Create Skills In C#

Declare a skill as a [Skill] method in its owning tool harness:

using HPD.Agent;

public sealed partial class DataTools
{
    [AIFunction(Name = "validate_dataset")]
    [AIDescription("Validates the active dataset and reports problems.")]
    public string ValidateDataset()
    {
        return "Dataset is valid.";
    }

    [Skill]
    public static Skill DataAnalysis() => Skill.Create(
        name: "data_analysis",
        description: "Helps validate and analyze datasets.",
        instructions: SkillInstructions.FromText("""
            Read the validation guide first.
            Validate the dataset before analyzing it.
            Explain any validation failures.
            """),
        capabilities:
        [
            SkillCapabilities.Function<DataTools>(nameof(ValidateDataset)),
            SkillCapabilities.Resource(
                name: "read_validation_guide",
                description: "Reads the dataset validation requirements.",
                content: "Every row must have a unique identifier.")
        ]);
}

Register the harness normally:

var agent = await new AgentBuilder()
    .WithToolHarness<DataTools>()
    .BuildAsync();

Authoring Rules

name is the model-visible activation function name. description is visible before activation, so keep it short and focused on when to select the skill. instructions are returned only after activation.

Use SkillCapabilities.Function<T>(nameof(Member)) for C# function references. The referenced function may belong to another generated harness, but that harness must be available to the agent. HPD validates missing functions and conflicting model-facing names during generation or agent construction.

Every model-selectable resource and script requires a useful description. Describe what it returns or changes and when it should be selected; do not merely repeat its filename.

Only SkillActivationLifetime.MessageTurn is currently supported. Other enum values fail validation instead of silently using different behavior.

Optional Reinforcement

Activation instructions are always returned in the tool result. For critical rules that benefit from higher-priority emphasis, add reinforcement:

reinforcement: SkillInstructions.FromText(
    "Never analyze a dataset that has not passed validation.")

Enable delivery through configuration:

var config = new AgentConfig
{
    Skills = new SkillRuntimeOptions
    {
        InstructionDelivery =
            SkillInstructionDelivery.ToolResultWithSystemReinforcement
    }
};

Reinforcement supplements the activation result; it does not replace it.

Generated [Skill] declarations require source generation. They do not use the JIT reflection fallback available to some ordinary tool functions.