Docker ships 11 open-source skills for AI coding agents

Your coding agent has read a million Dockerfiles on the internet, most of them bad. That is why it writes a four-layer image that ships your .git directory. Docker's answer is to stop hoping the model guessed right: the company published Docker Skills, an open-source set of 11 agent skills that install into whatever coding agent you already use.

What actually happened

Docker released the collection at github.com/docker/skills under Apache License 2.0, with documentation on Docker Docs. Installation is one command through the skills CLI:

npx skills add docker/skills

The 11 skills are grouped by product area:

• Build and Dockerfile: docker-project-foundations, docker-build-strategies
• Compose: docker-compose-patterns
• Sandboxes: docker-sandboxes-lifecycle, docker-sandboxes-network-credentials, docker-sandboxes-env, docker-sandboxes-kits
• Docker Agent: docker-agent-config, docker-agent-run, docker-agent-deploy
• Cross-product: docker-destructive-guardrails

Supported agents include Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI, Google Antigravity, OpenCode, plus Docker's own Agent and Sandboxes. That list is the point. Skills are plain SKILL.md directories on disk, discovered through standard paths, so one authored file works across every compliant agent instead of once per vendor. The repo uses main as a rolling channel with tags as immutable snapshots — pin a tag if you want reproducible agent behavior.

Why portable agent skills matter for your business

Vendor documentation is becoming executable context. Docker did not write a blog post about best practices. It shipped the practices in the format an agent loads at runtime. Expect your cloud provider, your payment processor and your database vendor to do the same, because it is cheaper than answering support tickets about generated code that does not work.

docker-destructive-guardrails is the one to read first. An agent with shell access and a Docker daemon can remove volumes. A skill that tells it which commands are one-way is the kind of control that belongs in the agent's context, not in a policy document nobody loads. If you run agents against real infrastructure, this is the file to read before you install the rest.

The real lesson is to write your own. Docker knows Docker. It does not know your deployment target, your naming conventions, your staging environment, or the three things that break when someone touches your Compose file. That knowledge currently lives in a senior engineer's head and gets re-explained in every chat session. Written once as a SKILL.md in your repo, it loads automatically for every agent and every teammate. We do this for client stacks and it is the single highest-leverage hour of agent setup there is.

Apache 2.0 and plain files means no lock-in to audit. You can read every skill before it reaches your agent's context, fork the ones that do not match your setup, and pin a tag so a rolling main does not change how your agent behaves mid-sprint. Do that last part. An agent whose instructions silently updated is an agent whose output you cannot reproduce.

Sources: Docker Docs, docker/skills on GitHub.