The skills

The skills: 29 guides, two families

What a skill is

A directory under ~/.claude/skills/ holding a SKILL.md: frontmatter with a name and a description, then the content. Some skills carry extra material in references/, scripts/ or examples/ subdirectories.

The description is the part that matters most operationally. Claude Code reads it to decide whether the skill is relevant, so a vague description means a skill that never loads. The ones here all name their trigger keywords explicitly.

Family 1: the fourteen DevOps guides

Reference material for infrastructure work. Each one follows the same shape: an operating workflow in numbered steps, a recommended project structure, versions to pin, a table of pitfalls with symptom and remedy, and reference commands.

Skill	Domain
terraform-guide	Infrastructure as Code, modules, state, backends
helm-chart-builder	Helm charts, templates, values, releases
ansible-playbook-builder	Playbooks, roles, inventories, vault
docker-swarm-guide	Services, stacks, overlay networks, secrets
argocd-guide	GitOps, sync, App of Apps, ApplicationSets
prometheus-grafana-setup	Metrics, alerting, PromQL, dashboards
azure-cloud-advisor	Choosing Azure services, architecture, costs
github-actions-expert	Workflows, matrix builds, OIDC, reusable workflows
gitlab-ci-guide	Pipelines, stages, runners, artifacts
bash-script-builder	Stream contract, guard rails, idempotence, exit codes
makefile-builder	Orchestration entry point, exported config, output contract
logging-setup	Console output format, run journal, redaction, rotation
metrics-setup	Naming, metric types, cardinality, RED and USE
tracing-setup	Propagation, span naming, sampling, correlation

These carry dated knowledge on purpose. A model answers from its training when it has no fresher source, and training data ages: it will pin a Terraform version that was current two years ago. The skills exist to override that, which is why they name exact version constraints and call out recent deprecations, such as native S3 locking replacing dynamodb_table, or the removal of the public Bitnami chart catalogue.

That also means they need maintenance. A skill full of stale versions is worse than no skill, because it is trusted.

Family 2: the fifteen method skills

These do not describe a technology, they describe how to work. Most are adapted from obra/superpowers.

Skill	What it enforces
brainstorming	Explore intent and requirements before building anything
writing-plans	Turn a spec into a written plan before touching code
executing-plans	Run a written plan with review checkpoints
subagent-driven-development	Execute independent tasks of a plan in one session
dispatching-parallel-agents	Split genuinely independent work
test-driven-development	Red, green, refactor, before implementation code
systematic-debugging	Diagnose before proposing a fix
requesting-code-review	Check work against requirements before merging
receiving-code-review	Verify feedback rather than agreeing performatively
verification-before-completion	Run the commands before claiming success
finishing-a-development-branch	Decide how work gets integrated
using-git-worktrees	Isolate feature work from the current workspace
using-superpowers	Find and invoke the right skill in the first place
writing-skills	Create and verify new skills
clean-user-facing-text	Audit and polish text meant for readers

The one that matters most is verification-before-completion. The failure it targets is an agent announcing that something works without having run anything, which is the single most expensive kind of wrong answer: it looks finished.

How they trigger

Claude Code matches the request against the skill descriptions and loads what fits. In practice, three things make that reliable:

1. Descriptions that name the keywords. Every skill here lists its triggers, which is why the DevOps ones say things like "also triggers on terraform, tfstate, HCL".
2. A routing reminder in CLAUDE.md. The # SKILLS section states that DevOps work means consulting the relevant skill before proposing commands or configuration, and that development work means following the brainstorm, plan, TDD and review sequence.
3. Explicit invocation. A skill can always be called by name when the automatic match does not happen.

This is worth contrasting with Mistral Vibe, where skills do not trigger on keywords at all and the whole routing has to be forced through a custom system prompt. See the Your-Mistral-DevOps wiki for how that plays out.

Adding your own

Use the writing-skills skill, which covers creating one and verifying it actually loads. The minimum bar: a description that names the situations and keywords that should trigger it, and content with a concrete method rather than generalities. A skill that says "use best practices" helps nobody and costs tokens.