A colleague and I were discussing a complex design document that he and Claude had produced. I’m finding it difficult to review such documents as they often end up with way too much detail, because it’s so easy for the AI to keep writing.
I suggested trimming it down and remembered an agent skill I had come across recently called visual-explainer that generates little web pages or slide decks with short overviews of a document or even an entire PR.
An agent skill, if you’re not chronically AI obsessed, is a simple standard
for injecting knowledge into your AI. It’s a text file that you register for
your coding agent. Once registered, the AI will see an instruction like “Read
<this SKILL.md file> if you are asked to visually explain a complex idea.” It
reads the instructions and uses that knowledge to do better work.
So after selling my colleague on how cool this skill was, he asked me to just run it right now on the call.
Uhh… oh… wait. Which coding agent did I install that in?
I’m trying to use every coding agent out there and I’m constantly changing where and how I work:
- I use Claude Code
- I use Codex
- I use the Codex Mac App
- I use Copilot CLI
- I use Amp
- I do some work on my Mac
- I do some work in VS Code
- I do some work in devcontainers
- I do some work in VS Code in a devcontainer, or even a CLI coding agent in a devcontainer
- I do some work on my Omarchy Linux PC
- I do some work in VMs in the cloud
So I have all these coding agents in all these locations. Which one has the visual-explainer skill?
I couldn’t even remember how to install one. It is such a simple idea (just a text file!) but there is a whole website directory, skills.sh, several tools, and each coding agent seems to have a different way to do it. It really sucks.
After an afternoon of research and experimenting I think I’ve got it cracked: my dotfiles and some bash.
All these locations are set up with my dotfiles repository which means I have configuration (like my shell prompt) consistent everywhere. This works out of the box with devcontainers.
A config file lists my skills. A script installs them across all my coding agents.
When I find a skill I want, I run this:
$ agent-skills add https://github.com/nicobailon/visual-explainer/tree/main/plugins/visual-explainer
added github.com/nicobailon/visual-explainer//plugins/visual-explainer@61f06e818e8fd49a86351d5981dcc6eeeae2f890
syncing visual-explainer
This added and installed the skill. Now I do need to update my dotfiles if I want this persisted. But the next time I have a host where I want all my skills installed, I simply run:
$ agent-skills sync
synced playwright dc48aff8208131776c9937326002bd60cf572ab6
synced openai-docs dc48aff8208131776c9937326002bd60cf572ab6
synced nerv-ui c2c890b023e7d7b6d9498b67b82ffb3a3c92340f
synced visual-explainer 61f06e818e8fd49a86351d5981dcc6eeeae2f890
Each of these skills is pinned to a git SHA, so if you are security conscious (and you should be, this is the best vector for a prompt injection attack ever) then you can at least feel somewhat secure that the skill isn’t going to become malicious out from under you.
I’ve extracted just the agent-skills tool from my dotfiles here:
agent-skills. I’ve been using it for
several minutes now and it seems pretty good so far.