Most of my time working with AI is spent defining tasks (and reviewing the results). The hardest part is remembering everything that needs to be thought through beforehand.

So now I start almost every new feature the same way.

I create a "docs/backlog/feature" file in the project repo and write something like this:

I want to build feature X. Here's how it should work.

Ask me every question that needs to be answered before implementation in the following format:

Question

1. Option 1
2. Option 2 ... Answer:

After I reply, generate a complete specification based on my answers, then remove the questions and answers section.

What happens next:

• AI reviews the codebase and asks questions I wouldn't have thought of.
• It suggests possible answers for each question.
• Most of the time I just reply with 1, 2, or 3.
• Then AI turns everything into a complete specification that's ready to implement.

I like this workflow because it turns writing a spec into a simple 10-minute interview.

If you aren't doing this already, give it a try.

---

Hi, I'm Nick, a designer with 15 years of experience, formerly at Yandex. Now building Leaflo, a mental health app for self-help and reflection.

LinkedIn