Skip to content

nicholaswagner/skills

Repository files navigation

Skills

My personal collection of guardrails for models that have never doubted themselves.

I use these primarily with locally hosted models using Pi, but they work equally well with Claude and Codex.

Installing skills

Frontier skills live at the repo root. Crisp skills live under crisp/ as compact, explicit variants for bounded-context models. Either variant is made available to an agent by symlinking it into that agent's skill directory:

Agent Skill directory
Claude Code ~/.claude/skills
Codex ~/.codex/skills
Pi ~/.agents/skills

The linkr skill manages these links (it is also handy as a plain CLI — symlink linkr/scripts/linkr.sh onto your PATH as linkr):

linkr list                  # state of every skill × agent
linkr link tv codex         # make a skill available to one agent
linkr link herdr --crisp    # link the crisp variant everywhere
linkr enable yt-dlp         # link a skill for all agents
linkr disable herdr pi      # park a link without losing it
linkr unlink vhs claude     # remove a link

Anything added to this repo with a SKILL.md is picked up automatically: root skills show as frontier, while crisp/<skill>/ entries show as crisp.

The skills

Skill What it does
analytical-wit An experiment in creating a stable AI voice inspired by the analytical, conversational style of my favorite comedian, Dara Ó Briain. It combines routing, examples, and drift controls so humor emerges from reasoning instead of being stapled on afterward.
apple-define Looks up definitions from the local macOS Apple Dictionary through DictionaryServices.
apple-tv Controls Apple TV devices on the local network with pyatv for pairing, power, app launching, and remote input.
conventional-commits Inspects Git changes, drafts Conventional Commit messages, and creates commits only when explicitly asked.
convert-pdf Uses Poppler utilities to convert PDFs to text, HTML, or images; extract embedded images; and split or merge documents.
herdr Controls a running herdr instance (a terminal multiplexer for agents) from inside one of its panes — managing workspaces, tabs, and pane splits, spawning agents, reading pane output, and waiting on state changes over its local unix socket.
html-to-markdown Converts files, pasted HTML, and web pages into clean Markdown or other text formats, (removing images by default). html-to-markdown
linkr Manages which skills in this repo are available to each agent (Claude Code, Codex, Pi) by creating, removing, and parking symlinks in their skill directories — link, unlink, enable, disable, per agent or globally.
motion-skills A set of focused Motion skills for React and vanilla JS animation: core props, gestures, layout, scroll, hooks, exit animations, vanilla APIs, and performance.
session-recall Answers "what did we do earlier / yesterday / last week" questions by querying Claude Code's on-disk session transcripts and prompt history instead of claiming no access to past sessions.
tv Controls a TCL Roku TV over the local network with plain curl via Roku's External Control Protocol — status, key presses, app launching, and text entry.
vhs Writes and validates VHS tape files for reproducible terminal GIFs, videos, and screenshots.
yt-dlp Downloads video, audio, playlists, subtitles, and metadata using task-specific yt-dlp recipes.

License

Licensed under the GNU General Public License v3.0.

About

My personal collection of guardrails for models that have never doubted themselves.

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages