Agent Skills: CI/CD Pipeline Generator

CI/CD Pipeline Generator

Generate a CI/CD pipeline config for the detected project stack, covering lint, test, build, security, and (when requested) deploy. Start by parsing the arguments passed to this skill invocation (platform, --deploy, --monorepo): see Phase 0 below.

Anti-hallucination guidelines

Pipeline configurations must always match actual project tooling. This is critical:

1. Discover before generating: never assume the project's tooling, check what it actually uses first.
2. Verify commands exist: check the project's own config (e.g. package.json scripts, Makefile targets) before referencing them.
3. Match real versions: use the actual language/runtime version pinned in the project's own config files, not a guess.
4. No phantom dependencies: only include services (Redis, PostgreSQL, etc.) confirmed in project dependencies.
5. Platform-specific syntax: each CI platform has distinct YAML structure, so never mix syntax between platforms.

Workflow

Phase 0: Parse Arguments

Extract configuration from the arguments passed to this skill:

Arguments:
- <platform>: Target CI platform (default: auto-detect from existing config)
 - "github" or "gh": GitHub Actions
 - "gitlab" or "gl": GitLab CI
 - "circle" or "circleci": CircleCI
 - "jenkins": Jenkins (Jenkinsfile)
- "--deploy <target>": Deployment target (optional)
 - "vercel", "netlify", "aws", "gcp", "azure", "docker", "k8s", "fly", "railway"
- "--monorepo": Generate matrix/path-filtered workflows
- No args: Auto-detect platform from existing CI config files, default to GitHub Actions

If no platform specified, detect from existing files:
- `.github/workflows/*.yml` → GitHub Actions
- `.gitlab-ci.yml` → GitLab CI
- `.circleci/config.yml` → CircleCI
- `Jenkinsfile` → Jenkins
- No CI config found → Default to GitHub Actions

Phase 1: Project Stack Detection

Explore the codebase to discover the complete project technology stack.

Phase 2: Research Best Practices (only if needed)

references/platform-patterns.md already ships current, comprehensive patterns for Node/Python/matrix/Docker/deploy across all four platforms: check it first. Only reach for WebSearch when the detected stack/platform combo isn't covered there (e.g. an unusual language or a deploy target not in the reference):

Use WebSearch:
- query: "[detected platform] CI/CD best practices [detected language] [current year]"

Use WebSearch:
- query: "[detected platform] security scanning pipeline [detected language] [current year]"

Focus on:

• Caching strategies for the detected package manager
• Recommended runner images and versions
• Security scanning tools appropriate for the language
• Deployment best practices for the target platform
• Matrix testing strategies if multiple versions needed

Phase 3: Design Pipeline Architecture

Based on discovery and research, derive the stages from what the repo actually uses. For platform-specific triggers and syntax conventions (caching included), see references/platform-patterns.md.

The generated pipeline must meet these goals. Map each to the repo's own tooling from Phase 1; skip a goal only when the repo has nothing to satisfy it, and say so in the summary:

• Fails fast: lint and type checks run first or in parallel, and a failure there cancels the slow jobs.
• Runs the test suite the repo already uses, with its real command and any service containers (database, cache) the tests need, plus coverage if configured. Use a runtime-version matrix only if the project supports several versions.
• Builds what ships (app build, Docker image if a Dockerfile exists) and keeps the artifacts later jobs need.
• Catches vulnerable dependencies and leaked secrets, plus static-analysis findings where the language has a standard tool. Scan the container image if one is built.
• Caches dependency installs keyed on the lockfile.
• Pins third-party actions, orbs and images to a SHA or exact tag, never latest or a moving branch.
• Grants least-privilege permissions: a read-only token by default on GitHub Actions, widened per job only where a step needs it.
• Deploys only when --deploy is given or deployment config is detected: staging and production are separate, a health check follows the deploy, and the built artifact reference is threaded into the deploy step. The Build job must expose the image tag/digest it produced (job outputs, GITHUB_OUTPUT, an artifact file, etc.), and the deploy step must consume that same reference: rendering it into a task definition, helm upgrade --set image.tag=<ref>, kubectl set image deployment/<name> <container>=<ref>, or equivalent. Never emit a blind restart (aws ecs update-service --force-new-deployment, kubectl rollout restart with no image change) as the whole deploy step: if the target pins an image tag/digest, a blind restart just re-pulls the OLD image and ships nothing new.

