🤖 Claude Agent Skills Stack

Self-contained skills for autonomous AI agents. Each skill bundles the instructions, validators, and reference material an agent needs to do one job well, with no dependency on the others.

Installation • Skills • Quick start • Development • Architecture

🎯 Available skills

⚡ Quick start

Skills are discovered automatically once installed. Ask in plain language:

"Create a Dockerfile for my Python FastAPI application"
"Generate a Helm chart for this container image"
"Analyze my Dockerfile for best practices"
"Set up Claude Code autoMode for this project"

Or drive the validators directly. uv run reads each script's PEP 723 header and prepares the environment, so there is no pip install step:

# Analyze a Dockerfile, then a Compose file
uv run skills/dockerfile-best-practices/scripts/analyze_dockerfile.py ./Dockerfile
uv run skills/dockerfile-best-practices/scripts/analyze_compose.py ./compose.yaml

# Validate a Helm chart
uv run skills/helm-bjw-s-chart/scripts/validate_chart.py ./my-chart/

# Inspect autoMode state, then scan for adoption candidates
uv run skills/automode-config/scripts/inspect_automode.py
uv run skills/automode-config/scripts/scan_project.py

automode-config also ships a natural-language slash command:

/automode-edit add a hard_deny rule that forbids pushing to main
/automode-edit allow deploys to the staging namespace
/automode-edit drop the soft_deny entry about npm install

The agent interprets the query, builds a full four-section proposal, shows a diff, and commits through apply_automode.py behind the critique gate, the hash gate, and an atomic flock-protected write. Full contract: commands/automode-edit.md.

📦 Installation

Recommended: the skills CLI

The skills CLI resolves any GitHub repo and wires the bundles into the right agent directories (Claude Code, Cursor, and others). Run it via npx, no global Node install required.

# Interactive: pick scope and agents
npx skills add obeone/claude-skills

# All skills, user-global (~/.claude/skills)
npx skills add obeone/claude-skills -g --all

# All skills, current project (./.claude/skills)
npx skills add obeone/claude-skills --all

# A single skill (project-scoped; add -g for global)
npx skills add obeone/claude-skills --skill dockerfile-best-practices -y

# List what the repo offers without installing
npx skills add obeone/claude-skills -l

Update with npx skills update, remove with npx skills remove, inspect with npx skills list. Full options: npx skills -h.

Claude.ai (web)

1. Download the .skill bundles from Releases
2. Go to Settings then Skills
3. Click Upload skill and select each .skill file

Manual, from a release bundle

For hosts where Node is not an option:

mkdir -p ~/.claude/skills

for skill in dockerfile-best-practices helm-bjw-s-chart automode-config apple-shortcuts; do
  curl -L "https://github.com/obeone/claude-skills/releases/latest/download/${skill}.skill" \
    -o /tmp/skill.zip && unzip -o /tmp/skill.zip -d ~/.claude/skills/
done

Swap ~/.claude/skills/ for .claude/skills/ to install project-scoped.

From source

git clone https://github.com/obeone/claude-skills.git
cp -r claude-skills/skills/* ~/.claude/skills/

Other platforms

The skills CLI auto-detects supported agents; pass --agent <name> to target one specifically (for example --agent cursor). For an unsupported agent, extract the .skill bundle into that agent's skills directory.

How it works: skills are packaged by Skill Pack on every release. The action builds .skill bundles (ZIP archives) and uploads them to the GitHub release.

🧩 Skill anatomy

Every skill follows the same layout, so an agent that knows one knows them all:

Design principles

🔍 What each skill covers

dockerfile-best-practices

• Language templates: Python/uv, Node.js, Go, Rust, PHP, Debian
• Security patterns: non-root users (UID/GID above 10000), secret mounts, SBOM and provenance attestations
• Performance: cache mounts, multi-stage builds, layer ordering
• Static analyzer: 30+ rules, including defects a green build never reveals (a --chown silently discarded to root, a healthcheck calling a binary the image does not ship)
• Compose support: V2 practices (no version: field, runtime hardening, container_name: only where you will never --scale)
• Linting, ours and everyone else's: where docker build --check, hadolint, DCLint and this repo's analyzers overlap, and where each is the only tool that catches a given defect
• Compose file naming: all four names Compose V2 accepts are valid, so the skill matches whatever the project already uses, asks once when there is no precedent, and remembers the answer

