How I use AI

People have a lot of different, big feelings about AI.

For my part, I’m really excited because I can use it to make more things. For example, I can build personal software projects for my own use in the space of a few weeks. In the past, it would’ve taken me months and that cost would’ve been too high for a lot of ideas, sometimes leaving me paralyzed to decide what was actually worth all that time.

I’m writing this for friends who might be curious why I like AI or folks who already use it and are looking for ideas on how to use AI differently or better. This also serves as a bit of a record for myself to look back on how things are changing because it interests me as a UX professional how workflows, UI, and tooling changes what we choose and are able to do.

I pay for a chatGPT subscription and a Cursor subscription, about $20 each per month. I haven’t been hitting limits and I find it very worthwhile. I use chatGPT on the web and on my phone. I’ve dabbled a bit with Claude on the web and phone but it’s been a while. I use my openAI API endpoints for chatGPT and other models in a couple apps I’ve developed - so far my usage of these API has been so low as to be basically free. I also use a few CLIs - codex, cursor, pi agent, and claude. More about that and sandboxing below.

I’m going to just use the term “AI” as a catch-all for LLMs, image models, text-to-speech models, and all the complex systems that are integrating these models.

Search

AI (LLMs, chatGPT, Claude, Perplexity) is great for searching the internet for things that don’t search well in Google. This is either things that require a whole sentence to describe (“that video from when I was a kid where they flew around with a lion”) or topics full of crappy blog posts with optimized SEO to capturing ad revenue.

• Medical info - this doesn’t replace a GP but it gives a better overview and answers more complex questions (such as alternative possibilities) than the SEO-optimized medical information websites. I like that I can ask it to dive into details like the biology or physical mechanisms behind a symptom. Once it offered me “would you like to know why snot is green sometimes?” Yes! I do!
• Travel recommendations - the search results are full of surface level overview blog posts from travel influencers and I’d prefer the Lonely Planet style overview of the basics and then tailored descriptions of “here’s an odd thing” or “how this compares to a place you’re familiar with”.
• How it works, How its made - general questions about how or why something is the way it is.
• Project Research - For example, I wanted to know more about immigration statistics in the Netherlands so I asked what data sources are available and what they contain.
• News summaries - I avoid reading all news other than the Amsterdam city weekly newsletter (which is about what the city is doing, not the news in the city). Once in a while, someone will mention something I feel like I need to get a sense of so I ask Perplexity “what’s been happening with X in the last month or two” to get a summary. I only want the high level, not the day-by-day handwringing. This protects me from the news cycle and I can actively decide whether I need to know more or not.

Art

I want to create objects to fill my life with things I want to see. I don’t consider myself an artist, as an identity, so I’m not creating things to represent myself to the world. I’m not creating to sell things.

Dabbled briefly with Midjourney but I’m not very interested in creating detailed, complex artwork.

I’ve been happy using chatGPT to generate icons for my projects and also designs that I want to use for t-shirts and stickers of birds being silly. I gave it a reference style image and it has done a good job following that style.

Honestly, I wouldn’t have sat down to draw these things. Instead of another project disappearing down the list of things I want to do, I’ve got a prototype, something to motivate me to explore further.

Travel

I ask ChatGPT to look for upcoming events in an area I will be visiting, places that might be of interest specifically to me (off-the-beaten path museums, street art, places that sell locally designed stickers or zines, factory tours), comparisons of flights vs trains, alternative flight routes, local events like conferences that might affect flight and hotel prices, etc.

On a recent trip (October 2026), I was using chatGPT to plan the museums I wanted to visit for a week in NYC and the itinerary for a campervan trip in the Dolomites. The threads get really long and information that I want to hang on to gets spread out through the thread – a place I liked earlier in the convo, opening hours later – mixed in with information that I want to discard. I found it useful to ask for the information to be put into an html file (a markdown file is nice too). The chatGPT web interface can open the html file next to the thread so I don’t have to mess with downloading it and opening it in my browser. This way, I could ask for tables of information (e.g. opening hours for each museum over the days I’m in town) and even interactive day planners to place the activities I’m considering. They could improve a lot but they’re promising enough that I had chatGPT create prompts for creating them (heavily edited by myself) to create a skill for trip-planning.

You can browse the artifacts:

• Trip planning skill
• NYC museums
• Dolomites trip

In the future, I’m interested to see if AI will get better access to flight data so it can give me better information for choosing flight routes and dates than the simple date pickers the aggregators have now. For example, one time Kevin and I were headed from Seattle to the UK and were flexible on where we could go. He found a $300 one way flight to Manchester, half the price of the others, because it was a special flight to deliver a plane to an airline (passengers were a bonus). I want AI to be able to find those things for me.

Language Learning

I use chatGPT for help learning Dutch in a couple ways. Most of the time, I’m relying on Google Translate or other online dictionaries. When I’m curious to see a set of examples about how a phrase might be used, chatGPT is pretty good at that. I’ve also had chatGPT create a couple practice tests for me while I’m in class based on slides from class and a previous test that I scanned in.

