Recap and Practice
Key takeaways
- A feature is a stack of small vertical slices, not one giant — build, run, and see each slice work before moving on.
- Run the build loop every time: slice → pick one step → prompt → review the diff → run the app → → repeat.
- Review diffs like a reviewer, not a compiler — watch for files it shouldn't have touched, silent behavior changes, and a diff far bigger than the ask.
- Let the AI run on small, cheap, common steps; steer manually (smaller prompts, closer review) wherever the blast radius is large.
- Commit only working states in small increments — if a slice goes sideways, throw it away and re-prompt instead of piling a fix on a broken base.
Try it
Pick one small feature you want in a project and slice it into 3–5 vertical steps, writing each as a single short sentence. Build only the first step: prompt for it, run git diff and skim the change, run the app to confirm it works, then commit with a one-line message. Stop there for now. The goal is to feel a single clean turn of the loop — one runnable slice, one diff you actually read, one commit you can return to.
I want to add this feature: [describe the feature in one or two sentences].
Don't write any code yet. First, slice it into 3 to 5 vertical steps, where
each step is something I can build, run, and SEE working on its own. Order
them so the app keeps working after every step. For each step give me a
one-sentence description and name the first file you'd change. Then wait — I'll
ask you to build them one at a time.