Skip to content

About

No description, website, or topics provided.

Resources

Stars

7 stars

Watchers

0 watching

Forks

Latest commit

 

History

54 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agent skills

What it is

This repository is a collection of skills for coding-agent harnesses (Grok, Claude Code, Cursor, OpenCode, and similar). A skill is a directory with a SKILL.md file. The YAML frontmatter has name and description. Some skills also ship scripts/ and references/ next to that file.

Clone the repo and point the harness at the clone, or copy the skill directories into the harness skill folder. The agent then reads the playbooks from the files. AGENTS.md is the cross-harness bootstrap for the graphical workstation loop.

The package is this git repository. It does not build an application, and it does not contain a GNOME or Plasma container image. Supply that image from a separate lab project. Shared skills use placeholders (<host>, <port>, <user>) for lab endpoints.

Prerequisites

  • Git.
  • A harness that can read SKILL.md or AGENTS.md.
  • Workstation drive on Windows: Windows PowerShell. windows-ui-inject/scripts/Invoke-MstscConnect.ps1 prefers Windows PowerShell 5.1 because it uses Add-Type and System.Drawing. The desktop image is separate from this repo.
  • Local Hyprland (drive-hyprland): a Hyprland session, python3, hyprctl, grim, wtype, and ydotool with ydotoold and a writable /dev/uinput. drive-hyprland/scripts/install.sh host installs packages with omarchy or pacman when one of those is on PATH.
  • PDF training scripts: Python 3 and PyMuPDF (import fitz). OCR is optional (pytesseract and Pillow). to_datum.py, train_lora.py, and test_finetuned.py need the Tinker SDK (pip install tinker). The skill text records tinker 0.23.0 as verified.
  • mcp-server/SKILL.md is a Python MCP server specification. It documents Python 3.10+ and the MCP Python SDK. This repo does not ship a server package for it.
  • buzz-relay-self-host/scripts/verify-pair-endpoint.sh needs curl. jq is optional.

Install, check, and run

There is no root package.json, Makefile, lockfile, or test runner. Checks below are the commands the scripts actually accept.

Clone

git clone https://github.com/timsonner/agent-skills.git
cd agent-skills

Point Grok at the clone

From agent-workstation-drive/references/harness-install.md:

# ~/.grok/config.toml
[skills]
paths = ["<absolute-path-to-this-clone>"]

Claude Code and Cursor can use the same clone if the agent is allowed to read it. Optional copies go to ~/.claude/skills/<name>/SKILL.md or ~/.cursor/skills/<name>/SKILL.md.

Copy the workstation-drive family (Windows)

agent-workstation-drive/scripts/Install-ToHarness.ps1 copies these directories: agent-workstation-drive, gnome-wayland-rdp, kde-wayland-rdp, windows-ui-inject, browser-read-bypass, wsl-containers.

powershell -NoProfile -ExecutionPolicy Bypass -File .\agent-workstation-drive\scripts\Install-ToHarness.ps1 -SourceRoot (Get-Location) -Harness all

-Harness is grok, claude, cursor, or all (default all). Destinations are %USERPROFILE%\.grok\skills, %USERPROFILE%\.claude\skills, and %USERPROFILE%\.cursor\skills.

Hyprland install script

bash drive-hyprland/scripts/install.sh --help
bash drive-hyprland/scripts/install.sh harness
bash drive-hyprland/scripts/install.sh harness cursor
bash drive-hyprland/scripts/install.sh verify

harness symlinks this skill into ~/.grok/skills, ~/.claude/skills, and ~/.cursor/skills. The second argument is grok, claude, cursor, or all (default all).

verify checks hyprctl, python3, grim, wtype, ydotool, /dev/uinput, the ydotool socket, and ydotool.service.

host installs packages, adjusts /dev/uinput, and starts the user ydotool service. Run it on the Hyprland machine.

Shell syntax

find . -name '*.sh' -not -path './.git/*' -print0 | xargs -0 -n1 bash -n

PDF to Tinker data

chunk_pdf.py imports PyMuPDF before it parses arguments, so --help fails until fitz is installed.

python3 pdf-to-tinker-training/scripts/chunk_pdf.py --pdf <path-to.pdf>
python3 pdf-to-tinker-training/scripts/chunk_pdf.py --pdf <path-to.pdf> --output <out.jsonl> --pages 10
python3 pdf-to-tinker-training/scripts/to_datum.py --input <chunks.jsonl>
python3 pdf-to-tinker-training/scripts/train_lora.py --data <datum.jsonl>
python3 pdf-to-tinker-training/scripts/test_finetuned.py --checkpoint <checkpoint-name>

to_datum.py also takes --output and --model (default thinkingmachines/Inkling). train_lora.py also takes --model, --epochs, --lr, --batch_size, and --checkpoint. test_finetuned.py also takes --model, --prompt, --num_samples, --max_tokens, and --temperature.

