failureMode: underspecified-task excerpt: "Do the thing" only works if there's exactly one obvious thing.

The hidden cost of "do the thing"

When a prompt assumes a default action, the LLM picks one. If your infrastructure makes the default unambiguous, fine. If not, you get whatever the model decides is most likely - which may not match what you wanted.

Underspecified tasks are the most common prompt failure mode in professional workflows. They don't fail with an error. They fail silently, producing work that looks right until someone looks closely.

The antidote isn't exhaustive specification in every prompt. It's a shared convention layer: CLAUDE.md files, workspace manifests, voice rules, prior decisions as context. With those in place, "do the thing" resolves to the right thing.

(placeholder body - flesh out before launch)