How I use AI in 2026

August 4, 2026

When it comes to AI use, I see a lot of commentary about how these tools will either bring about end-times or provide overabundance and ease. There’s not much written about what people are using AI for in their day-to-day lives and the practical use cases they’ve found.

I left the corporate workplace at a super weird time for the tech industry. When I started working on my own in January 2025, most AI tools for programming were pretty nascent, and involved copy-pasting code in and out of ChatGPT. Beyond the UX inefficiencies, the models themselves didn’t read from the Internet and it made working with fast-moving tech changes clunky to use. Mere months later, tools like Claude Code became ubiquitous, and you could suddenly work like a bulldozer, rather than a scalpel, on large codebases.

I use AI tooling every day to build Table Slayer, Counter Slayer, and this site. Given the rapid pace of advancement, I thought it might be useful to document how I use AI at this particular time (August 2026) for both code and personal use. My assumption is that this post will feel just as dated as my previous posts on this subject in another year or two.

The models I use
I use AI to build complex web apps
I use AI to troubleshoot my Linux workstation
I use AI to research my media hobbies
I use AI to learn the mechanics of board games
I use AI to create quick TTRPG encounters
I use AI for tip of my tongue searches
I use AI for copywriting
I don’t use AI to create artwork
I don’t use AI for social needs, and don’t share personal information
What I’m worried about with AI

The models I use

I generally use the latest model from Anthropic available on their max subscription tier. Right now that’s Fable, which, at least for code tasks, seems like a pretty large jump over previous models. Outside some occasional explorations into esoteric models for other tasks, I’m pretty happy in the Claude ecosystem and haven’t found a reason to change.

I use AI to build complex web apps

I primarily use Claude Code in a terminal, jumping to Neovim when I need to review or make manual edits. My projects are open source, and I don’t really hide my AI use in those repos. They all come with claude.md files that give some instructions for how to work within the repo, and I’ve started to build repo-specific skills to form patterns and prime Claude for the overall architecture of the projects. The latter has really helped avoid AI footguns and provides historical context for architecture decisions.

Both “Slayer” sites are high-complexity web apps, and Table Slayer in particular has a multitude of hard problems: “3D” canvases, post-processing, networked multiplayer, and game-level browser performance tuning. I add the usual web architecture niceties: a mono-repo, custom design system, auth system, billing, PR previews off of prod data with GitHub workflows…etc. I try to avoid dependencies, and outside some core platform choices (Svelte, Turso, Partykit, Cloudflare, yJS) I try to have the code live in the repos. I’m pretty happy with the stability and code quality of the apps I build. I feel I prune and rework as much as I did when I was the primary writer of the code.

Despite handing off most code duties, I’ve learned more about web technologies in the last two years than I likely did in the previous ten. Two years ago I was a UX designer who was an expert in CSS, with a healthy amount of experience in React component systems and the usual smattering of web worker skills (git, vim, Linux…etc). In the past two years, with the help of Claude, I now understand quite a bit about the following subjects:

Database modeling, migrations and performance
JavaScript API patterns and architecture
Auth systems and general cryptography patterns
Three JS capabilities and limits
Webhooks, workers, and crons
Websockets and complex state passthru
Docker setups and generalized web hosting
Region-level hosting with embedded replicas
Billing systems and subscriptions

I admittedly came into this period of AI discovery with a lot of prior knowledge (nearly 20 years in the industry with multiple successful startups), but the level of acceleration in my raw skills feels absolutely baffling to me. In eighteen months I’ve graduated into a makeshift full-stack engineer. That I learned this much at age 46 and not 26 sort of blows my mind.

In some ways I feel AI tools like Claude Code are a perfect fit for people like me. I’ve always known what I wanted to build, and always struggled with how to build it. I tend to excel at learning through exploration and reverse-engineering rather than coursework. Claude Code is the perfect tutor for me. It doesn’t judge my dumb questions, and I feel comfortable going off on tangents to understand a subject before applying it.

Learning while I code

A good example of this study was working through a problem I had a few months ago: trying to get Table Slayer’s fog of war syncs to be faster. Previously, the fog of war was a stored black and white image per map that was used as a mask for the 3D “clouds” to mimic. This required a round-trip upload / download from my media server and an update in the DB. Aside from the 2-second lag this caused, I had to add lots of cache busting on the images themselves, because I didn’t want to store a new image every time the fog changed.

I was smart enough to recognize the problem, but I didn’t really know a better way to solve it. I could blob the images and store URIs? But that would be a lot to store in my DB. I started bugging Claude about other options. I’d previously moved to an embedded SQLite database to make my DB operations fast, but the image round trip was always going to be slow on the network.

