Traditional multi-stage Dockerfiles are extremely powerful but become difficult to reason about at scale:

• Linear file order + implicit dependencies via FROM foo / COPY --from=foo.
• Lots of boilerplate for “download in parallel for cache” patterns (see the giant torch/Dockerfile and torch-extras/Dockerfile in coreweave/ml-containers).
• Scripts must be COPYed (or put in heredocs) even when you only want them for the build.
• Secret mounts are verbose and easy to get slightly wrong (leaking into layers or logs).
• Hard to express “these N independent things can be built in parallel and then combined”.

Yamlfile makes the graph explicit:

• targets are named first-class citizens.
• from: and copy.from: (sibling targets today; the component:target form for multi-file references is accepted but not yet supported at runtime) declare dependencies.
• Only the targets you actually need are built.
• run.script lets you run a script from your build context without leaving a copy of it in the final image.
• Secrets have a clean declarative form that maps directly to type=secret (file or env=).

You still get the full power of BuildKit (caching, multi-platform, provenance, etc.).

See the Syntax Reference for how targets, dependencies, and steps are declared.