Over the past year, I’ve rebuilt how I work. Between AI coding partners, better tooling, and a bit of structure, I’ve gone from half-finished side projects to shipping working products that actually make sense.
This post is a look at what my workflow looks like in 2025 and a peek into how tools like Claude Code, ChatGPT, and Linear fit together, and what I’ve learned about balancing speed and structure (while maintaining my sanity) along the way.
Slow start
Over the past two years, I’ve seen how new advances in Artificial Intelligence have unlocked the ability for hobby developers to build and develop tools and solutions at a speed previously unseen.
This acceleration has resulted in a Cambrian explosion — a sudden burst of creativity and experimentation that feels like everyone is building something new every week.
As someone who started their career in engineering and then took a deviation into the world of business and entrepreneurship, I saw tremendous potential in these tools to bring my (admittedly half-baked) ideas to life.
In my journey through entrepreneurship, I always had a list of business ideas or projects of various scales, which I had bouncing around my head or hastily written down in various notes apps.
There was always something blocking the pursuit of these initiatives (need for capital, resources, not scalable, etc.), but most often it was the need for development skills, which I didn’t really have.
While I could speak engineering, I knew that when staring at pages of code, I lose focus and don’t have a lot of patience with the minutiae of code.
If these tools could serve to abstract away this need for technical expertise, they might reduce the barrier to entry and enable me to build and experiment faster with lower risks.
At first, I was chasing speed and was trying to build faster, but I realised that this was not getting me very far since I was directionless. What I really needed wasn’t just speed — it was momentum.
First Steps
After realizing the potential of these tools, I decided to get my hands dirty.
I initially tried the cohort of “vibe coding” apps — a term used for AI-assisted IDEs like Cursor, Lovable, and Replit that let you code by describing intent rather than syntax.
While I found these tools easy to use, they didn’t really give me confidence that I could build something stable.
Things would look correct, but as I kept working on the products, I would keep running into issues I couldn't get out of.
My code would develop into spaghetti and become unmanageable, which really put me off the idea that “vibe-coding” was the answer.
New Launches
Those first attempts were messy, but they helped me understand what I actually needed from an AI coding partner.
This all changed with the launch of Claude Code.
All of a sudden, the AI was in my codebase, and I could instruct it with a much larger context window directly in my CLI.
This was much closer to what I was looking for, and I jumped on it to see how far I could take my projects.
I loved it. It was to the point and was able to solve just about everything I threw at it. The code was modular and manageable (to the extent that even I could understand it).
More importantly, it gave me direction. I could finally see how my projects fit together — not as random bursts of output, but as connected steps toward a tangible product.
All of a sudden, the gap between my ideas and execution felt smaller — like I finally had some momentum.
Together with Claude Code, I was able to build a few hobby projects and deploy some experiments, but it was at this point that I got to the next hurdle.
But with greater speed came a new problem: managing scope.
An Explosion of Features
Once I found a setup that worked, I leaned in — but it turned out I might have leaned on it a bit too hard.
It was now really easy to build new features. Whatever I could think of could be built, and I would do it.
If I thought it would add value, I would try it and push my code into longer and longer solutions.
Then it would break.
And I was right back where I started. I realized that I now had more power, but my scope was exploding since I didn’t have a good plan around what I needed to be doing.
My mindset was “can it be done?” not “should it be done?”
I didn’t have a good definition of what success was or what “done” was.
New features could just be added with minimal effort — I needed a way to rein things in and give my work some direction.
Once I started layering in structure, I noticed something interesting — I could go after larger projects without losing control.
Structure gave my work more mass. It added weight and stability, which meant I could build bigger systems or even juggle multiple projects concurrently without things collapsing under their own speed.
That’s when momentum really started to take shape — not just motion, but motion with direction and weight.
Bringing Order to the Chaos
I needed to slow down, add some structure, and start defining what “done” actually meant.
Now that I had a better idea as to what I was capable of building, I went back to my well of ideas and started working through them with Claude Code to scope them out.
Could they be built? Did they make sense? Would they take too much effort to maintain and scale?
Initially, I turned to Claude Code to help with this, but I found that for these sorts of brainstorming exercises, the answers weren’t as inventive as I’d like.
On some advice, I tried out ChatGPT as an alternative and found it to be a better thought partner for these sorts of exercises.
Together with the newer models from OpenAI, I was able to have nuanced conversations evaluating the scope and feasibility of my ideas and breaking them into manageable pieces with a defined scope of work.
Adding Structure
Once I knew how to evaluate ideas, the next challenge was managing the actual execution.
The next piece that I needed to do was to prioritize the projects and track the work.
I turned to Linear as a project management tool. With it, I was able to scope my projects into epics and issues, set up milestones, tags, priorities, and dependencies.
This helped me get a better handle on how much complexity a task would take and how it would impact other parts of the project.
I was also able to better timebox by enabling sprints, where I could evaluate how many story points I could complete in a given session or day.
Claude Code’s time-out windows actually came in handy here — they served as natural five-hour bursts of focused context before the session expired.
This enabled me to more accurately track my progress and plan where I’d be in the short term.
To manage feature development, I moved more strictly to Git- and TDD-based development.
I instructed Claude Code or OpenAI Codex to write the tests before writing any code.
I also set up Git instructions to always develop in new branches and submit work as pull requests that I’d review in GitHub.
This enabled me to review what was done, identify any issues, and request corrections within the same context window before merging.
Switching to this workflow helped prevent errors from propagating through the codebase and minimized the risks of context windows being polluted since I was starting fresh with every new issue or branch.
Tools in My 2025 Development Workflow
| Category | Tool | Purpose |
|-----------|------|---------|
| AI Coding Partner | Claude Code | Core development assistant for building and debugging |
| Brainstorming & Ideation | ChatGPT | Helps scope, refine, and challenge project ideas |
| Model Foundation | OpenAI Codex | Generates and tests code during TDD |
| Project Management | Linear | Planning, sprinting, and tracking progress |
| Codebase Management | GitHub | Version control and pull request review |
| Early Experiments | Cursor, Lovable, Replit | Low-friction AI coding environments for quick tests |
What I’ve Learned About Working With AI
Working with AI as a coding partner has been incredibly enlightening.
I’ve learned that:
- Scope management matters more than ever. As the barriers to build come down, it’s even more important to keep focus on the end goal.
- Context is a currency. Keeping sessions small and focused makes AI output exponentially better.
- Burnout hides in productivity. Just because I can build faster doesn’t mean I should keep going nonstop.
- The most valuable thing AI gives me isn’t speed; it’s momentum.
Momentum is mass times velocity — speed with direction. And that direction is what makes the work feel meaningful.
So this is where I am now.
Juggling between bursts of writing code, reviewing pull requests, scoping future work — with childcare sandwiched in between.
And it works for me.
What’s next? Playing around more with Claude Code and Codex on the web; exploring Claude Skills; and continuing to build and ship — this time with a clearer sense of direction and enough structure to keep the momentum going.
This is a living document. Last updated: October 2025