I’ve also tried making my own iOS app to improve on the realtime chat experience of chatGPT. You can read about that in my post about Dutch Buddy.

Data Collection

I’ve been using Grokbot a little bit, mainly to test it out the workflow of persistent background tasks.

I have it running two tasks:

• job search
• weather

I keep daily notes so I like having Grokbot looking up the weather at the end of the day so I have the local high and low.

The job search task is to gather job listings that fit a couple different sets of criteria so that I can review them and mark which ones I’m interested in. This will also give me a dataset of salaries and an idea of how the market might be changing over time for the specific roles I’m interested in.

I barely spend time with either of these. So far they do their thing but I haven’t found a use case of high value for myself.

Code

In early 2024, I began trying out Copilot for autocomplete in IntellijIDEA but half the time it wasn’t good as the Intellisense that it was replacing.

At the time, I dreamed that I’d be able to upload images of sketches of UI and AI could code them. It’s come true!

In early 2025, I loved Cursor for giving me a good diff view of its changes, similar to a diffing tool. For the first time, it was easy and reliable to ask an LLM to edit code in a repository using the repository (selectively, at first) for context. This enabled me to throw together full interactive prototypes of new products within a couple days, including mocked data.

For context, I have been funemployed this year so the following is a description of how I use AI to code on my own projects. I’m not using the institutional tools that spin up agents in the cloud like my friends are at work.

2026 has been an exciting year for playing with Codex, Claude, Cursor, and Pi Agent. I like being able to let an agent loose on my codebases but I find the sandboxing of Cursor and other tools lacking and I don’t have much confidence in them. I’m not worried about a malicious AI agent stealing my data but I am aware of the mistakes software teams make and it’s easy to imagine that the AI can read all my local files which means they are sent and stored on a company’s servers where they can accidentally be kept or not properly anonymized, or sold for ad profiles.

After some fits and starts, I now use Kevin’s vibe sandboxing tool to contain coding agents and allow them access to my selected code repositories. Within this sandbox, I have tried out Claude (but stopped because I don’t want to pay for it), Codex, Cursor, and Pi Agent CLIs. I’ve more fully switched to using CLI to chat with agents about code projects.

When I work with an agent on a code project, I start by asking it to make a plan and putting that in a markdown file in a /docs folder of my repository. The readme provides an outline of the work and links to these plan files. I review the plan, which includes a list of tasks divided into “git commit sized chunks”, each numbered. Then I can start a new session to clear the context and ask the agent to implement the next chunk on the list, marking it done in the file so the current status is known. I still like to skim over the changes (using Cursor, so the VS code git diffing tool) and make the commit myself. I’ve also tried out a workflow of giving the agent access to the github-cli tool so it can create PRs in a github repo that I can then review and merge. I find the github PR process easier to review in the web interface than git log but I haven’t revisited that process because I don’t feel I need that level of reviewable history.

I’ve played around a little bit with agentView and it’s an excellent tool for reviewing agent sessions but I haven’t been inclined to review the tool usage, token usage, and context size of my sessions. I think it would be a really useful tool, in addition to asking an agent to review and summarize sessions, to learn how to make use of agents more efficiently.

I like Pi Agent a lot because its customizable with plugins, self-documented, and model-independent. I’ve played around with customizing it to limit its tool usage and attaching it to a local model so that I would feel comfortable using it to analyze my bank records. That project is paused but promising because I’d like to do something similar for health records and my personal Obsidian vault. That is all data that I definitely do not want being sent to a company as part of the context for an AI chat thread, regardless of their stated policies for privacy and training.

Things I hope to improve

I’d like a better tool for reviewing diffs - I care a lot about being able to see and understand the changes an agent is making. This is one reason I like working in text files, not just chat threads. The thread is ephemeral and messy, the text file is the saved record which should be versioned. I want to be able to edit the diffs - this is the downside of a lot of git diffing tools. By far, the IntellijIDEA plugin for git has been my favorite interface for reviewing and editing diffs because it’s easy to merge conflicts, stage single lines, and review a single file’s history.

I’d like to be able to review diffs of markdown files in their preview mode.

I’d like to be able to work with an agent on a formatted document. I’d like to be able to put it in suggestion mode so it could only suggest changes so that I could be certain what was changed and I could approve/deny/edit those changes. I’d like to be able to highlight and comment on the text to ask clarifying questions or instruct larger changes. I like the Google Docs interface for all this but I want the underlying file format to be markdown and something I can work on locally with my agent. My use case for this is developing research documents. For example, when I want to learn about a topic like AFib, I want to start with a summary from the agent, ask more questions, remove the bits I don’t care about, and then I will have a file I can save in my Obsidian vault that I can use as reference or edit in the future.

Open Questions

• Will IntellijIDEA come up with a better integration of AI into my favorite IDE? I want to check out their Air plugin.
• Would it benefit me to work with cloud agents?