Claude suggested storing run-length encoding data in the database directly. Essentially, rather than store physical images, or blobs, we could store a shorter string made of the “runs” of white and black pixels. A 6x2 image of 4 black pixels, followed by 2 white ones and 6 more black, can become 4B 2W 6B. Here’s what it looks like in practice.

The fog of war in action, the 1920x1080 mask it generates, and the run-length encoded string that replaces it.

The original mask is 1920x1080 and, despite being a black and white image, would generally be around 120kb in size. When we transfer that information instead to a run-length encoded string, we get something like the example beside it, which is bytes in size. The latter, converted to base64, provides a level of compression that is easy to send over the wire in JSON across Partykit and store directly in my DB for quick retrieval.

“RLE” was a new concept to me. I came into computers organically at home; I didn’t learn this stuff, and I work primarily on my own. I just knew my problem, and was smart enough to know what type of solution wouldn’t work. It suggested plenty of other solutions before coming to this one. I had to be active enough in the decisions to research and validate the concepts proposed. I can tell you that I tried to solve “make Table Slayer faster” several times in multiple PRs before eventually figuring out this specific bottleneck. For me, I was happy to learn about it through this method.

My overall point is that I know what run-length encoding is now. I can guarantee I wouldn’t have known about it without AI exploration, and moreover, I never would have taken on a complicated project like this in the first place if I didn’t have AI tooling to support me. Yes, I use AI to do the grunt work of coding, but I use it primarily every day to accelerate my own learning; the code is just a byproduct.

I don’t let AI make commits

Call me old-fashioned, but I don’t like Claude creating or signing the commits to my repos. Likely, my own pride is in the way here, but I also spend so much time cleaning up its initial ideas and code that at best its a shared credit. I still make branches, make commits, push the files, test things manually and then verify the PRs and review them there. I always look over the code to make sure I understand it, and spend most of my time trying to get AI to follow my own preferences in code-style than anything else.

I’ve read quite a bit about people who are more lenient than me, and let their agents run overnight without breaks. That’s not me. I watch the stream in real-time and stop it occasionally when I see it doing something I don’t like. I don’t give it access to my DBs, don’t let it run SQL and don’t let it generate or run migrations of any kind, even on dev databases.

I have a feeling this is the line of mine that move the most over the next coming years. Eventually the code itself won’t matter and I’ll give in.

I use AI to troubleshoot my Linux workstation

I’ve written before about how “I’m a Linux” and made the transition to Linux workstations many years before AI came around. I’m comfortable in pretty bare-bones systems, and even made a fairly popular getting started video series for folks to help them transition to distros like Arch. That said, I don’t know everything, and often it’s a lot easier to query AI about how certain systems work, or feed it some logs to troubleshoot.

I’m more conservative than most here. I mostly do traditional question-and-answer chats for this and feed it information, or give it targeted access to certain logs while we diagnose. It’s rare that I give AI access to system level files.

I also use AI to manage my Neovim config and Tinty theming. I used to dread major Neovim bumps because it was hard to keep up with the various Treesitter and theme changes. Now it’s fairly easy to talk directly with Claude and tell it the issues I’m seeing, feed it a couple logs, or let it run in headless mode to get Neovim running the way I prefer.

Lastly, I’ve used AI to write dozens of small scripts to convert data between systems or run command line tools. This past year I converted my DVD and music libraries to be accessible from my NAS over Plex. Rather than install bulky tools that would only be used once, I put together small workflows to automate a lot of these tasks. It also helped to create backup solutions and run health checks along the way.

Trying to remember ffmpeg commands

can you give me a one liner for fish on how to convert an mp4 to a gif using ffmpeg

ffmpeg -i input.mp4 -vf 'fps=15,scale=640:-1' -loop 0 output.gif

Both work fine in fish as-is since the filter graph is quoted. Adjust fps and scale to trade size for quality — GIFs balloon fast, so 10-15 fps and 480-640px wide is usually the sweet spot.

I use AI to research my media hobbies

The most surprising usage of AI for me is that I have a pretty long-running chat going about art-house and independent cinema. I feed Claude a lot of my movie likes and dislikes and try to wrestle out recommendations for movies that I wouldn’t have known about. I used to run a few movie websites, so there’s not a lot I’ve missed, and I’m pretty picky about what I watch.

Finding indie flicks I might have missed

Let’s talk about Clockwatchers and how it’s not as well remembered as Office Space. What else am I missing of that era that is female led?

There's a deep bench of female-led 90s indies that got similarly underweighted. Roughly grouped:

The Jill Sprecher follow-up — Thirteen Conversations About One Thing (2001) — She and her sister Karen wrote it; Matthew McConaughey, Alan Arkin, John Turturro, Clea DuVall in interlocking stories about luck and unhappiness. Quieter than Clockwatchers, possibly better. Almost no one talks about it.

