dataverket-skills

Skills for coding agents. One folder per skill: SKILL.md, and scripts/, references/, evals/ where the skill needs them.

Skill	What it does
bash-style	Bash style for scripts committed to a repo, after postmodern's ruby-install and chruby: tabs, function blocks, `
playwright	Browser automation in a persistent container. Screenshots, computed styles, DOM and network checks.
zmx	Long-lived processes in detached sessions: dev servers, watchers, databases. Start, wait for ready, scan output, kill by project.

Install

This repo is the canonical copy of each skill. Agents read symlinks to it from their own skill folders.

Folder	Read by
~/.claude/skills	Claude Code
~/.agents/skills	Other agents. Claude Code does not load skills from it.

git clone ssh://git@git.dataverket.org/dataverket/skills.git
cd skills
task link     # first run: symlink every skill into both folders

Afterwards:

task          # help: the list of tasks
task list     # skills and symlink status per folder
task update   # git pull, then link
task link     # add missing symlinks, remove stale ones

task link leaves real folders and symlinks to other places alone. Needs Task.

Conventions

• SKILL.md under 150 lines. Longer material in references/, pointed to from the body.
• Skill scripts: no host dependencies beyond podman or docker. Versions pinned in the script.
• bin/ follows bash-style, in-repo layout. Verify with its shellcheck command for bin/ alone, from the repo root.
• Taskfile.yml lists commands. A task that needs conditionals, loops or computed variables becomes a script in bin/.
• Comments: facts, decisions, references. No reasoning prose.
• evals/evals.json: test prompts for the skill-creator loop.
• Verify a skill before commit: shellcheck scripts/*.sh, node --check scripts/*.js, one real run.