Buzz pairing check

bash buzz-relay-self-host/scripts/verify-pair-endpoint.sh https://<buzz-fqdn>
bash buzz-relay-self-host/scripts/set-agent-avatar.sh <path-to-image.png> --name <name> --about <text>

set-agent-avatar.sh reads BUZZ_RELAY_URL and BUZZ_PRIVATE_KEY from the environment. With no arguments it prints usage and exits 2. The same is true for verify-pair-endpoint.sh.

Desktop session scripts

Both scripts exit 2 until a password is set in the environment. They are meant to run as the desktop user inside the lab image.

GNOME_RDP_PASS=<password> bash gnome-wayland-rdp/scripts/start-gnome-headless-rdp.sh
KRDP_PASS=<password> bash kde-wayland-rdp/scripts/start-krdp.sh

Optional GNOME variables in start-gnome-headless-rdp.sh: DESKTOP_USER, GNOME_WIDTH (default 1920), GNOME_HEIGHT (default 1080), GNOME_RDP_PORT (default 3389), GNOME_RDP_USER, GRD_TLS_DIR.

Optional KRdp variables in start-krdp.sh: DESKTOP_USER (default kdeuser), KRDP_USER, KRDP_PORT (default 3389), KWIN_WIDTH, KWIN_HEIGHT, KRDP_TLS_DIR, KRDP_QUALITY (default 40).

Other scripts in those folders (configure-desktop-ui.sh, promote-krdp-output.sh, start-kwin-gl.sh, and the rest) are invoked by the matching SKILL.md. Read that file before running them.

Windows helpers

Offload one job to agy, opencode, copilot, codex, or hermes:

powershell -NoProfile -ExecutionPolicy Bypass -File .\harness-offload\scripts\Invoke-HarnessOffload.ps1 -Worker opencode -Prompt "<prompt>" -Dir (Get-Location)

Optional switches on that script: -AllowTools, -Session, -Model.

Connect mstsc and save screenshots (click Connect, and read the skill before repeating a click):

powershell -NoProfile -ExecutionPolicy Bypass -File .\windows-ui-inject\scripts\Invoke-MstscConnect.ps1 -HostAddress <host> -Port <port> -User <user> -Password <password> -ShotDir $env:TEMP\rdp-shots

Optional: -RdpPath, -NoClipboard.

Usage

  1. For a graphical Linux workstation from Windows, read AGENTS.md, then agent-workstation-drive/SKILL.md. The first-session prompt is agent-workstation-drive/references/first-session-prompt.md.
  2. On a local Hyprland session, load drive-hyprland/SKILL.md and use drive-hyprland/scripts/install.sh.
  3. For any other topic, open the SKILL.md in the directory from the layout below. OpenCode topics live under opencode/. Tool skills live under pentest/tools/.
  4. Keep lab hosts, ports, users, and passwords in the environment or in the command you type. Do not commit them into SKILL.md.

Authorized assessment notes live under pentest/, defender-opsec/, and wslc-pentest-workflow/. Use them only inside a scope you are allowed to test.

Layout

Path Role
AGENTS.md Bootstrap for the workstation-drive loop
agent-workstation-drive/ End-to-end GNOME Wayland RDP plus Windows GUI drive
gnome-wayland-rdp/ Headless GNOME Shell with gnome-remote-desktop
kde-wayland-rdp/ Headless Plasma on kwin_wayland with KRdp
windows-ui-inject/ Windows mouse and keyboard injection, including mstsc
browser-read-bypass/ Alias of windows-ui-inject
drive-hyprland/ Local Hyprland workspaces, click, type, screenshots
wsl-containers/ wslc container build, run, and publish notes
podman/ Podman create, run, exec, compose, tear down
proxmox-system-administration/ Proxmox VE host administration
proxmox-vm-recovery/ Stuck UEFI/OVMF VM boot recovery
agy-harness/ Drive the Antigravity (agy) CLI
opencode-harness/ Drive the OpenCode CLI
opencode/ OpenCode install, configure, and usage skills
copilot-harness/ Drive the GitHub Copilot CLI
copilot-sdk/ Python JSON-RPC notes for the Copilot CLI
codex-harness/ Drive the Codex CLI
hermes-harness/ Drive the Hermes Agent CLI
harness-offload/ One-job dispatcher for those CLIs
ralph-wiggum-pattern/ Ralph Wiggum feedback-loop coding notes
mcp-server/ Python MCP server specification
read-canvas-browser/ Canvas and WebGL pages with an empty accessibility tree
buzz-relay-self-host/ Self-host Block Buzz, pairing, and buzz-acp
buzz-hermes-gateway/ Hermes gateway to a Buzz community
pdf-to-tinker-training/ PDF to Tinker supervised-pair scripts
pentest/ Methodology guide plus per-tool skills under pentest/tools/
defender-opsec/ Windows Defender privacy settings for an authorized lab
wslc-pentest-workflow/ Authorized pentest workflow inside WSL containers

License

Skill frontmatter sets license: MIT on every SKILL.md except pdf-to-tinker-training/SKILL.md, which has no license field. This repository has no root LICENSE file.

About

No description, website, or topics provided.

Resources

Stars

7 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages