Skills

A skill is the smallest user-visible unit of capability in Relay-E. It bundles:

• A system prompt — the persona and rules.
• A tool set — the functions the model is allowed to call.
• A preferred model tier — fast / balanced / premium.
• Optional examples — few-shot guidance.

Customers (your end users) compose skills. Their assistant might be the financial-advisor skill plus the customer-support skill, sharing a session.

Defining a skill

import { defineSkill } from "@relay-e/engine";

skills.register(
  defineSkill({
    name: "financial-advisor",
    description: "Reads balances, analyses spending, proposes transfers.",
    systemPrompt:
      "You are a personal finance assistant. Use the tools to fetch real " +
      "data — never invent balances or transactions. For destructive actions " +
      "(transfers), explain what you're about to do and rely on the approval flow.",
    toolNames: ["get_balance", "analyze_spending", "transfer_funds"],
    preferredTier: "balanced",
    examples: [
      {
        input: "How much did I spend on food last month?",
        output:
          "Calling analyze_spending(days=30) and reporting dining + groceries.",
      },
    ],
  }),
);

Picking a tier

The model router resolves the preferred tier against the configured providers:

If you've configured OpenRouter or OpenAI instead, the router falls back to their equivalents. See Concepts → Providers.

Composing skills at the request

A request can activate multiple skills:

curl -X POST .../v1/sessions/demo/messages \
  -d '{
    "prompt": "...",
    "skills": ["financial-advisor", "customer-support"]
  }'

When multiple skills are active:

• System prompts are concatenated under # Skills headings.
• Tool sets are unioned (deduped by name).
• The first skill's preferredTier wins.

When to make a new skill vs. add a tool

Skill-as-config (roadmap)

Today, skills are TypeScript: they live in bootstrap/registries.ts and require a fork to extend. The roadmap is to make skills loadable from JSON / DB / env so customers can author their own without forking. The schema is already designed in packages/engine/src/skills/index.ts — it's serialisable.

When that lands, the same schema will power:

• A skill management API (POST /v1/skills)
• A web UI for authoring (in the dashboard, when that ships)
• Versioned skill marketplaces

Reference

interface SkillDefinition {
  name: string;
  description: string;
  systemPrompt: string;
  toolNames: string[];
  preferredTier?: "fast" | "balanced" | "premium";
  requiredModalities?: Array<"text" | "image" | "audio" | "video" | "document">;
  examples?: Array<{ input: string; output: string }>;
  config?: Record<string, unknown>;
}