Skill, system prompt or MCP: where an expert’s judgment goes

You have a way of making calls in your field, and you want an AI to work by it. There are three places to put it: a system prompt, a skill, or an MCP connection.

Most comparisons of the three are about tools and procedures. This one asks a narrower question: where should one person’s judgment live?

The three, in one line each

• System prompt. Text loaded at the start of every conversation. It is always on, and it stays as you wrote it.
• Skill. A folder with instructions that the assistant loads when the task matches. The format has been an open standard since December 2025, and several assistants read it.
• MCP. A connection to something outside the chat. The assistant calls it when needed, and the material stays at the other end.

Anthropic’s own comparison puts it this way: “MCP connects Claude to data; Skills teach Claude what to do with that data.”

Judgment is not a procedure

A procedure is the same every time: how to format a report, how to run a release. It is written once and rarely argued with. A skill suits it well.

Judgment is a record. It is the calls one person made, the reason for each, and the conditions that would have changed them. It grows every month, it gets revised, and it belongs to somebody. Those three properties decide where it should live.

Five questions that pick the place

1. How much is there?

Ten rules with clear boundaries fit in a system prompt, and nothing else is needed. A few hundred decisions do not. Every line of a prompt is loaded on every turn, whether the question needs it or not. A skill and an MCP are both read on demand.

2. Who will use it?

If it is you and your team, a skill in your own repository is the simplest thing that works. If it is other people, they each need a way to get it.

3. How often does it change?

A skill that somebody copied is a copy. Your edit reaches them when they update. An MCP address returns the current version the next time an assistant calls it.

4. Who may read it?

A file is read by whoever holds it. A connection can be public, limited to a team, or opened with a key, and it can be closed. Closing it stops the next read. An assistant that has already read a decision has read it.

5. Whose name is on it?

A prompt and a skill are usually anonymous by the time they reach a third person. If the value is that these are a specific person’s calls, the place needs to say who published them and when they last changed.

A rule of thumb

• Your own short list of rules: a system prompt.
• A procedure your team repeats: a skill.
• One person’s decisions, used by people outside the team and updated over time: an MCP connection.

They also combine. A single line in a prompt or a skill can tell the assistant when to consult a model, while the MCP holds what is consulted.

What none of the three fixes

The place does not write the content. A skill full of general advice and an MCP full of general advice give the same general answers. What makes any of them specific is the decisions inside: what was decided, why, and what would have changed it.

And none of them guarantees the assistant applies the material well. Ask it a question you know the answer to, change one condition, and read what it does.

See the third option

avva publishes one person’s decisions in one field as a Decision Model, read over MCP. You can put a question to one without a key. If the judgment is yours, here is how to build one.