trying to achieve: I wanted to consolidate my audience metrics across all platforms and figure out where conversion actually happens.
What happened: I built a unified report across Google Search Console, WordPress, LinkedIn, and Kit, discovering that LinkedIn drives reputation, while email is where real engagement converts.
Reflection: Running a small business that is basically built on blog posts means sometimes having to do things like reading audience stats. I'm not a huge fan, so I've built this thing which works across my three main sources of data: LinkedIn, mailing list, and WordPress.
Prompt:
im trying to access something at http://100.92.132.105:8800/ over my tailnet and i can see it on my phone but not on this macbook
What I was trying to achieve: I wanted to debug why my tailnet access worked on my phone but not my MacBook.
What happened: I discovered a single-digit typo in the IP address (100.92 vs 100.90); once corrected, access worked immediately on the MacBook.
Reflection: oops my bad.
Prompt:
gather up some articles on the openai/hugging face attack, the claude "hacks", the meta attacks, anything about jailbroken chinese open source cyberthreat LLMs from the past few weeks, as well as my own article claude hacked ny website but it didnt go rogue. create a new folder in writing/blog posts and put downloaded copies of articles in there...
What I was trying to achieve: I wanted to research recent LLM security incidents for a draft article exploring whether 'rogue AI' is a real concern.
What happened: I gathered articles on OpenAI, HuggingFace, Meta, and Chinese open-source LLM attacks, then verbally drafted a ~3,000-word WordPress post with evidence and structured argument.
Reflection: I don't like AI for writing (or any human-to-human comms), but I'll happily outsource research. I like this approach: claude code finds webpages and downloads local copies to my shared folder, which I can then access on the move. Internet is sketchy on the farm, but with this approach I can create a fast, local repository of dozens of articles I can read at my leisure.
Prompt:
i need to update the banner image on my homepage that has the whonive worked with logos. do you see the one i mean?
What I was trying to achieve: I wanted to update my homepage banner with client logos and verify nothing broke elsewhere.
What happened: I updated the banner and verified CSS was properly scoped—the blog received only inert custom properties, so additional CSS didn't overwrite anything.
Reflection: Over a few prompts I went from a static, grayscale, outdated banner to a full colour one with motion and updated clients. You can see it on the homepage at https://leonfurze.com/
Prompt:
have a think about this one. using what you know abojt me and my capabilities and area of expertise, i want a study program into more advanced AI territory. i dont have a technical background so i appreciate it might mean learning some fundamentals in ML, maths, coding etc... people like gary marcus and yann lecun agree (and they often DISagree) that LLMs arent the future, but what is, and how do i position myself ahead of the curve? i started a PhD in LLMs months before ChatGPT arrived. this is partly what i mean. This might go into projects/personal/future-study. im only interested in valid pathwa […]
What I was trying to achieve: I wanted to build an advanced AI study program beyond LLMs, informed by experts who rarely agree but both distrust LLM-focused paths.
What happened: I curated 26 resources (9 books including Deisenroth, ISLP, and Sutton & Barto; 17 papers) covering fundamentals and advanced topics, and uploaded them to my Kindle library.
Reflection: The most interesting part of this was how claude code ended up tackling the kindle part. Files were too big to send via email, so it instead used the send to kindle website in google chrome. I've got a fair bit of reading to do now...
Prompt:
(no prompt captured)
What I was trying to achieve: I wanted to create a header graphic for the rogue AI blog article.
What happened: I located the Pexels padlock image from the published post and generated a 2240×1260 header graphic for it.
Reflection: this one just works... it's the skill I use to create blog header images with the semi-transparent black layer and white text. I use these on socials.