Skill

DeployGuard Guard Skill

Validates Kubernetes manifests and Terraform configs against the DeployGuard policy ruleset. Acts as a senior reviewer: every violation explains why it matters and what to fix, not just what's wrong.

Trigger: /guard

---

Installation

~/.claude/skills/guard/SKILL.md     # global (available in all projects)
.claude/skills/guard/SKILL.md       # project-local

Prerequisites: deployguard package must be importable in the Python environment where Claude Code is running.

pip install dg-deploy            # from PyPI
# or from the repo:
pip install -e /path/to/DeployGuard

Run dg doctor to confirm the package and its tools are ready.

---

What it checks

Rule	Default level
Missing resources.requests / resources.limits	ERROR
Missing livenessProbe / readinessProbe	ERROR
Missing pod securityContext.runAsNonRoot	ERROR
:latest or untagged image	ERROR
runAsUser: 0 (root)	ERROR
privileged: true	ERROR
automountServiceAccountToken not disabled	WARN
Malformed YAML (yaml_parse)	ERROR
Schema violations (kubeconform)	ERROR (when installed)
Misconfigs (trivy)	ERROR/WARN (when installed)

---

Instructions for Claude

When the user invokes /guard [args]:

Step 1 — Determine what to guard

• Path argument (e.g. /guard k8s/ or /guard deployment.yaml): use that path.
• Pasted manifest (YAML block in the same message, no path given): write the YAML to a temp file (/tmp/guard_input.yaml), guard it, delete it after.
• No argument, no YAML: ask once — "What should I guard? Provide a path or paste a manifest."

Step 2 — Check if deployguard is available

Run:

python3 -c "import deployguard.guard" 2>&1

• If it succeeds: proceed to Step 3.
• If it fails: tell the user deployguard is not installed and show the install command. Then fall back to Step 3b (manual review).

Step 3a — Run the guard (package available)

Run the following Python snippet via Bash. Replace TARGET with the resolved path:

python3 - << 'PYEOF'
import sys
from deployguard.guard import guard, format_result
from rich.console import Console
import io

TARGET = "TARGET_PATH"

DEFAULT_RULES = {
    "require_resource_limits": "error",
    "require_probes": "error",
    "require_security_context": "error",
    "no_latest_tag": "error",
    "no_root_user": "error",
    "no_privileged_containers": "error",
    "iam_least_privilege": "warn",
}

result = guard(TARGET, DEFAULT_RULES, run_kubeconform=True, run_trivy=True)

buf = io.StringIO()
Console(file=buf, markup=True, highlight=False, width=100).print(
    format_result(result, explain=True)
)
print(buf.getvalue())
sys.exit(0 if result.passed else 1)
PYEOF

Read the output and present it per Step 4.

Step 3b — Manual review (fallback, no package)

Read every .yaml / .yml file under the target path. For each manifest, apply the rules in the table above using your own analysis. Produce the same output format as Step 4 — grouped by file, with why-it-matters and what-to-add for each violation. Make clear at the top that this is a manual review (no guard binary available).

Step 4 — Present the results

Group violations by file. For each file that has violations, show:

── k8s/deployment.yaml ────────────────────────────────────────
ERROR  [require_resource_limits]
  Container 'api' is missing resources.requests and resources.limits.
  Why it matters: Without resource limits one pod can starve the node and bring
  down its neighbours. Kubernetes will not schedule your pod if the node is
  overcommitted.
  What to add: Add resources.requests and resources.limits to every container spec.

WARN   [iam_least_privilege]
  ServiceAccount 'default' does not explicitly disable automountServiceAccountToken.
  Why it matters: Auto-mounted tokens give every pod API server access by default.
  If a pod is compromised the token is the attacker's first lateral-movement tool.
  What to add: Set automountServiceAccountToken: false on the ServiceAccount.

End with a summary line:

Guard result: 3 error(s), 1 warning(s) — FAILED.

or

Guard result: 0 error(s), 0 warning(s) — PASSED. All manifests are clean.

Rules for how you present findings:

• Never list a bare rule name with no explanation. Every violation gets its full why-it-matters and what-to-add text.
• Do not suggest fixes that haven't been explained. Explain first, prescribe second.
• If a file has no violations, do not list it.
• If the guard passed, say so clearly in one line — do not pad with generic advice.
• Act as a senior reviewer, not a linter. Tone: precise and helpful, not alarming.

---

Example invocations

/guard k8s/
/guard deployment.yaml
/guard .
/guard
<paste YAML here>

---

Example output (3 violations, 1 file)

── k8s/deployment.yaml ────────────────────────────────────────────────────────

ERROR  [no_latest_tag]
  Container 'payments-api' uses image 'payments-api:latest' — no pinned tag.
  Why it matters: :latest is mutable — the image you tested is not guaranteed to
  be the image that deploys. Pinned tags make rollbacks and audits reliable.
  What to add: Pin every image to a specific digest or semantic version tag.

ERROR  [require_resource_limits]
  Container 'payments-api' is missing resources.requests and resources.limits.
  Why it matters: Without resource limits one pod can starve the node and bring
  down its neighbours. Kubernetes will not schedule your pod if the node is
  overcommitted.
  What to add: Add resources.requests and resources.limits to every container spec.

WARN   [iam_least_privilege]
  ServiceAccount 'default' does not explicitly disable automountServiceAccountToken.
  Why it matters: Auto-mounted tokens give every pod API server access by default.
  If a pod is compromised the token is the attacker's first lateral-movement tool.
  What to add: Set automountServiceAccountToken: false on the ServiceAccount.

Guard result: 2 error(s), 1 warning(s) — FAILED.

---

Customising rule levels

Create .deployguard/config.yaml in your project to change rule levels:

guard:
  strict: true          # false = warnings only, never block
  explain: true
  rules:
    iam_least_privilege: error    # promote to error for your team
    no_privileged_containers: off # suppress if your workload requires it

The skill picks up this config automatically when deployguard is installed.