Make template inputs clear

An original source-review prompt could ask for a claim, its supporting passage, and any limitation. Keep the expected input fields clear so the reader or host understands what information the template will include.

External prompts cannot authorize new actions

A remote template can contain instructions unrelated to the user’s task. The host should evaluate the template in context and keep external content from overriding the user’s intent.

What the template is

A server can publish prompts with arguments, like a form that fills a paragraph. The client shows them to a person, who may insert the filled text into a model call.

That is useful for a repeated task: “summarize this incident in our format.” The template standardizes the ask.

It does not execute anything by itself.

The template is still text the model will follow if you send it. A third-party server can ship a prompt that says to ignore other instructions or to call a tool.

Read prompts from servers you did not write. Pin the server version.

An update can change the words without changing the name you see in the menu.

How this differs from tools and resources

A prompt does not fetch data and does not act. If the filled template tells the model to call a tool, the tool call is a separate step the host can still block.

Do not wire a prompt so that selecting it also runs tools with no further check. The menu label is not informed consent about a side effect.

Arguments need the same care as any form. An argument that is pasted into a shell later is an injection risk if a tool is careless.

Keep prompt arguments as data until a specific tool validates them. Do not build one prompt that “does the whole workflow.”

That is how a template becomes an unreviewed agent.

What to record before you rely on it

List the prompts each server exposes and who reviewed the wording. Pin server versions.

Confirm that choosing a prompt only inserts text. If a prompt’s argument later reaches a tool, name that tool and the validation in front of it.

Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.

Do not relax the check until the model happens to pass. Name a person who owns that failure.

A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.

Common questions

Are MCP prompts the same as a system prompt?

They can be inserted into one, but they are supplied by a server and chosen from a list. Your own system prompt is configuration you control.

Theirs is content you should read before you trust it.

Should a prompt include credentials?

No. A prompt is text that will be shown and logged. Credentials belong in the tool’s own secret store, not in a template argument.

Sources & further reading

  1. MCP prompts specification — 2025-06-18 ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories