This is how I actually work. Not aspirational. Not what I think sounds impressive. Just the stack and habits that have stuck after a year of building with AI.
The Core Stack
I work across two main modes:
- Deep building — Cursor + a good model + my own brain
- Research and planning — Perplexity or web search + my own notes
Everything else is either support or distraction.
Cursor as the Centre of Gravity
My editor is Cursor. Not because it’s magical, but because it removed enough friction that I stopped context-switching.
The features I actually use:
- Tab completion for the boring stuff—boilerplate, repetitive patterns, obvious next lines
- Inline editing (Cmd+K) for refactoring a block without losing flow
- Composer for larger architectural changes where I need to see across multiple files
What I don’t do:
- I don’t ask AI to “build me a feature” and walk away
- I don’t accept generated code without reading it
- I don’t let AI write my business logic unsupervised
The rule: AI speeds up typing. It does not replace thinking.
Model Choice
I switch between models depending on the task:
- Claude for architecture, API design, and anything where taste matters
- GPT-4 for debugging, regex, and quick syntax questions
- Local models for sensitive data or when I don’t want to send code to a third party
I don’t have a religious attachment to any model. They’re all wrong sometimes. The trick is learning which one is wrong in the way you can catch.
The Research Layer
Before I write code, I usually need to understand something:
- A library I’ve never used
- A pattern I’m not sure about
- Whether someone already solved this problem
My workflow:
- Perplexity or web search for quick factual lookup
- Read the actual docs — AI summaries are decent, but docs are the source of truth
- Search GitHub for real-world usage examples
- Take notes in plain text — not in some fancy system, just a scratchpad
The notes matter. They force me to process what I found instead of bookmarking and forgetting.
Project Scaffolding
I created Dev Foundry because I was tired of setting up the same Astro + Tailwind + auth + database stack for every new project.
Now I start from a boilerplate that already has:
- Auth (OAuth + magic links)
- Database schema patterns
- Basic UI components
- Deployment config for Vercel
This saves me a day per project. The AI helps me customise from there, but the foundation is pre-thought.
The Writing and Shipping Loop
My actual daily loop looks like this:
- Morning: review what I planned yesterday, adjust if needed
- Deep work block: build the core thing—no AI, just focus
- Afternoon: use AI to fill gaps, refactor, write tests, fix edge cases
- End of day: deploy to staging, review, note what broke
The key insight: AI is for the second half of the day. Not the first.
If I start with AI, I end up with a mess I don’t understand. If I start with my own code and use AI to clean it up, I end up with something I trust.
What I Still Do Manually
- Code review: I read every diff before committing. No exceptions.
- Architecture decisions: AI can suggest, but I decide. The blame is mine.
- User-facing copy: AI drafts are always wrong. I rewrite everything.
- Tests for critical paths: AI writes decent tests, but I write the ones that matter most.
What’s Changed Over Time
Six months ago, I was using AI for everything. Now I’m more selective.
I noticed a pattern: the projects where I leaned hardest on AI were the ones I understood least. That felt fast in the moment, but slow later when I had to debug something I didn’t build.
Now I use AI as an amplifier, not a substitute. I still have to know what I’m doing. The AI just helps me do it faster.
The Workflow in One Sentence
Think first, build second, use AI to clean up the gap between the two.
That’s it. Nothing revolutionary. Just a discipline that took me too long to learn.
Next: The Trap of AI-Assisted Mediocrity — Why more output isn’t always better output.