AI coding agents are only as good as the context they are given. This week Docker published an open collection of agent skills on GitHub that teach assistants such as Claude Code, Cursor, Copilot, Codex and Gemini how to build, test, debug and ship containerised applications the right way. It is a small repository, but it illustrates a pattern every engineering team should be adopting: packaging expert knowledge as reusable skills that agents load on demand.
What Agent Skills Are
A skill is a self-contained folder, typically built around a SKILL.md file, that describes a task and the rules for doing it well. Compliant agents discover skills automatically from standard paths and load one only when its description matches the job at hand. That keeps the agent's context window lean while giving it deep guidance exactly when needed.
Docker's repository follows this model. Skills are written once as portable SKILL.md directories, released under the Apache 2.0 licence, and each one triggers from its own description, with no master skill to load first.
What Docker Ships
The collection is grouped by Docker product and includes:
- docker-project-foundations for initialising and structuring a Dockerised project.
- docker-build-strategies for efficient, secure and optimised image builds.
- Compose guidance for wiring multi-container stacks into maintainable configurations.
- docker-agent-config, docker-agent-run and docker-agent-deploy for authoring, running and shipping agents built with Docker Agent, including sandbox isolation, approval modes and CI regression checks.
- docker-destructive-guardrails, a cross-product policy requiring confirmation before irreversible operations.
That last skill is worth noting. With rogue-agent incidents in the headlines, a skill that forces an agent to pause before deleting volumes or pruning images is a cheap, practical safeguard.
A Five-Step Playbook for Your Team
1. Start with your most repeated mistakes
Review recent pull requests and incident notes. Where do agents, or new engineers, keep getting things wrong? Build conventions, migration steps, logging standards and deployment checklists are ideal first skills.
2. Write narrow, well-described skills
Each skill should do one job. The description is what the agent uses to decide whether to load it, so make it specific: "Use when writing or modifying Dockerfiles for Python services" beats "Docker help." Keep instructions concrete, with explicit do and don't rules.
3. Reuse vendor skills before writing your own
Official packs like Docker's encode the vendor's own best practice and will be maintained as products change. Community collections also exist, such as the full-stack-skills/docker-skills pack with 16 skills. Review any third-party skill before installing it, since a skill is effectively instructions your agent will follow.
4. Add guardrail skills early
Mirror Docker's approach with policies for your own riskiest operations: database migrations, production deploys, secret handling and deletions. A guardrail skill should tell the agent to stop, explain the action and ask for confirmation.
5. Version and test skills like code
Store skills in the repository, review changes through pull requests, and check agent output against a small set of representative tasks whenever a skill changes. Treat a regression in agent behaviour the same way you would treat a failing test.
Where Skills Fit Next to MCP
Skills and the Model Context Protocol solve different problems. MCP connects an agent to live tools and data, such as a database, a ticketing system or a container runtime. Skills tell the agent how to use those tools well. Docker ships both: its MCP Gateway plugin runs MCP servers in isolated containers, while the new skills pack supplies the know-how. The strongest setups pair the two.
Common Pitfalls to Avoid
- Overloading one skill. Giant skills bloat context and blur triggers. Split them.
- Vague descriptions. If the agent cannot tell when a skill applies, it will either ignore it or load it everywhere.
- Stale guidance. Skills that reference old tool versions actively mislead agents. Assign owners.
- Blind trust in external packs. Audit third-party skills for risky commands or network calls before adoption.
Why It Matters
Model upgrades get the headlines, but much of the day-to-day quality gap between teams using AI coding agents comes down to context. Skills turn tribal knowledge into something an agent can apply consistently, across tools, without every developer re-explaining house rules in each prompt.
Docker's move also signals where the ecosystem is heading. As more vendors publish official skills for their platforms, much as they once published SDKs and documentation, the agent that knows your stack best will increasingly be the one with the best skills installed, not necessarily the one running the largest model.
