[Feature] Support MCP Prompts for agent hinting

[eamonwoortman]

Overview

It seems that there's a MCP protocol that can assist agents with decision making on what tool to use.
From modelcontextprotocol.io;

Prompts enable servers to define reusable prompt templates and workflows that clients can easily surface to users and LLMs. They provide a powerful way to standardize and share common LLM interactions.

There's another project, mcp-cli that does a clever thing and incorporates prompts in their system prompt before sending the user prompt to the LLM.
See this video for more details.

Goal

I would expect it to aid the agent to be better at decision making on what tools to use.

Prompt example

An example of an MCP server implementing a prompt is Blender MCP.
Their prompt looks like;

@mcp.prompt()
def asset_creation_strategy() -> str:
    """Defines the preferred strategy for creating assets in Blender"""
    return """When creating 3D content in Blender, always start by checking if integrations are available:

    0. Before anything, always check the scene from get_scene_info()
    1. First use the following tools to verify if the following integrations are enabled:
        1. PolyHaven
            Use get_polyhaven_status() to verify its status
            If PolyHaven is enabled:
            - For objects/models: Use download_polyhaven_asset() with asset_type="models"
            - For materials/textures: Use download_polyhaven_asset() with asset_type="textures"
            - For environment lighting: Use download_polyhaven_asset() with asset_type="hdris"
        2. Hyper3D(Rodin)
            Hyper3D Rodin is good at generating 3D models for single item.
            So don't try to:
            1. Generate the whole scene with one shot
            2. Generate ground using Rodin
            3. Generate parts of the items separately and put them together afterwards

            Use get_hyper3d_status() to verify its status
            If Hyper3D is enabled:
            - For objects/models, do the following steps:
                1. Create the model generation task
                    - Use generate_hyper3d_model_via_images() if image(s) is/are given
                    - Use generate_hyper3d_model_via_text() if generating 3D asset using text prompt
                    If key type is free_trial and insufficient balance error returned, tell the user that the free trial key can only generated limited models everyday, they can choose to:
                    - Wait for another day and try again
                    - Go to hyper3d.ai to find out how to get their own API key
                    - Go to fal.ai to get their own private API key
                2. Poll the status
                    - Use poll_rodin_job_status() to check if the generation task has completed or failed
                3. Import the asset
                    - Use import_generated_asset() to import the generated GLB model the asset
                4. After importing the asset, ALWAYS check the world_bounding_box of the imported mesh, and adjust the mesh's location and size
                    Adjust the imported mesh's location, scale, rotation, so that the mesh is on the right spot.

                You can reuse assets previous generated by running python code to duplicate the object, without creating another generation task.

    2. If all integrations are disabled or when falling back to basic tools:
       - create_object() for basic primitives (CUBE, SPHERE, CYLINDER, etc.)
       - set_material() for basic colors and materials
    
    3. When including an object into scene, ALWAYS make sure that the name of the object is meanful.

    4. Always check the world_bounding_box for each item so that:
        - Ensure that all objects that should not be clipping are not clipping.
        - Items have right spatial relationship.
    
    5. After giving the tool location/scale/rotation information (via create_object() and modify_object()),
       double check the related object's location, scale, rotation, and world_bounding_box using get_object_info(),
       so that the object is in the desired location.

    Only fall back to basic creation tools when:
    - PolyHaven and Hyper3D are disabled
    - A simple primitive is explicitly requested
    - No suitable PolyHaven asset exists
    - Hyper3D Rodin failed to generate the desired asset
    - The task specifically requires a basic material/color
    """