Generate One Layer at a Time, Not the Whole Feature

A workflow where you ask Claude to build a feature one layer or slice at a time — data model, then business logic, then the endpoint, then the UI — verifying each before moving on, rather than requesting the entire feature in a single generation.

A whole-feature dump is unreviewable: hundreds of lines across many files where a wrong assumption at the data layer silently poisons everything above it, and you find out only when nothing works. Layer by layer, each step is small enough to actually read, the error surfaces at the layer that caused it, and Claude carries a verified foundation into the next step instead of compounding a guess.

Any feature that spans layers, unfamiliar frameworks where you can't eyeball correctness, and work you intend to maintain rather than throw away.

Small, single-file additions, or well-worn boilerplate you can regenerate cheaply if it's wrong. The ceremony isn't worth it below a certain size.

We'll build <feature> layer by layer. Do NOT jump ahead.
Start with just the data layer: the schema/models and the functions to read and write them, plus a quick way for me to verify it works. Stop there and wait.
I'll confirm before we do the logic layer, then the API, then the UI.

Building a comments feature, Claude first ships the comments table and a tested createComment/listComments pair; you run them, confirm, and only then does it build the endpoint — so when the UI misbehaves later, you already know the storage and logic are sound.

Keep a short running 'contract' note between layers (the function signatures and invariants each layer exposes) and paste it into the next step, so later layers build against the real interface rather than Claude's memory of it.

• Saying 'build the whole thing' and then trying to review 600 lines at once.
• Skipping the verify step, which defeats the point — an unverified layer is just a delayed bug.
• Letting the layers drift apart because the interface between them was never written down.