mcpField
J
jh-create-skill

jh-create-skill

github.com/jasonharmongitAgent workflows038 installsUpdated 4d ago
Repository

About jh-create-skill

A skill-authoring guide for structuring agent skills and writing their metadata and content. It focuses on keeping instructions task-oriented, stable across runs, and free of one-off procedures.

Key features

  1. 01Defines skill folder and file layout
  2. 02Specifies frontmatter fields and naming
  3. 03Explains skill dependencies and dependents
  4. 04Separates reusable tasks from one-off work
  5. 05Requires concise, header-free skill content

Use cases

  1. 01Creating a new agent skill
  2. 02Reviewing or restructuring an existing skill
  3. 03Deciding what belongs in a skill versus runtime context

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/jasonharmongit/skills --skill jh-create-skill

After installing, type/jh-create-skillin Claude Code to confirm it loaded.

Details

From the project README

Structure

File structure
skill-name/
├── SKILL.md
└── scripts/          # optional

Put every step, rule, and example in SKILL.md. Do not add other sibling docs.

Frontmatter
---
name: your-skill-name
description: Brief high-level description of what this skill does.
disable-model-invocation: true
metadata:
  dependencies: other-skill
  dependents: downstream-skill
---
  • name: kebab-case, max 64 chars, lowercase letters, numbers, hyphens only.
  • description: A single sentence giving a very high-level summary of what the skill does. Do not include specifics about how the skill works, when to invoke it, or trigger terms.
  • metadata:
    • dependencies: skills this one tells the agent to read and follow (**`skill-name` skill**, **`skill-name`**, or a link to skill-name/SKILL.md). Skill-to-skill only - not hooks or sub-paths.
    • dependents: the reverse. Keep both sides in sync.
    • Comma-separated name values. Omit empty fields.
    • Do not infer dependencies from artifact names or shared tooling unless the body explicitly names the skill.

Content

Read the writing-for-agents skill and follow it - this skill covers skill-specific mechanics only.

  • Do not include any kind of header. No title, no description, no starting section, no introduction, no preface.
  • List a skill in dependencies only when this body routes the agent through that skill's workflow; an isolated fact or rule stays inline here.
  • When a skill is listed, remove from this body any step or rule that dependency already owns - do not restate it.
Task, not one run

A skill tells the agent how to do a kind of task. It is not a write-up of one session you or the agent already ran.

Put in the skill

  • Ordered workflow and scope limits (what to do, what to skip).
  • Hard rules the agent breaks without them (environment traps, "never halt remsh", approval gates).
  • Pointers to other skills when that workflow is their job.

Leave out

  • Procedures that belong to a single feature unless the skill name and description are explicitly about that feature only.
  • Copy-paste from a past run: sample message bodies, TypeIDs, SQL, Oban worker names, API routes, Redis keys, "cancel job then reset status" recipes, and similar one-off scripts.
  • Details the agent can load at run time from the user's request, tests, or the codebase.

At run time the agent reads the user's ask and the repo for specifics. The skill stays stable when those specifics change.

Comments

Markdown supported