Agent Skills

Open-source skills I've written for coding agents. Install one, then ask for what you want in plain English.

>> github.com/liwoo

How a skill works

  1. 01

    Install

    A skill is a folder with a SKILL.md in it: instructions, reference docs and a few scripts. It sits in your repo or your agent's skills directory.

  2. 02

    Ask

    When what you ask for matches the skill, your coding agent loads it. Claude Code, Cursor, Codex and Copilot all read them.

  3. 03

    Run

    The agent makes the judgement calls. Small scripts handle the parts that have to be exact, like mapping a repo or rendering a report.

The governance scorecard for Northstar, a sample refunds agent

skill: agent-governance-scorecard

Agent Governance Scorecard

Find out what your AI agent can get wrong today, and which lines to change.

It reads your agent's code along the path between "the model suggested an action" and "the system actually did it". You get 48 checks across the 6 safeguards from my Governance House talk, every gap traced to a file and line, and a fix list ranked by what can hurt you.

npx skills add liwoo/agentic-governance-skills

Then ask your agent: "Run a governance scorecard on this repo."

What happens

  1. IndexA tiny zero-dependency script maps the repo: frameworks, tools, anything that moves money or data, tests and CI. Secrets get masked.
  2. MapYour agent reads the code and writes down what it can actually do, from model output to side effect.
  3. Scope5 yes/no questions decide which checks apply. A Q&A bot only gets checked on Reliability; an agent that issues refunds gets the lot.
  4. AssessRoom by room, each check comes back present, partial, missing or unknown, with evidence. A system prompt saying "be careful" earns nothing.
  5. RenderThe script validates coverage and builds an HTML report plus a scorecard.json you can diff between releases or fail CI on.

See it work

The sample refunds agent asks a human before refunding, but the approval edge is unconditional, so a "no" still pays out. The scorecard catches it in 1 pass and scores the agent 8/100.

View on GitHub
A frame from truescreen's example reveal: the real customers list, a fingertip about to tap a row

skill: software-reveal-video

truescreen

Product launch videos made from your real UI, with zero mockups or screen recordings.

A Remotion toolkit plus a skill that turns your coding agent into a video editor. It compiles your actual React components into the film, feeds them demo data frame by frame, and poses them in 3D with headlines, taps and camera push-ins. When the UI changes, you re-render.

git clone https://github.com/liwoo/truescreen
cp -r truescreen/skill .claude/skills/software-reveal-video

Then ask your agent: "Make a 30-second reveal of my dashboard and settings pages."

What happens

  1. ScriptOne idea per beat, about 4 beats for 30 seconds. Every claim on screen gets checked against the code before it ships.
  2. ShootHeadline on the left, your app in a tilted window on the right, then the camera pushes in on the one detail that matters.
  3. WireIt finds the real components behind each screen and compiles them into the Remotion bundle, with the framework runtime stubbed out.
  4. ReviewIt renders a contact sheet of 1 frame per second and reads it for half-loaded screens and taps off target before the final render.

See it work

The repo ships Orbit, a tiny Next.js app, and an 18-second reveal built from its components. Every pixel of UI in the film is the app itself.

View on GitHub
skip to content