Directives are not enforced by prompting

You can ask the model nicely and it will mostly listen most of the time. “DO NOT SKIP THIS STEP!!!” is not engineering. If a regex, a parser, or a tool-call hook can enforce the rule, use it.

Need a list of actions to be taken every time? Make each a tool call <step_one>, with the loop/pipeline not finishing until the final <end_run> tool is called. At the “end” of the conversation you can parse the tool calls to enforce that they were all called. If not, then either abort and retry the run, or if appropriate add a followup turn + directive to run the missing steps. Some agent frameworks give you a version of this when using their execution primatives, but its also something not terribly difficult to implement.

For SKILLs and personal/local agents this overhead may not be needed, but for production-facing agents this level of enforcability and operational insight is necessary.