I get asked “what are you using to make this stuff?” often enough to make it a post, so here goes. This is How I AI in 2026. I think the tech bros say “what’s your stack?”

My general principle is inspired by Ghostbusters – don’t cross the streams. Some people are agent power users; I prefer to know what each workspace is for, have control of what I want, and organise a lot of thinking by hand.

Sites & Apps, Bits & Bobs

The site (https://ifyouuseme.ai/) is managed and updated with Lovable. This was recommended to me by Kathleen Naglee and I’m thankful. I pay about £22 a month for this and it is very useful. Through Lovable’s connection with IONOS, I registered the domain and manage hosting, which is about £100/year for the ifyouuseme.ai address. I find Lovable really easy to work with, update, host and test ideas with public sharing. I find it a useful place to spin up and test other projects too.

This blog and i-Biology are still hosted on WordPress.com. After 19 years, I just can’t let them go, but beyond hosting and domains, I don’t pay for any AI or other features.

I use Claude a lot and pay about £90 a month. I know I could do much of this work in Lovable as well, but my personal preference is to do use Claude as my workhorse and I’ve cut most other paid subscriptions. I set up Projects with context, memory and rules, and work a lot with .md files. I learned a lot about Claude last year, some in this post.

I also learned that for most of my app-like projects (loads are buried in the site), forcing it to work in HTML works for me. It reduces dependencies on other types of code, streamlines my thinking, I can read it and turns out really lightweight files. It might sacrifice a little bit of flashiness, but I don’t need that. For example, the Environmental Footprint Estimator is only a 47kb file. Another complex visualisation of competencies was under 1MB. The entire IB Biology review, statbook and mind-map was about 3MB.

When I’m happy with the product, I export it as HTML and load it up in Lovable. It is all backed up in GitHub, but I’m very time-poor and can’t be bothered with extra steps. It also helps me partition my thinking: do the work in the background (Claude) and then send the product to public (Lovable). If I update the app, I just re-send the HTML and publish. If needed, I can send a file to someone by email or Teams and they can try it, locally, with no risks.

I don’t use agents or give anything too much access to my computer. I do have some folders connected to Claude CoWork, but they are safely in their own space. Maybe I’m too cautious, but I am a pathological worrier, so this is good enough for me.

Sub-Editing & Research-Checking

I love writing and hate handing it over to AI.

Having said that, I will ask Claude to act as a sub-editor on longer pieces. For example, the book had about 500 references and took months of evenings, weekends and breaks. Within this time, I wanted to rearrange sections, update sources and just rethink some bits. I use a version of CSE superscripted numbers as it tracks well for a reader, but when parts got changed, the citation-reference order went out of whack. Claude was very helpful in renumbering and double-checking. It was also useful to give feedback on consistency of tone, redundancies and general grammary bits.

In one recent project, writing for formal publication, the process started over a year ago and the piece is still not out. Claude was helpful in finding where pre-print papers from the first draft (Dec 2025) now live as actual publications in Aug 2026. This would have taken hours; and of course I checked and re-read the newer versions of the papers.

In another work project, which took a tonne of time and effort over 60 pages and a couple of weeks, I put a “DRAFT” watermark on the Word file and opened comments for feedback. The next day, the document had crashed and was unrecoverable. Doom. My laptop nearly went through the window; not even our amazing IT team could work it out. Claude found it immediately: a Word for Mac bug had left hundreds of layered watermarks on top of each other. I asked it to strip out all the problem-causing elements and give me the exact same document back to keep working. Problem solved, doom averted.

Search

Perplexity. I’ve been using this since it was a test app. I don’t pay for it any more though. But for a few key searches a day, and with a good prompt, it is good enough. Quick, focused and not filling up my Claude.

Multilingual Videos

I have a lowish-level work account for HeyGen, used very sparingly and for high-importance tasks. For example, this week I used it to do 3-4mins of really clear, information-dense and accurate videos for our Chinese and Korean parents to explain learning at the school, what we’re working on and so on. This plays a few roles: demonstrating updates to the capabilities of AI each year; giving the live translators a break during a long session; adding a bit of novelty to a session; and creating an asset that can be shared with parents who missed the session.

My workflow is still quite involved: updating the video avatar and voice; writing the script carefully, translating and checking the script with experts (particularly important with educational homonyms); creating the video and getting it checked; downloading the captions and burning them in with HandBrake; and then using Canva for branding and export.

I know their are other ways, but I want to check every step. I don’t use video-video translation because of the homonym issue; for example “agency” could be interpreted many ways. We need to get the right characters in place.

When I use video or explain how, I make sure to note the intentionality of use and mention the footprint.

Data Analysis & Dashboards

Here’s an example. Claude is really useful for this, as we collect a lot of data. Always using only anonymised data, it can help cluster responses, visualise themes and do a little bit of qualitative content analysis. Previously this would have taken many hours of cutting and pasting (clustering) and clicky-clicking (Excel), to get a mediocre product. Now I can start from good in a couple of hours and make it better by hand.

School Stuff

We have Canva, Toddle, Flint and MS CoPilot at school. I’ll keep using these to make sure I know what they can do, and anything student-facing is done mostly in Flint. There are many much better users of these tools at WAB than I. We also have some other projects to play with, with special interest groups (e.g. AI cars).

Other Stuff

I play with other tools a lot, checking back in on them to see what they can do. Z.AI, Kimi, DeepSeek, Gemini. I dip back into ChatGPT every now and then, but switching and setting up again, to make the most of it, is too much effort, mentally, to keep. I’ve found workflows that work for me, so other tools are for curiosity and awareness. I have HuggingFace, GitHub and Vercel as well, though these are reserved mostly for the very rare moments I have to just play.

Anyway, that’s about it. I’ve reduced the number of paid subscriptions, but got closer to optimising the tools I do use. I’m always curious but less likely now to hand over a payment than I was in 2023.