More Joyful Building: My AI Development Stack, Q1 2026

How LLMs made building more playful, and the workflow I’m using now in 2026

Coding is fun, but I didn’t know it could be this fun. I’ll have an idea in the morning and it’s tested and running after I finish my first cup of tea. I try something ambitious and it works. I try something slightly unhinged and it mostly works. It feels less like work and more like playing in a creative lab. As CTO, I coded less as my company grew. After moving into an advisory role in 2025, I started coding more with LLMs for fun and realized just how much the paradigm has shifted. It’s fast. It’s fun. I’m obsessed.

Here’s how I’m feeling most delighted with development right now.

Ideation & Planning

For each project I have chats organized within their own folder in ChatGPT, Claude, and Gemini. I pay for Claude Max and ChatGPT Plus and use Gemini for free.

When I have a new product or feature idea, I bounce it off all three services using their chat interface and workshop it among them, debating and haggling until we’ve created a technical spec I’m happy with.

I talk to all three services using VoiceInk and only type when my environment demands it. Voice allows me to express myself more fully and more quickly, and it feels like I’m chatting with a set of advisors, which is more enjoyable.

Implementation

I use Rails as a monolith with RSpec for testing. It’s just so productive for me and I love the semantics of Ruby. Shoutout to the amazing Rails and Ruby communities.

To make implementation smoother, I have tight testing guidelines and style guides (code and UI/UX) covered in a CLAUDE.md file in each project’s repository.

Both Claude Code (CC) and OpenAI Codex are running via CLI on the project repo.

I put the final spec file in my project’s /docs/specs/active directory, double-check it, and have CC review it in Plan mode and answer any questions it has. Then I feed that review back into CC if needed and repeat until I’m satisfied.. Then it’s time to execute on the plan.

While CC is fiddle-faddling or flibbertigibbeting, I get a snack, do a quick chore, or get another agent to work.

Then I check CC’s work. Make adjustments and follow-ups as needed using the CLI, do some manual smoke tests. I then have Codex do a code review of the changes and feed that back to CC if necessary.

When CC struggles, I’ll see if Codex handles it better.

If there are front end issues or changes I want that I can’t accurately explain in a sentence, I take a screenshot, annotate it with Affinity, and feed it back to CC by pasting the image into the CLI. If I can explain the issue in a sentence I’ll still include a screenshot for context.

If any manual code adjustments are needed I make those in Cursor.

Feedback & Iteration

I’ve been conducting UI/UX tests using agents (both GPT-5.2 and Sonnet-4.6) with vision capabilities (but not DOM access). From my app’s admin, I spin them up and have them log in using different personas, from tech-inept to technical cofounder, across different devices. They navigate the app according to that persona’s goal and then give UI/UX feedback with screenshots and markup. They’ve already helped me identify and update a number of things to make the UI/UX easier and more intuitive.

Deployment

I use Git and GitHub for version control. Version control, testing, and LLM coding go together like puzzle pieces. Don’t skip them.

Pull Requests on Github generate a preview environment for testing with its own ephemeral database on Render.

I commit with a Rubocop autoformat hook via overcommit, push to GitHub where CI runs, and then Render pulls, containerizes, and deploys to production. Sentry provides monitoring and logging. Resend provides a nice email API and Firebase runs SMS auth. Rails handles the rest.

If I’m on the go and want to do something fairly straightforward, I’ll use the CC coding agent on my phone connected to the project’s GitHub repo. I recently created and deployed a feature while eating pizza at a restaurant. Lest you fear I was ignoring my company: I was dining alone.

The Catch

The joy is real, but you still have to steer. Left alone, the AI, much like an over-eager junior dev, will happily over-engineer, abstract too early, or invent solutions to problems I don’t have. I’ve had to talk it down many times. It can be fun to let it run wild, but the process is best when I’m guiding it. That’s when it becomes a collaborative instrument.

–

This workflow will evolve and change. And yes, there’s plenty more to automate, but I actually like being involved in these flows. I’m having fun building. What’s working for you?