How I use AI

Like so many others, I've been experimenting with a number of different AI-powered tools to find the ones that work best for my style of working and how I learn.

For folks trying to get some ideas of how to start applying AI in their workflows, I hope this quick run down of how and where I've been using AI so far is helpful.

Where I use AI

Coding on problems where I have a good understanding of the solution

I've found I've had the most success with AI coding when I've prompted Claude Code to propose a plan on how to solve a certain problem with a decent amount of direction from me on how to do so. After the first proposal, I iterate on it a bit until it's where I want it to be, and then allow the AI to make the edits.

I've noticed I have to be careful not to let the AI spin its wheels for too long, otherwise it occasionally goes off-track. But for the most part, this has been good for small features and bug fixes.

Helping me understand a codebase

In this respect, the new AIs have been a godsend. There are a number of different occasions where asking Claude specific questions about the codebase has saved me significant amounts of time.

I used to use Cursor for this, but I found that Claude Code was giving me better results overall and so eventually cancelled my plan with Cursor to go all-in on Claude Code.

Web searches

Sometimes I struggle to think of concise keywords for what I want to search for, and Perplexity has been great for this, particularly because it provides citations for the results it returns. I often verify what it says by visiting those citations, and discover and learn new things beyond what I was initially searching for by doing so.

Code reviews

For my local code changes, I leverage Claude Code to do a quick pass. For code changes within my team at my day job, I use Cubic.

It was initially not picking up much when I first introduced it to the engineering organization but over the past few months, it's gotten real good. It's learned enough from feedback by other engineers that when it leaves a comment, it's usually right.

Project updates

It used to take me about 10-15 minutes a day to write weekly project reports to share with the wider organization at my job.

Leveraging MCP servers like Github and Linear, and a Claude Skill that's specific to what I want from a project update document, has made this much quicker, and ensures that I don't miss any of the work my team has done since the last update.

Where I don't use AI

Any kind of writing

I had tried to use it to generate social media posts since I absolutely hate writing them, but they sounded so canned and "thought-leader-y" in a cringy way that after trying a couple of times, I decided to continue doing it my way. They may not result in as engaging of a social media post, but at least they feel more authentic to me.

In cases where I'm writing something more lengthy (like documentation), the temptation is there to give the AI a bullet point list of what I'd like to write about, and to get it to provide me with an outline to get started with.

However, I've noticed that my writing is worse off and requires more rounds of edits than without the use of AI because, as I write, I think about things I didn't initially consider. The adage of "writing is thinking" very much holds up in my experience.

Where I want to learn more

Agents

I feel like I haven't learned to effectively leverage agents to their maximum capabilities yet. I'm planning on spending some time here to see what workflows these can unlock for me.

Spec-driven development

I've seen folks get good results from building a project using AI and spec-driven development through something like spec-kit. I have a couple of small side projects that would be good candidates to try this out on, and am excited to see how they turn out.