dzhng/skills

What it is

Skills is a personal library of reusable, domain‑agnostic agent skills that can be plugged into many AI‑coding assistants (Claude Code, Codex, Cursor, Duet, etc.). Each skill is a small, self‑contained folder that describes a single piece of workflow – e.g. exploring unknown requirements, writing a spec, running automated research, reviewing code, or checking visual designs. The repo ships the definitions, documentation and a tiny CLI that lets you install the collection into the skill directory of your chosen harness.

Why it matters

Modern AI‑driven development is moving from single‑task prompts to software factories – autonomous agents that iterate on a goal until the result is trustworthy. To make that possible you need a set of well‑defined, composable actions that an agent can call, verify, and chain together. This repository provides exactly that: a curated set of “building‑block” skills that break a large feature into independently verifiable slices, keep a living spec, and audit every decision an agent makes. By reusing these skills you can:

• Reduce the amount of custom prompting needed for each project.
• Ensure each step (exploration, spec‑writing, implementation, review) is repeatable and auditable.
• Plug the same library into many different AI coding environments.

How to get started

1. Install the collection – either via the npm registry:

   npx skills add dzhng/skills
   

   or from a local clone:

   npm run install-skills   # installs into ~/.agents/skills and links into ~/.claude/skills
   

   You can target a specific repo with npm run install-skills -- ../my-app.
2. Pick a skill – run npm run list-skills to see the categories (engineering, visual, authoring, graphics) and names.
3. Invoke it – most harnesses expose a slash command, e.g. /explore-unknowns or /implement-spec. The skill will interactively ask questions, generate artefacts, and write files under a specs/ folder.
4. Chain them – a typical end‑to‑end flow is:
   1. /explore-unknowns → map the problem.
   2. /write-spec → turn the map into a verifiable spec.
   3. /goal /implement-spec specs/<feature> → let the agent build the feature unattended.
   4. /review (or /audit-choices) → get a decision ledger and close the loop.

What’s inside

Category	Example skill	What it does
Engineering	explore-unknowns	Interviews you to produce a four‑quadrant map of knowns, unknowns, assumptions and risks.
	write-spec	Breaks a feature into independent, testable slices with clear API seams.
	implement-spec	Drives the agent to build each slice, optionally using a goal‑mode harness.
	auto-research	Runs fast experiments, evaluates metrics, and returns a parameter‑effect map.
	code-review, audit-choices	Audits diffs and the rationale behind every AI‑made decision.
Visual review	compare-screenshots	Generates a script that diffs screenshots against a baseline and flags visual regressions.
Authoring	write-skills	Helps you create or edit new skill definitions.
Graphics	renderer	Provides architecture guidelines for GPU rendering with WebGPU/TypeGPU.

Who should use it

• Developers building AI‑assisted tooling who want a ready‑made catalogue of prompt‑driven actions.
• Team leads looking to standardise how AI agents plan, implement and audit work.
• Researchers experimenting with autonomous coding loops and needing reproducible skill definitions.

License

MIT – free to use, modify, and redistribute.

---

All information above is taken directly from the repository’s README.