Allison Anders' run — Gas Food Lodging (1992) is the key one, Brooke Adams as a single mom raising two daughters (Ione Skye and Fairuza Balk) in a New Mexico truck-stop town. Adapted from a Richard Peck novel. It's about female interiority in a way American film almost never bothers with. Her Mi Vida Loca (1993) is the East LA companion piece, about Latina gang members, also worth seeing. Anders had a real moment and then Hollywood mostly stopped calling.

...

Over the past year I’ve gotten some great recommendations out of AI. Although some of these films saw some manner of award coverage originally, for the most part I lost track of them for one reason or another. Conversations with AI gave me great pointers to check them out.

The Hit (1984)
Leave No Trace (2018)
Wendy and Lucy (2008)
Thunder Road (2018)
The Rider (2017)
Shotgun Stories (2007)
Mikey and Nicky (1976)
Mississippi Grind (2015)

If anything, I found AI’s knowledge of indie cinema pretty awesome. While it’s easy to get impressed by Claude Code blazing away in my repos, it’s just as impressive to have the entire corpus of film criticism at hand for a quick chat. This means that if I want to have a conversation about Pauline Kael and Cassavetes movies in the 70s it’s totally game. This is the closest I ever get to AI filling a void in my social life. Despite my wife being a film minor, and having several friends who are deep into movies, I don’t have anyone to bounce against when talking about micro-niches and eras of the movie industry. Definitely not at 2:30am.

When I was in my twenties I remember using editorial recommendation databases like AllMusic to do this kind of “similar to” and “inspired by” research, but AI tooling has made it immensely faster and more specific. I really like AI for this, and even without all the code tooling I use it for, I would likely pay just for this service alone.

I use AI to learn the mechanics of board games

Over the last year I’ve fallen in deep with board games as my primary recreational hobby. I usually pick up a new game a month, and while part of the fun is learning new game mechanics, I’m not someone who likes poring over rule books. Typically, I learn by watching a few playthroughs on YouTube, then load PDF versions of the rules into a new Claude project so that I can query the rules as I play. This use case is the only one where I actually like to use a microphone to perform most of my queries. I can play the game and just randomly ask questions as I run into issues.

Clarifying rules for Star Trek Captain's Chair

Do i run the action step for the bot after i run my own Captain's actions? Or do i go all the way through the my steps, doing the bot actions after my own clean up phase?

Fully alternating turns — you never interleave. You run all four of your steps start to finish, then the Bot runs its entire turn, then back to you.

The core rules state that play alternates with each player completely resolving all four steps of their turn before passing play to the other player, and the solo rules explicitly say you play the game the same way as the core rules, with the Bot just resolving each step differently. Since you take the Starting Player token during solo setup, you always go first.

So a round looks like:

Your turn: Resupply → Control → Action → Clean-up

Bot's turn: (no Resupply) → Control → Action → Clean-up

I use AI to create quick TTRPG encounters

I DM two games currently. One is a Pathfinder 2E Kingmaker campaign, and the other is a DND Phandelver campaign. It’s a lot of work, and can sometimes be a bear to manage when my players go off the rails. Like my board games, I will often upload the source books for my campaigns into Claude so that I can query specifics (monster stats, NPC names…etc) on the fly while we play. More interesting is that during prep I’ll often build custom encounters for the next game. Typically I already have the encounter figured out and just want stat blocks and some game mechanics to match whatever I come up with.

What AI is most helpful for here is generating printable documents that I can use as reference sheets during my game. Running a game is mostly about improv and I don’t use these sheets literally. Mostly I just have them as a general outline of events that can happen, and then freeform build off them during the game. Below is an example of building a slightly skewed encounter from source material.

I use AI for tip of my tongue searches

As I get older my memory sometimes fails me and I’ll forget the name of a movie or book I read. I can remember plot points or actors, but not the name of the movie itself. This type of searching is fantastic for LLMs, which can usually figure these things out quickly when I give them enough information.

Hunting for a movie title

I’m trying to remember the silent film with the train stunts. I know there are a lot of movies like that, but I’m thinking of the most famous one with the giant wood ties being used at precise times.

That's The General (1926), Buster Keaton's Civil War comedy. The scene you're describing is one of its most famous gags: Keaton, riding on the cowcatcher at the front of the locomotive, holds a railroad tie and uses it to knock another tie off the tracks at exactly the right moment — all done for real, with precise timing and no trickery.

