agent-skills

Open-source agent skills for shipping and operating real infrastructure. Each skill is a self-contained SKILL.md plus the references, scripts, and templates an agent needs to do the job well. They trigger on intent — you describe what you want, the relevant skill fires.

Cross-agent (Claude Code, Cursor, Codex, OpenCode, …) via the npx skills CLI, and installable as a full Claude Code plugin.

Skills in this collection

More to come.

ec2-docker-host at a glance

Treats one EC2 instance as a small self-hosting platform. Four modes:

• Provision a new host from nothing (AWS CLI: instance, Elastic IP, security group, a dedicated data

volume, Docker, the shared DB + proxy stack, backups).

• Adopt an existing host — read-only discovery first, then a conformance read, so you never overwrite

a neighbor's workload.

• Deploy a project (monorepo or polyrepo) as its own isolated unit: unique host port, own Compose

project, own database + role, own secret namespace, an additive proxy vhost, validate-then-reload.

• Maintain — logical DB dumps + volume snapshots + off-host sync, restore drills, deliberate upgrades.

Defaults are lightweight (any OCI registry + host-side env files); an AWS-native path (ECR + SSM + GitHub OIDC) is documented as an opt-in upgrade.

Installation

Two install paths — pick what fits.

Path A — npx skills add (recommended for quick install)

# Install every skill in this repo to ~/.claude/skills/ for Claude Code, globally
npx skills add juncoding/agent-skills -g -a claude-code

# Browse what's in the repo without installing
npx skills add juncoding/agent-skills --list

# Install just one specific skill
npx skills add juncoding/agent-skills --skill ec2-docker-host -g -a claude-code

Path B — install as a Claude Code plugin

Inside Claude Code, run /plugin and add this source:

github:juncoding/agent-skills

Or edit ~/.claude/settings.json (user-level) / .claude/settings.json (project-level) directly:

{
  "plugins": {
    "agent-skills": {
      "source": "github:juncoding/agent-skills"
    }
  }
}

Restart Claude Code; the skills become available and trigger on intent.

Path C — clone locally (for iterating on the skills themselves)

git clone [email protected]:juncoding/agent-skills.git ~/Dev/agent-skills

Then point a plugin source at the local path:

{
  "plugins": {
    "agent-skills": {
      "source": "/Users/you/Dev/agent-skills"
    }
  }
}

⚠️ Safety

ec2-docker-host bundles shell scripts that SSH into a host and run `docker`, `psql`, and `caddy` with `sudo`. They are written to be additive and neighbor-safe (validate-then-reload the proxy, pull images before swapping, scope --remove-orphans per app), but they act on real infrastructure.

• Read a script before you run it the first time on a host.
• Test against a throwaway host before pointing them at production.
• The skill always discovers live host state before changing anything, and re-verifies neighbors after —

trust that flow, and stop if discovery shows something unexpected.

• Provided as is, without warranty (see LICENSE). You are responsible for what runs against

your servers.

Repository layout

.claude-plugin/plugin.json   # Claude Code plugin manifest
skills/
  ec2-docker-host/
    SKILL.md                 # the skill (router + safety doctrine + isolation invariants)
    references/              # per-mode deep dives
    scripts/                 # idempotent, neighbor-safe ops scripts
    assets/                  # fill-in templates (compose, vhost, entrypoint, host config)
    evals/                   # scenario test cases

Contributing

Add a skill as a new directory under skills/<name>/ with a SKILL.md (a name + a specific, trigger-rich description in the frontmatter) and any references/scripts/assets it needs. Keep skills self-contained and intent-triggered. PRs welcome.

License

MIT © Bill Zhou