Skills

Skills are YAML files that become sub-agents. Six ship built in; you can add your own by dropping YAML files into a directory and pointing SKILL_DIRS at it.

Built-in skills

Name	Display name	Focus
`security`	Security Operations	Security auditing, RBAC analysis, network policy enforcement, compliance checks.
`administration`	Cluster Administration	Deployments, scaling, rollouts, upgrades, namespace management, resource lifecycle.
`networking`	Networking Operations	Services, ingress, DNS diagnostics, network policies, connectivity troubleshooting.
`monitoring`	Monitoring & Observability	Resource usage analysis, HPA behaviour, event monitoring, Prometheus integration.
`troubleshooting`	Troubleshooting & Incident Response	Root cause analysis, crash diagnostics, scheduling failures.
`cost_optimization`	Cost Optimization & FinOps	Right-sizing, idle workload detection, orphaned resource review. See the guide.

All skills use the same four executor tools (run_kubectl, run_helm, run_shell, read_resource), so the safety and approval rules apply identically regardless of which skill is acting.

Anatomy of a skill

A skill definition has a handful of fields. Abridged from the built-in cost-optimization skill:

name: cost_optimization
display_name: Cost Optimization & FinOps
description: >
  Autonomous cost analysis — resource right-sizing, idle workload
  detection, orphaned resource cleanup, and efficiency recommendations.
category: finops

system_prompt: |
  You are a Kubernetes FinOps specialist focused on cost efficiency.
  - RIGHT-SIZING: Compare actual CPU/memory usage to requests and limits.
  - IDLE DETECTION: Find deployments with 0 ready replicas, unused
    Services, Jobs that completed long ago but weren't cleaned up.
  ...
  Rules:
  - Never delete resources without explicit confirmation.
  - Always show current vs recommended values.

documentation: |
  # Quick Reference
  - Pod usage vs requests: compare `kubectl top pods` with requests
  - PVC status: `kubectl get pvc -A` — look for non-Bound entries

system_prompt shapes how the sub-agent investigates; documentation is reference material included with the skill.

Custom skills

At startup Kopilot loads YAML files from the built-in directory plus any directories listed in the SKILL_DIRS environment variable (colon-separated on Linux/macOS, semicolon-separated on Windows):

SKILL_DIRS=/etc/kubedevaiops/skills

A skill is only activated if its name also appears in ENABLED_SKILLS. To add a custom skill named backup_review:

ENABLED_SKILLS=["security","administration","networking","monitoring","troubleshooting","cost_optimization","backup_review"]
SKILL_DIRS=/etc/kubedevaiops/skills

In-cluster, mount a ConfigMap containing your skill YAMLs and point SKILL_DIRS at the mount path. The Helm chart exposes this as the extraSkillDirs value.

AISkill CRD

The Helm chart installs an AISkill CRD (kubedevaiops.io/v1alpha1). Loading skill definitions from cluster resources is planned but not implemented yet — today skills load from YAML files only. The portable skill manifests exposed at /skills/portable declare this honestly as a planned_sources entry; see Interoperability.