How I use AI in practice.
Tools like GitHub Copilot and Claude Code have changed every phase of how I work — not just how fast I type, but how quickly I can think through a problem and move to something working.
If AI only saved keystrokes, I would not care much about it. The bigger shift is cognitive. Context gathering, codebase traversal, option comparison, first-pass architecture — all of that moves faster now. What used to be an afternoon of chasing documentation and tracing dependencies can take twenty minutes.
I use tools like GitHub Copilot and Claude Code. Both do the same core thing well: read context, reason about it, generate useful output. Between the two, I have settled into preferring GitHub Copilot as my primary tool. It lives inside VS Code, the editor I already think in — I do not context-switch, I do not break flow, it is just there. The pricing is also meaningfully better for the volume of work I push through it every day.
Where AI earns its keep
It is not autocomplete. Before I write a line of code: reading an unfamiliar codebase, tracing how a feature actually works end to end, comparing two implementation paths before committing to one. During implementation: schema changes, repetitive handlers, file-to-file refactors, test scaffolding, boilerplate that has to be correct but does not need to be interesting. After: reviewing the diff, catching edge cases I missed, spotting places where I over-engineered.
The whole loop — from problem to shipped code — moves differently now. Refactors that used to cost hours become short feedback cycles. I can explore an option I would have skipped before just because the cost of trying it is so low. That changes what I build, not just how fast I build it.
Where judgment still lives
I build integrations, realtime features, and payment flows. These are systems where "often correct" is not good enough. A sync conflict, a race condition, a missing idempotency key — the AI will generate plausible code for all of these and sometimes generate it wrong. Every output gets read. Every pattern gets checked against how the actual system behaves.
The tools are fast and frequently right. But I still decide what ships. I still read the error and trace the system. I still catch the case where a generated pattern is technically valid but wrong for the product. That work has not changed — only the speed at which everything around it moves.
The frame that works for me: AI is leverage across the full loop, from research to delivery. Not a replacement for knowing what you are doing, but a serious force multiplier on the parts of the work that did not need to be slow.