skills

Agent skills shared across repositories. Written for Claude Code, and deliberately free of anything tying them to one repository, so they are useful outside the repos they came from. Some cover shared infrastructure (a self-hosted CI pool, a multi-GPU host); others cover work that is not infrastructure at all, which is why they are grouped into plugins by subject.

What is here

All are host-independent. They name no repository, no organization, and no host values, which is what makes them shareable at all. Anything genuinely specific to one repository belongs in that repository's AGENTS.md, and the skills defer to it by name rather than guessing.

Two ways to consume them

Pick one deliberately. They differ in who gets the skills, not in what the skills say.

They coexist without colliding, because plugin skills are namespaced and vendored ones are bare. The subtler hazard is that both can be listed and invocable under different names with different content, whenever a vendored copy has fallen behind. That is invisible in a way a collision is not, which is what the drift check below exists to close.

Option 1: install as a plugin

/plugin marketplace add speediedan/skills
/plugin install common-infra-skills@speediedan-skills

Nothing else is needed. Installs are recorded per user rather than per clone, so a colleague cloning your repository will not have them.

Option 2: vendor the skills into a repository

Copy the skill directories into the consuming repo's .claude/skills/, so they travel with a clone:

cp -r plugins/common-infra-skills/skills/gpu-lease         /path/to/repo/.claude/skills/
cp -r plugins/common-infra-skills/skills/az-pipelines-ops  /path/to/repo/.claude/skills/

Then adopt the drift check below, because a vendored copy is a copy and copies diverge.

Keeping vendored copies honest

Two different things can go wrong, they need different checks, and the split is by who can act on the failure.

Someone edited the vendored copy

Actionable by whoever trips it, and it needs nothing but the local files. A local edit to a vendored skill is silently discarded the next time the copy is refreshed: the person who wrote it loses the work and nobody finds out why.

consumer/check_vendored_skills.py is a pre-commit hook that catches exactly this. To install it:

1. Copy consumer/check_vendored_skills.py into the consuming repo, conventionally scripts/.

2. Add a manifest at .claude/skills/.shared-skills.sha256, in sha256sum format:

   # Vendored from https://github.com/speediedan/skills (plugin: common-infra-skills).
   # Regenerated when the vendored copies are refreshed. Do not hand-edit.
   <sha256>  .claude/skills/gpu-lease/SKILL.md
   <sha256>  .claude/skills/az-pipelines-ops/SKILL.md
   

   Generate it with sha256sum .claude/skills/*/SKILL.md from the repo root.

3. Wire the hook in .pre-commit-config.yaml:

     - repo: local
       hooks:
         - id: vendored-skills-unmodified
           name: Vendored skills are not edited here
           entry: python3 scripts/check_vendored_skills.py
           language: system
           files: ^\.claude/skills/
           pass_filenames: false

Whatever refreshes the vendored copies must regenerate the manifest in the same operation. This is the one requirement that makes the hook work rather than backfire: without it every legitimate refresh trips the hook, and a hook that fires on correct behavior teaches people to pass --no-verify, which disables every other hook in the repo.

A missing manifest is treated as a failure rather than a skip, deliberately. A skip would be indistinguishable from a clean pass, which is the ambiguity the check exists to remove.

Upstream has moved ahead of the vendored copy

Not actionable by a contributor and not detectable offline, so it is deliberately not in the pre-commit hook: that would mean a network call on every commit, or a hook that fails on a clean clone, which is the --no-verify problem again.

Check it when you choose to, by diffing against a clone of this repository:

diff -r plugins/common-infra-skills/skills/gpu-lease  /path/to/repo/.claude/skills/gpu-lease

If you maintain several consuming repos, script that comparison once rather than remembering it. The useful shape is a status mode that reports every consumer's drift in one run, refuses when its own clone of this repository is stale (otherwise every consumer looks drifted at once, which is a precondition failure wearing a content failure's costume), and distinguishes formatting-only differences from real ones.

Contributing back

If you fix something in a vendored copy, send it here rather than leaving it local. A local fix is discarded by the next refresh, and every consumer keeps the bug.

Skills in this repository are checked on commit by one gate, so an edit that breaks a property will be refused rather than published:

• Structure + spec: the marketplace manifest resolves, every plugin has a plugin.json whose name agrees, and every SKILL.md opens and closes its frontmatter with a spec-valid name (matching its directory, lowercase alphanumeric with single hyphens) and description (1-1024 chars).
• Neutrality: no repository, organization, host, or credential names in any file under plugins/. This is what makes them publishable, and once pushed it cannot be undone, so it is a gate rather than a report.
• Normalization: each SKILL.md is a fixed point of mdformat==0.7.17 with mdformat-gfm and mdformat_frontmatter (separate script, scripts/check_normalized.sh, since it skips when the formatter is unavailable). At least one consuming repo runs that formatter over its own copy, so an unnormalized file here would be rewritten on contact there and re-create drift on every commit.

Run them yourself with scripts/check_skills.sh and scripts/check_normalized.sh.

Requirements

gpu-lease describes working with a host-wide lease and assumes one exists, with GPU_LEASE_CMD pointing at it. The lease implementation is host and operator infrastructure and is not distributed here. The skill is a complete no-op when GPU_LEASE_CMD is unset, so installing it on a machine without a lease costs nothing.

az-pipelines-ops assumes an Azure DevOps project and a self-hosted agent pool. Its commands are parameterized on ${ORG} and ${PROJECT}; supply those from your own repository's AGENTS.md.

Licence

Apache-2.0. See LICENSE.

{
  "$schema": "https://code.claude.com/schemas/marketplace.json",
  "name": "speediedan-skills",
  "owner": {
    "name": "speediedan",
    "url": "https://github.com/speediedan"
  },
  "plugins": [
    {
      "name": "common-infra-skills",
      "source": "./plugins/common-infra-skills",
      "description": "Host-independent infrastructure skills shared across repos that run CI on a self-hosted agent pool"
    },
    {
      "name": "docs-authoring-skills",
      "source": "./plugins/docs-authoring-skills",
      "description": "Skills for authoring documentation that must render correctly on more than one engine"
    }
  ]
}