skill

Go Skill Discovery Library

A high-performance, token-efficient discovery system for LLMs.

Overview

This library provides a minimalist, SQL-based mechanism for LLMs to discover and invoke tools ("skills"). It is designed to minimize token usage in the initial system prompt by providing a compact index and allowing the LLM to query for details on demand.

Features

• Token-Efficient Index: GetIndex() returns a highly compact string (e.g., files(read,write), net(ping)) for the initial context.
• SQL-Based Discovery: LLMs can query the catalog view to retrieve detailed parameter schemas only when needed.
• Idempotent Registration: Register() handles upserts and auto-provisions categories.
• Minimalist Schema: Uses cats, skills, params tables and a catalog view.

Usage

Installation

go get github.com/tinywasm/skill

Initialization

Initialize the store with any *orm.DB connection (e.g., github.com/tinywasm/sqlite).

package main

import (
	"context"
	"fmt"
	"log"

	"github.com/tinywasm/skill"
	"github.com/tinywasm/sqlite"
)

func main() {
	db, _ := sqlite.Open(":memory:") // or file path
	store := skill.NewStore(db)

	// Initialize schema
	// Make sure to create necessary tables and the catalog view using db.ExecSQL
}

Registering Skills

ctx := context.Background()

mySkill := skill.Skill{
    Category: "weather",
    Name:     "check",
    Info:     "Get current weather",
    Parameters: []skill.Parameter{
        {Name: "city", Type: "string", Info: "City name", Required: true},
    },
}

if err := store.Register(ctx, mySkill); err != nil {
    log.Fatal(err)
}

LLM Integration

1. Initial Context

Inject the compact index and instructions into your system prompt.

index, _ := store.GetIndex(ctx)
// index example: "weather(check), files(read,write)"

systemPrompt := fmt.Sprintf(skill.LLMInstruction, index)

The skill.LLMInstruction constant provides the standard prompt:

"Available: [INDEX]. Query catalog table for args schema. Example: SELECT args FROM catalog WHERE name='read';"

2. Skill Discovery

When the LLM needs to use a tool, it can query the catalog view or you can use Search() programmatically.

// Programmatic search
skills, _ := store.Search(ctx, "weather")

3. Execution

The LLM can query for the arguments schema directly:

SELECT args FROM catalog WHERE name='check';

Returns JSON: [{"n":"city","t":"string","r":true,"d":"City name"}]

Schema

• cats: id, name (unique)
• skills: id, cat_id, name (unique), info
• params: id, skill_id, name, type, info, req
• catalog (VIEW): cat, name, info, args (JSON array of parameters)