AI-Assisted Development

AI tooling is part of almost all of my work. I use it as a thinking partner and execution accelerator, not as a replacement for judgment.

Daily drivers:

• Cursor — AI-native IDE with seamless integration between inline assistance and background agents. Works well enough that I can make meaningful progress from my phone.
• Wispr Flow — Voice-to-text that actually works for technical content. Saves hours of typing, especially for documentation and longer explanations.
• Coder — For longer-running tasks or when I need multiple agents working in parallel on different parts of a problem.

Process: One of the first things I do in any repository is create an agents.md file that describes the project structure, tooling conventions, and how to navigate the codebase efficiently. This gives AI assistants the context they need to be useful instead of generic.

The goal is to stay in the driver's seat while offloading the mechanical parts of coding.

When I use it

• For idea soundboarding, concept challenge, and generating alternatives before committing.
• For API scaffolding and fast iteration on implementation plans.
• For repetitive refactors, boilerplate-heavy changes, and documentation synthesis.

When I avoid it

• When external stakeholder input is missing and cannot be inferred safely.
• When decisions are high-stakes and cannot be validated with clear review criteria.
• When a claim cannot be verified and would require guessing.

One real example

When defining a new API surface, I use AI to propose endpoint variants, challenge hidden assumptions, and draft acceptance criteria. I then select a constrained version, implement against it, and iterate based on real integration feedback.