I use AI for copywriting

Although I have no problems using AI to write code, I never use it to write blog posts or emails. Though I spent a lot of time building up skills as a programmer, the output of those skills was always the product itself. With writing, the output is the writing. I’ve similar spent years developing my voice and don’t want that augmented.

I do make tons of errors, and want to avoid them. Typically, I’ll run a blog post like this through AI tooling to fix grammatical errors. Hilariously, like John Scalzi I was always a fan of em dashes and semicolons, and chuckle every time someone mentions them as a “tell”.

I don’t use AI to create artwork

Table Slayer has a lot of illustrations, and I wanted to invoke the feeling of 1980s metal albums with its art. There was never a point where I thought “I’ll just use AI for that”. I knew immediately I wanted to work with an illustrator and would pay them appropriately for their time. Cinnamon Devil did all the fantastic work on the site, and it took months to get everything done. I love her art and even if an AI could create that sort of look I probably wouldn’t use it.

I’ve fooled around with Midjourney and other, local modals. Maybe they’ll improve eventually, but for now it all seems pretty fake.

A couple years ago I remember suggesting to a coworker that it would be interesting to see each others chat logs, to see how we used it differently. They blushed and admitted quite frankly that they used AI to talk through personal quandaries and social interactions.

This woke me up to the idea that people use these tools very differently than me. I think this is why we all feel so differently about AI usage. For me, it’s just an input / output tool for work. For others, it helps them in places a therapist might. I have no judgements here, and part of my own blog posting and journaling is a mechanism to think through my own decisions. Chatbots in a way are no different. It’s the privacy that is concerning.

What I’m worried about with AI

I’m not an expert in global economics or energy consumption. My engineer’s mind mostly looks at AI as a new technology, and an advancement over a long line of tools built by humans (based on the work of other humans) to speed up work. I’m not qualified to talk beyond that, and my evaluations of AI are in the tool output, rather than the byproduct of the tool. Likely you have your own opinions, and this post isn’t meant to change them.

I’m not particularly worried about AI removing the joy of creation and craft. That may be a general, societal problem at large, but it’s not one for me in my personal use. If anything, I’ve gotten more creative and challenged myself more than at any other point in my career. I’m not looking for wholesale invention from these tools, but a work partner who can teach me as I go. Right now, it’s doing its job.

That said, AI has become useful enough to me personally that I’m worried about how it eventually gets limited or polluted. Because LLMs are summarization tools primarily, I’m worried about what source data gets summarized, and how I can tell the difference between something being weighted in one direction or the other. Using a dumb example from earlier in this post, I’m worried what movie recommendations I might get once a movie studio decides to pay to shift the weights in one direction or the other. What happens when a startup pays to have their solutions and skills built directly into these generalized tools?

I’m old enough to remember a more independent Internet, and the consolidation of content that came after. That wave largely destroyed whole industries; some, like journalism, never really recovered, creating a net negative for society. I have the same fears here, except that I’m worried that as we move more and more to summarized LLM content, the actual source information that powers it will get harder to get a hold of. Kept clean, these tools are amazing, but there’s plenty of historical precedent to understand that data pollution because of politics or monetary policy is coming.

I’m also mature enough, and have enough background education, to understand the pitfalls of these tools and what can happen if you use them to build things, rather than learn things as you go. Talk five minutes with any school teacher and you’ll understand the problem. It’s not the cheating that they’re worried about, it’s the abandonment of learning. How can you understand the importance of an education if you don’t build up enough personal brainpower to recognize its usefulness?

Privacy is a big problem I can’t avoid though. By using these tools I’m volunteering quite a bit about myself. Even when Claude pulls up a “memory” of a previous conversation I get a little creeped out. Yes, I’m that weird Linux guy working on Table Slayer that wants to have a sidebar about 1970s crime cinema. I’ve sent it pictures of my car to diagnose engine troubles, asked it about vacation spots, and generally shared a lot about myself. It’s concerning. Right now I share a lot about myself publicly already. My activity page is essentially a timeline of what I do all day, but it doesn’t capture everything, and I’m pretty specific about what I do share publicly and keep to myself.

I do worry about general access to these tools. Their ubiquity in daily life will be common, but if governments, companies and individuals have differing access to them, it will create a level of classism that will be scary. Right now the Internet and libraries provide nearly same access source material to us. Will AI models provide the same? Will they destroy source information as a way to collect knowledge?

What I know is that these tools aren’t going away. The wave has already broken, and we’ll need to be careful, both as creators and as consumers of these new tools. I remain optimistic overall, and despite the noise of the current era, think that humanity tends to reconcile towards goodness in the long term. Then again, I watch a lot of Star Trek, and was raised on The Muppets. What do I know?

Sníd