I love using Projects to organise my chats. Aside from the organisational aspects of Projects, it is such a powerful way to interact with ChatGPT because it allows you to add custom instructions and files specifically related to the project. This is very important as it provides context for the model, and context is key when interacting with Large Language Models. Last weekend, I sat here trying to create the perfect custom project instructions and, despite my best efforts, I was not entirely satisfied with the output. I realised I was spending way too much time manually drafting custom instructions for my ChatGPT Projects. I kept running into the same problem: prompt structure. Then the penny dropped: I realised I could just use ChatGPT to help me build this, and so I built a Prompt Generator GPT. I worked iteratively with GPT-5, asking the model to ask me clarifying questions to help me design a prompt which will allow me to build project instructions for a wide range of subjects.

GPT-5 arrived with an enormous amount of expectation attached to it. By the time it launched, OpenAI had already set a precedent with GPT-4, GPT-4o, and the reasoning models that followed. Each release raised the bar, not just technically, but emotionally. As users, we have come to expect clear leaps forward, and GPT-5 seemed to miss the high benchmark mark for the majority of users. From a purely technical standpoint, GPT-5 is undeniably powerful. It improves on reasoning, long-context handling, and multimodal capability. Yet the initial reaction across the AI community was mixed, and in some cases openly critical. That contrast between capability and perception is what makes GPT-5 such an interesting release to reflect on.

Like most people, I spend a lot of time reading, researching, and collecting things from across the web. I am always aware of how easily useful content can disappear or be forgotten over time. Blog posts, documentation, bash scripts, YouTube videos, services I want to explore later, and images which sparked a bout of nostalgia. Over time, I learned that browser bookmarks alone were not enough. I needed a smart, searchable digital scrapbook and that is how I discovered Raindrop.

If there’s one AI tool I reach for every day without fail, it’s Perplexity. It has become my go-to assistant for staying up to date with the latest news and exploring new topics. I wanted to share how I use it, why I keep coming back to it, and a few features I think are genuinely worth exploring. My Two Main Use Cases: News and Learning I use Perplexity primarily in two ways:

In my previous post about Tailscale, I covered what Tailscale is and why it’s become a staple in my homelab toolkit. Today, I want to share a real-world example of just how handy it can be—specifically, how it helped me retrieve personal documents from my homelab while sat on a train, armed with nothing more than my iPhone and a 5G connection. The Situation: Needing Remote Access on the Go We’ve all been there: you need a file, but it’s tucked away on your home network, and you’re nowhere near your desk. For me, this happened while I was on a train, miles away from home. The documents I needed were only accessible from my local network—or so I thought.