Skip to content

programmatic-usage.md

Latest commit

 

History

History
96 lines (68 loc) · 2.2 KB

programmatic-usage.md

File metadata and controls

96 lines (68 loc) · 2.2 KB

Programmatic usage

The package exposes a small TypeScript/JavaScript API for programmatic access to the skill registry. It is ESM-only and ships with type declarations.

Install

npm install @krxgu/kernel-skills

API

import {
  listSkills,
  getSkill,
  getSkillPath,
  searchSkills,
  bundleSkills,
} from "@krxgu/kernel-skills";

import type {
  KernelSkill,
  KernelSkillMetadata,
  SkillSearchOptions,
  SkillDifficulty,
  SkillCategory,
} from "@krxgu/kernel-skills";

listSkills(): KernelSkillMetadata[]

Returns every skill's metadata. No I/O on the markdown content.

const all = listSkills();
console.log(`${all.length} skills available`);

getSkill(id: string): Promise<KernelSkill>

Returns metadata plus the full Markdown body. Throws if the id is unknown.

const skill = await getSkill("triton.write-triton-layernorm-kernel");
console.log(skill.name);
console.log(skill.content);

getSkillPath(id: string): string

Returns the absolute on-disk path to the SKILL.md. Useful when you want to stream the file yourself or hand the path to another tool.

const path = getSkillPath("cuda.write-cuda-gemm-kernel");

searchSkills(options): KernelSkillMetadata[]

Accepts either a query string or a SkillSearchOptions object:

searchSkills("rmsnorm");

searchSkills({
  category: "triton",
  difficulty: "advanced",
  tags: ["fp16"],
});

searchSkills({
  query: "warp",
  category: "cuda",
});

Filters compose with AND semantics. Tag filters require every tag to be present.

bundleSkills(ids: string[]): Promise<string>

Concatenates multiple skills into a single agent-ready Markdown string.

const md = await bundleSkills([
  "triton.write-triton-layernorm-kernel",
  "patterns.write-kernel-test-plan",
]);

// Send `md` as system context to your agent of choice.

Where the data comes from

Metadata is read from generated/skills.index.json, which is regenerated from skills/<category>/<skill>/skill.json on every build. The Markdown content is read lazily from the source SKILL.md files that ship inside the package.

The package never executes generated code, never compiles kernels, and never makes network requests.