Design decisions:

• Parallelism: lint, test, and security scan run in parallel when possible.
• Branch strategy: main/master triggers deploy, while PRs trigger lint+test+build.
• Artifacts: build outputs passed between stages where needed.

Phase 4: Generate Pipeline Configuration

Generate the complete CI/CD configuration file based on the designed architecture.

For platform-specific syntax and patterns, consult:

• references/platform-patterns.md - Detailed YAML patterns for each platform

File locations by platform:

| Platform | File Path | |----------|-----------| | GitHub Actions | .github/workflows/ci.yml | | GitLab CI | .gitlab-ci.yml | | CircleCI | .circleci/config.yml | | Jenkins | Jenkinsfile |

Generation guidelines:

1. Use discovered commands exactly (do not invent scripts)
2. Pin action/orb/image versions (see the pinning goal in Phase 3)
3. Include inline comments explaining non-obvious configuration
4. Set appropriate timeouts for each job
5. Use environment variables for configurable values
6. Follow the platform's recommended project structure

If an existing CI config exists:

• Read the existing file first
• Ask the user whether to replace or augment
• Preserve any custom configuration the user has added
• Merge new stages with existing ones where appropriate

Phase 5: Validate & Present

Step 5.1: Validate, Fix, Re-validate

Run the platform's validator on the generated file, fix every reported error in the file, and run it again. Stop when it is clean or after 3 rounds, then report what still fails.

| Platform | Validator | Install | |----------|-----------|---------| | GitHub Actions | actionlint <workflow_file> | brew install actionlint | | GitLab CI | glab ci lint | brew install glab | | CircleCI | circleci config validate | brew install circleci | | Jenkins | the declarative linter, if the Jenkins instance is reachable (curl -X POST -F "jenkinsfile=<Jenkinsfile" $JENKINS_URL/pipeline-model-converter/validate) | none |

If no validator is available, say so in the summary and do the minimum YAML parse for YAML platforms:

python3 -c 'import yaml,sys;yaml.safe_load(open(sys.argv[1]))' <config_file>

A Jenkinsfile is Groovy, not YAML: without the linter, re-read it for unbalanced braces and quotes.

Step 5.2: Cross-Reference Check

Verify all referenced commands and paths exist:

1. Every script/command in the pipeline exists in the project
2. Every referenced file path is valid
3. Service versions match project requirements
4. Environment variable names are consistent
5. If a deploy stage exists, its deploy step references the image tag/digest (or equivalent build artifact) produced by the Build stage, not a blind restart with no reference to what was just built

Step 5.3: Present Summary

Output a summary including:

• Pipeline architecture overview
• Stages and their purposes
• Trigger conditions (which branches, PR events)
• Required secrets/environment variables to configure
• Cache strategy explanation
• Any manual steps needed (e.g., setting up deployment secrets)

Do not estimate run time per stage: it hasn't run yet and any number would be a guess.

Handling Ambiguity

If encountering unclear requirements:

1. Ask the user to clarify platform choice, deployment target, or branching strategy.
2. Present options with trade-offs when multiple valid approaches exist.
3. Default to the most common configuration for the detected stack.

Usage Examples

# Auto-detect everything
ci-generate

# Specify platform
ci-generate github
ci-generate gitlab
ci-generate circleci
ci-generate jenkins

# With deployment target
ci-generate github --deploy vercel
ci-generate gitlab --deploy docker
ci-generate github --deploy aws

# Monorepo support
ci-generate github --monorepo

# Combined options
ci-generate github --deploy k8s --monorepo

Important Notes

• Discover first: never assume project tooling, always run Phase 1.
• Pin versions: use specific versions for actions, orbs, images, and tools.
• Secrets documentation: list all required secrets so users know what to configure.
• Existing config: always check for and respect existing CI configuration.
• Security by default: include dependency scanning and secret detection in every pipeline.
• Cache effectively: proper caching can reduce CI times by 50-80%.