Slowing Down to Speed Up with AI

Everyone is sharing about AI, so I definitely have FOMO and want to finally give my own AI talk.

Happy to kick off the JuniorDevSG July meetup by sharing what the past 6 months of working with AI have actually been like. There has been excitement, moments where it felt like we were all becoming 10x engineers, and countless moments of frustration. There is also the comprehension debt where I look at a function and genuinely cannot remember writing it. Or maybe I did not write it in the first place.

The biggest realization came when I vibe coded a new product flow at work. Everything looked fine in the browser, so I did not read the diff because that was the advice everywhere. Work at a higher level. Use more AI. Run more agents. Spend less time reading code and more time looking at the output.

I did exactly that.

When the backend was ready, I finally sat down to review the code before merging. I found five different components rendering the same pricing card. Cleaning that up was not fun. Whatever time I saved at the start, I spent cleaning everything up afterwards.

That experience changed how I think about AI. The tooling is incredibly fast at building things. The slow part is still mine: deciding what to build, and reading what it produced. Slow down in those two places, and you actually speed up.

That is what I will be sharing through a bunch of live demos and lessons learned from building real products with AI. I shared an earlier version of this at SMU .Hack last month, and I am happy to bring it to the JuniorDevSG community.

See you on July 14!

Vibe coding feels incredible. You describe a feature, it appears, and you keep going. It looks right in the browser, so you stop reading what the AI wrote.

AI is fast at the doing. The slow part is still yours: deciding what to build, and reading what it actually shipped. That part did not get less important. It got more important, because now a bad decision gets built in an afternoon instead of a week.

I learned this the hard way. At work I started a new product flow, home to listing to detail page. There was no backend yet, so it was almost all front-end on mock data. I vibe coded it for three to four weeks. No real plan, and I never read the diff. It looked fine, so I kept shipping.

Then the backend arrived and my front-end had to merge in. I finally read the code, and I did not recognize it. Five different components were rendering the same pricing card. I read, refactored, found more, refactored again. Whatever time I saved up front, I lost cleaning it up.

So I changed how I work. Now I slow down in two places. Before I hand the work to AI, I get clear on what I actually want. That is not extra process, it is the cheapest place to catch a mistake. The same goes for reading the diff myself before anyone reviews or deploys it. This runs through everything I do now, at work and on side projects, from that cleanup to the repetitive workflows I turned into skills.

None of this needs a complicated setup. Learn 20% of a language and you can say 80% of what you need. AI is the same. Pick one tool. Run Explore - Plan - Code - Commit. Keep an AGENTS.md. Build skills for what you repeat. That is what you will walk away with.