agentleFS
KhaledSaeed18KhaledSaeed18/dotclaude/skills/engineering/dockerfile-best-practices/SKILL.md

dockerfile-best-practices

Skill5 starsChanged 4 months agoMITView on GitHub

Write or review a Dockerfile and compose setup for a production service: multi-stage builds that ship only the runtime, pinned base images, layer order for cache hits, non-root user, minimal image size, correct signal handling and health checks, secrets kept out of layers, and language-specific patterns for Node, Python, Go, and Java. Use when containerising a service, when images are large or slow to build, when a container ignores SIGTERM, or when a security scan flags the image.

  • Reads credentials
  • Deletes or force-pushes
  • Installs packages

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 1 is distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Prefer -slim or distroless over alpine for glibc-dependent runtimes (Node native modules, Python wheels).

    line 41
  2. Very specific:

    No secrets in ENV, ARG, or copied files.

    line 52
  3. Very specific:

    Read-only root filesystem where possible (--read-only with tmpfs for /tmp).

    line 53
  4. Very specific:

    Check with docker image ls and dive.

    line 64
  5. Distinctive:

    Do not install dev dependencies, docs, or test files in the runtime stage.

    line 65

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

A production image contains exactly what the process needs to run, built reproducibly, running as a non-root user, and stopping cleanly when asked. Everything below follows from those four properties.

Structure: multi-stage

# syntax=docker/dockerfile:1.7
FROM node:22.11-bookworm-slim AS deps
WORKDIR /app
COPY package.json pnpm-lock.yaml ./
RUN corepack enable && pnpm install --frozen-lockfile --prod=false

FROM deps AS build
COPY . .
RUN pnpm build && pnpm prune --prod

FROM node:22.11-bookworm-slim AS runtime
ENV NODE_ENV=production
WORKDIR /app
RUN groupadd -r app && useradd -r -g app -d /app app
COPY --from=build --chown=app:app /app/node_modules ./node_modules
COPY --from=build --chown=app:app /app/dist ./dist
COPY --chown=app:app package.json ./
USER app
EXPOSE 3000
HEALTHCHECK --interval=30s --timeout=3s --start-period=10s CMD node dist/healthcheck.js
ENTRYPOINT ["node", "dist/server.js"]

Build stages hold compilers and dev dependencies; the runtime stage copies artifacts only.

Rules

Base images

  • Pin to a specific version tag (22.11-bookworm-slim), and for reproducible builds pin the digest (@sha256:...) with a tool (Renovate, Dependabot) to bump it.
  • Prefer -slim or distroless over alpine for glibc-dependent runtimes (Node native modules, Python wheels); alpine for Go static binaries is fine.
  • One process per container; no supervisord.

Layers and cache

  • Order from least to most frequently changed: base, system packages, dependency manifests, dependency install, source, build. Copy lockfiles before source so dependency layers cache.
  • RUN steps combined with && and cleanup in the same step (apt-get install ... && rm -rf /var/lib/apt/lists/*); a later rm does not shrink an earlier layer.
  • .dockerignore with .git, node_modules, build output, .env*, tests, docs; without it the build context and the cache invalidation are both wrong.
  • BuildKit cache mounts for package managers: RUN --mount=type=cache,target=/root/.npm npm ci.

Security

  • USER non-root before ENTRYPOINT; files copied with --chown. Bind to a port above 1024.
  • No secrets in ENV, ARG, or copied files; they persist in layers and history. Use RUN --mount=type=secret,id=npmrc for build-time secrets and runtime injection (env at run, mounted files, a secrets manager) for the rest.
  • Read-only root filesystem where possible (--read-only with tmpfs for /tmp); drop capabilities; no --privileged.
  • Scan (docker scout, trivy image) in CI; fail on critical CVEs with a fix available.

Runtime behaviour

  • ENTRYPOINT in exec form (JSON array) so the process is PID 1 and receives signals; CMD for default arguments only. Shell form wraps in sh -c and swallows SIGTERM.
  • The application handles SIGTERM: stop accepting, drain in-flight requests, close connections, exit within the orchestrator's grace period (default 30 s). If the runtime cannot (some shell scripts), use tini as the init (ENTRYPOINT ["tini", "--", ...]).
  • HEALTHCHECK (or the orchestrator's probe) hits a cheap endpoint that checks dependencies the process needs; readiness separate from liveness in Kubernetes.
  • Logs to stdout/stderr, unbuffered (PYTHONUNBUFFERED=1); no log files in the container.
  • EXPOSE documents the port; it does not publish it.

Size

  • Check with docker image ls and dive. Typical targets: Go static under 20 MB (distroless/static), Node under 200 MB, Python under 250 MB.
  • Do not install dev dependencies, docs, or test files in the runtime stage.

Language notes

  • Node: npm ci / pnpm install --frozen-lockfile; NODE_ENV=production; node directly, not npm start (npm does not forward signals); set --max-old-space-size from the container limit or use a version that respects cgroups.
  • Python: python:3.12-slim; pip install --no-cache-dir -r requirements.txt (or uv sync --frozen) in a build stage into a venv copied to runtime; PYTHONDONTWRITEBYTECODE=1; run with gunicorn/uvicorn directly; --workers from CPU limit.
  • Go: build stage golang:1.23, CGO_ENABLED=0 go build -ldflags="-s -w", runtime gcr.io/distroless/static-debian12:nonroot, copy the binary and CA certs.
  • Java: build with the JDK, run on a JRE (eclipse-temurin:21-jre), layered jar (spring-boot:build-image or jarmode=layertools), -XX:MaxRAMPercentage=75.

Compose for local development

compose.yaml with the service built from the Dockerfile's build stage (target: build), source bind-mounted for hot reload, dependencies (db, cache) with named volumes and health checks, depends_on with condition: service_healthy, env from .env (git-ignored) with an .env.example committed.

Review checklist

Pinned base; multi-stage; lockfile copied before source; .dockerignore; non-root; exec-form ENTRYPOINT; signals handled; health check; no secrets in layers (docker history shows none); scan clean; size within target; builds reproducibly twice with the same digest.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.