How do you stay fast with AI without losing understanding of your code?
AI is great for code search and brainstorming, but for design docs and implementation, carefully reviewing everything it writes makes me slower than doing it myself. How do people keep both speed and a real understanding of what ships?
Reviewing after the fact is the slowest place to build understanding. You are reverse-engineering decisions the model already made. It is usually faster to make the key decisions yourself up front and let the model fill in the parts you have already agreed on.
A practical split: talk through the design with the model first (options, trade-offs, edge cases), settle it, then have it implement in small steps. You keep the understanding because the shape of the solution came from that conversation, not from reading the final diff.
For the review itself, ask the model to prove its claims instead of trusting a summary. Run the tests, show the output, point to the lines that handle each edge case. That turns review into checking evidence rather than rereading everything.
And keep doing some parts by hand. Debugging an unfamiliar failure step by step is where most of the understanding comes from, so let the model assist that process rather than replace it.
Listings mentioned
- Brainstorming Ideas Into Designs · skill by obraWorks through the design and trade-offs with you before any code is written.
- Verification before completion · skill by obraMakes the agent run checks and show evidence before it claims the work is done.
- Systematic debugging · skill by obraA step-by-step debugging routine that keeps you in the loop on root causes.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.