helm-bjw-s-chart

• bjw-s common library: app-template v5 patterns, with v4 legacy support
• Complete chart structure: Chart.yaml, values.yaml, common loader, NOTES.txt
• Deployment patterns: single container, sidecars, init containers, multi-controller, StatefulSets
• Beyond the basics: HorizontalPodAutoscaler, ServiceMonitor/PodMonitor, NetworkPolicy
• Chart validator: verifies structure and bjw-s compatibility

automode-config

• The four official sections: environment, allow, soft_deny, hard_deny, all arrays of prose rules with a per-section $defaults sentinel. There is no ask bucket and no plain deny bucket; those belong to permissions, and mixing them up is the mistake this skill exists to prevent.
• Critique-gated writes: claude auto-mode critique runs once per commit. A zero exit is necessary but not sufficient, because the binary sometimes exits 0 having produced no critique at all; an empty or degenerate critique fails the commit too. Raw output is archived per run.
• Atomic and reversible: per-file flock, O_EXCL write, five rolling backups, sha256 commit predicate.
• Capability auto-detection: swaps ~/.claude/settings.json transiently when the CLI lacks --settings, with signal-handler restore.
• Recovery: --repair reclaims stale flocks and restores .preview-orig orphans.
• Requires Claude Code 2.1.83+.

apple-shortcuts

• Importable output: .shortcut and plist XML for macOS and iOS
• Action catalogs: native actions plus common third-party apps
• Tooling: validator, inspector, and App Intents discovery

🔨 Development

Repository structure

.
├── README.md             # This file
├── CLAUDE.md             # Claude Code guidance for this repo
├── AGENTS.md             # Agent guidelines
├── .github/workflows/    # CI and release automation
└── skills/
    ├── dockerfile-best-practices/
    ├── helm-bjw-s-chart/
    ├── automode-config/
    └── apple-shortcuts/

Mandatory requirements

1. SKILL.md front-matter carries name, description, and metadata.version:

   ---
   name: my-skill
   description: "Skill description"
   metadata:
     version: "1.0.0"
   ---
   

2. POSIX text: every file ends with a newline (\n) and uses LF endings.

3. Python via uv, with dependencies declared inline (PEP 723).

Workflows

Versioning

Two version spaces, deliberately decoupled:

Cut a release by picking the next micro for the current month, then pushing an annotated signed tag:

git tag -s "v$(date +%Y.%m).0" -m "Publish skills"
git push origin "v$(date +%Y.%m).0"

Adding a skill

1. Create skills/<skill-name>/SKILL.md with the required front-matter
2. Add scripts/ with validators and analyzers
3. Add references/ with the deep-dive documentation
4. Add assets/ with templates if the skill needs them
5. Open a PR; validate-skills checks the structure

Migration note: helm-chart-generator became helm-bjw-s-chart (v3.0.0)

The skill was renamed to make its scope explicit: it targets the bjw-s common library, not generic Helm charts.

npx skills remove helm-chart-generator -g -y
npx skills add obeone/claude-skills -g --skill helm-bjw-s-chart -y

Anything still pointing at the old name breaks on the next install or refresh.

🏗️ Architecture

flowchart TB
    subgraph repo["claude-skills repository"]
        S1[dockerfile-best-practices]
        S2[helm-bjw-s-chart]
        S3[automode-config]
        S4[apple-shortcuts]
    end

    repo --> CI["publish-skills.yml<br/>on tag vYYYY.MM.MICRO"]
    CI --> REL[".skill bundles<br/>GitHub Releases"]

    REL --> CLI["npx skills add"]
    REL --> WEB["Upload in claude.ai"]
    REL --> MAN["curl + unzip"]

    CLI --> DIR["~/.claude/skills<br/>or ./.claude/skills"]
    WEB --> DIR
    MAN --> DIR

    DIR --> AGENT["Claude Code · Cursor · claude.ai"]

📄 License

MIT. See LICENSE.

🙏 Credits

• bjw-s common library for the Helm chart patterns
• astral-sh/uv for Python dependency management
• Docker BuildKit for modern build features
• Skill Pack for bundle packaging

Built with 🤖 by autonomous agents, for autonomous agents. The skills, the validators, and most of this documentation are written by coding agents; a human reviews and ships.

Made by Grégoire Compagnon (obeone)