Skills, Agents, Prompts.

Prompts are not the whole system.

A prompt can shape a response. A skill can package a repeatable way of working. An agent can take bounded action inside a workflow.

Those boundaries matter more than the label. The useful question is what responsibility the system is allowed to carry.

Responsibility needs a place.

If nobody owns the exception path, an agent will inherit that confusion. If approval rules are vague, a skill will only make the ambiguity reusable.

Better prompts help only after the workflow says what should happen, who can decide, and when the system must stop.

If the work is unclear, the agent inherits the confusion.

Choose the boundary first.

Before deciding between prompt, skill, or agent, separate the work. What is judgment? What is execution? What is transfer? What must be escalated?

The technical shape should follow the workflow boundary.