How I use Generative AI going into 2026
Posted on:
As Generative AI capabilites get better over time, I want to wanted make a time capsule of what things I am using with AI right now, what I am staying away from, and then be able to come back to this in the future and see how my usage evolves over time.
Note: All of these use cases are for personal projects and usage. I can use Gen AI at work, but I am wanting to learn about it outside of work and not cause issues by using experimental tech during my day job. Also, I am not trying to use Gen AI to profit out of new projects I make.
What I am using
Coding and talking with Claude, but with human in the loop
I personally right now am using Claude as my primary LLM family. I have the $20 plan and I have personally yet to run into any limits just given I am using it alongside my existing development workflow. As such, that means primarily using Sonnet or Haiku 4.5 and very rarely Opus.
Unlike a lot of people that I've seen who have a workflow where the majority of the work is in Claude Code or similar agents where they are prompting the models, I am trying to keep the usage minimal to start. Particularly, I may prompt for a function generation or a small refactor, but overall something that would take under 5 minutes or so. I've seen people who say they can use Claude Code with things like Ralph Wiggum loops to run it for hours, but I feel I particularly want to be involved heavily in the code generation so letting it generate thousands of lines of codes without any involvement still doesn't roll with my mindset right now. I may move further to the right of letting the agents do more of the work as the models advanced, but I still want a good amount of control.
Trying to secure systems
One of the main hiccups I want to personally protect is these CLI agents have flags like --dangerously-skip-permissions in Claude Code that just let you run rampant and are never asked for if it can take some action on your computer. What I don't want is that to delete something or get hit with a prompt injection attack.
I've been experimenting with Docker and container tech to at least protect my filesystem at large. I drop the project folder into a container via a docker volume such that if the system is wiped, the only files deleted are that repo which I likely have in source control and can do a quick git clone to restart. Docker Sandboxes seem to be a good start for implementing this, though they do still have some rough edges given if you want to add additional CLI tools you have to install them in the container.
I personally still am unsure how to protect against bad bash scripts / tool calls. Do have setup filters for what URLs http requests can be sent to? Do I prevent certain bash commands via the claude settings / custom rules? I know that the harnesses around these models are going to improve over time and hopefully enough security is in place that this becomes less of a problem.
Claude for mockups
One place I am okay letting Claude do a bit of hands-off work is using their Artifacts feature in Claude for UI design. I may prompt it for initial draft of a design and work on it inside Claude on the web before I am happy enough with it. I then would copy the code and start wrapping a backend around it. Given if you don't prompt it, the mockups are just react components with Tailwind for styling and if I am using React I can make minimal edits or if I am using a different templating engine I can refactor it into such.
What I am not doing
Writing
I personally currently have not used any AI model for professional writing and currently have zero plans on doing such. All of the text on this website is written without the usage of AI models as I prefer to write prose by myself and have that authenticity. As well, I appreciate the actual motions of writing rather than just trying to get to the end of a finished piece. This also goes for the same with email communications or discussions over chat apps as I want to put the time in for such.
Image and Video Generation
I feel these personally don't have any use cases in my personal workflows and day to day life. More importantly, I don't wish to add to the problem that these tools can cause misinformation or make harmful imagery given the diffusion models from Google, Midjourney, among others are getting more and more realistic.
AI Browser usage
There are either new browsers or new capabilites in existing browsers to browse on your behalf. Tools like Gemini in Chrome, Copilot in Edge, the Claude Code Chrome Extension, or browsers like ChatGPT Atlas, Perplexity Comet, or The Browser Company's Dia. I am not currently using any of these as my thought right now is a lot of the web is authenticated and I feel without security guardrails as I discussed with coding agents earlier, I feel it is not a good idea to stick with not using these new capabilites just yet.
Conclusion
Overall, you can see I am taking a nuanced approach to adding AI into my dev workflows. I'm not going the approach of barely write code anymore, but I am at least giving the tools a chance and seeing where their limits are. As mentioned earlier, this space is so fast moving that the tools we are using today may not be what we use 3 months, 6 months, or later on, but I do wish to give these tools a good chance to see where they can responsibly enhance my productivity rather than declare digital god.