Compose Agent Skills from Single-Responsibility Parts

The most maintainable agent setups mirror good software design: each skill does one thing and hands off cleanly to the next. One component should draft content in the right voice; a separate component should handle the mechanical act of placing it where it belongs. When you keep “compose the message” and “deliver the message” as distinct steps, you can swap, test, and reuse each independently.

This separation also creates natural safety checkpoints. Any step that performs an irreversible action — sending, publishing, transferring — should pause and echo back its inputs for explicit human confirmation. If a critical detail was guessed rather than supplied, the right behavior is to stop and ask, not to proceed.

A recurring operational gotcha: don’t assume a runtime or binary lives on the system PATH in your execution environment. Wrappers that rely on an implicit PATH fail silently in automation. Invoke interpreters and tools by absolute path inside your scripts so the behavior is identical whether a human or an agent runs them.