The distinction between tools, resources, and prompts is the part I found most useful. Calling everything a tool hides whether an action changes anything, and that is where most of the risk sits. I also liked the warning about mixing examples from different protocol revisions. Pinning the revision and the SDK versions before drawing the sequence feels like extra work, and then it saves a long debugging session later. The diagnostic table is a good checklist to keep nearby when a call is rejected or a result lands on the wrong job.