DevOps运维技能包

claude-ops-skills

Claude Code slash-command skills for DevOps daily operations. For DevOps engineers and SREs who manage Kubernetes clusters, Helm charts, Terraform IaC, ArgoCD GitOps, and blockchain infrastructure — run read-only inspections, security audits, and best-practice research without leaving the terminal.

All operations are READ-ONLY. Mutating commands are blocked at the settings level and never executed.

Infrastructure skills assume the cluster/server is not directly reachable from the session: they emit copy-paste read-only command bundles for the operator to run, then interpret the pasted output. Chart/IaC/manifest work runs locally on files.

Table of Contents

• Quick Start
• Skills
• External Plugins & Skills
• Prerequisites
• Structure
• Safety
• Reference

Quick Start

git clone https://github.com/seokheejang/claude-ops-skills.git
cd claude-ops-skills
make install

Then restart Claude Code and try:

/k8s-ops my-cluster          # Inspect a K8s cluster
/helm-ops my-release          # Check Helm release status
/best-practice EKS CNI choice # Research industry best practices

Other commands:

make update     # git pull + reinstall
make uninstall  # Uninstall
make test       # Run PreToolUse hook tests

Skills

External Plugins & Skills

Third-party Claude Code resources kept external where possible, so upstream updates arrive by git pull instead of going stale in this repo. Vendored only when the copy has to be patched to work here — see Vendored below.

Plugins — consumed via /plugin (installed manually, not handled by install.sh):

Skills (no marketplace.json → can't use /plugin) — declared in configs/external-skills.tsv, reproduced by install.sh (clone → npm → symlink, cloned outside the repo, updated via git pull):

No active entries. The mechanism stays in place — add one TAB-separated line to configs/external-skills.tsv and re-run install.sh to register a new external skill. (Previously carried archify, disabled 2026-07-09 — see docs/external-plugins.md.)

Vendored — copied into skills/ because the files had to be modified. These do not receive upstream updates:

Why vendored instead of an external-skills.tsv entry: we reach Notion through the claude.ai connector, not the plugin's bundled MCP server, so the upstream tool names do not resolve here — the files need patching, which a clone-and-symlink cannot carry. Upstream has no LICENSE file. The router SKILL.md is ours; the full change list is in its 출처 section.

→ Install commands, update notes, and role-vs-builtin guidance: docs/external-plugins.md

Prerequisites

Only Claude Code and make are required. Other tools are needed only for their respective skills.

Structure

├── skills/           # Symlinked to ~/.claude/skills/
│   ├── k8s-ops/      # K8s inspection — emits kubectl command bundles
│   ├── k8s-craft/    # K8s manifest authoring & design (paired with k8s-ops)
│   ├── k8s-security/ # K8s security audit skill
│   ├── ssh-ops/      # SSH inspection skill
│   ├── troubleshoot/ # Evidence-based failure-diagnosis discipline (no-speculation, manual-remediation guardrails)
│   ├── db-ops/       # DB query skill
│   ├── helm-ops/     # Helm chart inspection + authoring guide
│   ├── terraform-ops/ # Terraform state/plan + IaC authoring guide
│   ├── argocd-ops/   # ArgoCD monitoring + GitOps manifest authoring
│   ├── sh-craft/     # Shell script authoring — bash default, POSIX fallback for minimal pod images
│   ├── ralph/        # Self-review loop (iterative verification, any task type)
│   │   └── references/ # Domain checklists (code/IaC/analysis/docs) + FAIL/NOTE policy
│   ├── catchup/      # Session context restore (git + CLAUDE.md/README/docs)
│   ├── best-practice/ # DevOps best practice research
│   │   └── references/ # 3 topics (sources, verification, output)
│   ├── grill-me/     # Plan/design stress-test via relentless interview
│   ├── write-a-skill/ # Meta-skill: scaffold new skills with conventions
│   ├── transfer-en/  # Korean → simple natural English (single-file skill)
│   └── notion-report/ # Work output → Korean Notion page (router SKILL.md, effort: high)
│       ├── knowledge-capture/       # ← vendored from makenotion/claude-code-notion-plugin @9847f2a
│       ├── meeting-intelligence/    #   tool names rewritten to mcp__claude_ai_Notion__*,
│       ├── research-documentation/  #   SKILL.md → GUIDE.md, evaluations/+examples/ dropped
│       └── spec-to-implementation/  #   (see the 출처 section in notion-report/SKILL.md)
├── configs/          # Configuration templates
│   ├── settings.json.template     # allow/deny rules
│   ├── claude.md.template         # CLAUDE.md merge content
│   ├── external-skills.tsv        # External skills to clone + symlink (not plugins)
│   └── settings.local.json.example
├── scripts/          # Install/management scripts
│   ├── install.sh    # Backup → symlink → merge
│   ├── uninstall.sh
│   └── update.sh
├── templates/        # Templates for new skills
└── _disabled/        # Skills kept in repo but excluded from install (not under skills/*)
    ├── mmdraw/       # Mermaid diagram generator — re-enable with `mv _disabled/mmdraw skills/`
    └── humanize-ko/  # Superseded by humanize-korean plugin (see docs/external-plugins.md)

// Delete this block in ~/.claude/settings.json → hooks → Stop
"Stop": [
  {
    "hooks": [
      {
        "type": "command",
        "command": "FRONT=$(osascript -e ...)  ..."
      }
    ]
  }
]

Safety

• All operations are READ-ONLY — no system-affecting commands are executed
  • K8s: get, describe, logs, top, exec (inspection only)
  • Helm: list, status, get, show, lint, template, history, diff (no install/upgrade/delete)
  • Terraform: state list/show, plan, validate, fmt -check, output (no apply/destroy/init)
  • ArgoCD: app get/list/diff/logs/history, repo list, cluster list (no sync/create/delete)
• Mutating commands are blocked by deny rules in settings.json (the enforcement layer) and restated in each SKILL.md
• When changes are needed, the skill provides commands as text guidance only — never executes them
• Chart/IaC authoring uses Write/Edit tools to create files locally — no infrastructure impact
• Sensitive data (kubeconfig contents, SSH keys) is never included in this repo
• Cluster contexts are supplied by the operator via --context — no cluster inventory is stored in this repo

Reference

DevOps Skills

• antonbabenko/terraform-skill - Terraform/OpenTofu best practices skill (testing, modules, CI/CD, production patterns)
• akin-ozer/cc-devops-skills - 31 DevOps skills (Helm/Terraform/K8s generator+validator pairs, severity classification)
• ahmedasmar/devops-claude-skills - GitOps workflows (ArgoCD 3.x, Flux, drift detection, ApplicationSet patterns)

Diagram & Visualization

• mermaid-to-excalidraw - Mermaid → Excalidraw web converter

Compound Engineering

• EveryInc/compound-engineering-plugin - Official Compound Engineering plugin (Plan→Work→Review→Compound)
• Compound Engineering: How Every Codes With Agents - Original article (Kieran Klaassen, 2025)

Claude Code Ecosystem

• travisvn/awesome-claude-skills - Curated list of Claude Skills, resources, and tools
• hesreallyhim/awesome-claude-code - Skills, hooks, slash-commands, agent orchestrators
• alirezarezvani/claude-skills - 180+ production-ready skills for AI coding agents

Architecture & Patterns

• affaan-m/everything-claude-code - Production-grade AI harness (agents, skills, hooks, AgentShield, instincts)
• wshobson/agents - Claude Code plugin/agent architecture
• claude-code-auto-approve - PreToolUse hook with compound command parsing (shfmt AST)