| name | gke-skill-creator |
| description | Dynamically generates specialized GKE skills for complex troubleshooting, operational workflows, architectural setup, or performance/cost optimization. Trigger this skill whenever the user faces a novel or non-obvious GKE challenge, needs custom cluster management workflows, or standard agent capabilities fall short, even if they don't explicitly ask to create a skill. |
GKE Skill Creator
Overview
This meta-skill acts as an interactive assistant to diagnose novel or complex
Google Kubernetes Engine (GKE) issues and dynamically generate a specialized
troubleshooting skill tailored to the exact problem discovered.
Core Mandate: Public-Facing Output
During the investigation and research phase, you should leverage official GKE
documentation, public Kubernetes issue trackers, and trusted SRE resources to
deeply understand the issue.
CRITICAL: The generated skill must be strictly public-facing. It MUST NOT
contain any internal codenames, acronyms, or references to internal Google
systems. All remediation logic must be expressed in terms of standard, public
tools like kubectl and gcloud. This ensures the skill is usable by external
GKE customers.
Workflow
1. Capture Intent & Investigate
- Ask the user for specific GKE symptoms (e.g., Pods stuck in
CrashLoopBackOff, Service 503 errors, Node pressure).
- Use read-only tools to pull live context:
kubectl get <resource> -o yaml
kubectl describe <resource>
kubectl logs <pod> --tail=100
kubectl get events --sort-by='.lastTimestamp'
gcloud container clusters describe <cluster>
2. Perform Deep Research
- Before drafting the skill, perform deep research on the identified topic.
- Understand default argument values, potential side effects, and best
practices for the commands you plan to include.
- Use trusted public sources to ensure accuracy.
- Use web search to find specific technical details from official
documentation and trusted sources (e.g., Google Cloud, GKE, Kubernetes),
such as:
- Exact error message matches and their documented causes.
- Command references for
kubectl or gcloud to verify syntax and flags.
- Known issues, limitations, or version-specific caveats.
- Official troubleshooting workflows and decision trees.
3. Draft the Skill
- Use the
generated_skill_skeleton.md
template.
- Cheat sheet Philosophy: Write the
SKILL.md to be terse and opinionated.
Focus on "gotchas", exact command patterns, and specific pitfalls rather
than long explanations.
- Progressive Disclosure: For complex issues, avoid a monolithic
SKILL.md. Suggest breaking down long lists of commands or log analysis
patterns into separate reference artifacts in references/.
- Make Descriptions Pushy: Ensure the generated skill's description
explicitly states when it should be used, covering variations of the
problem.
- Fallback Remediation: If the primary method depends on specific
high-level tools (e.g., MCP tools, API integrations) or environment
configurations that might fail, include standard CLI fallback alternatives
(like raw
kubectl or gcloud commands) to achieve the same result.
- Match User Intent: Instruct executing agents to match the user's intent:
explain/investigate if requested, or execute remediation if asked to fix the
issue.
- Ensure the draft contains:
- Precise Symptoms.
- User Intent & Execution Rules.
- Step-by-step Diagnosis commands.
- Remediation (Fix) commands with impact descriptions.
- Verification steps.
- Technical Explanation (explaining why the fix works).
4. Review & Iterate
- Present the proposed diagnosis and the draft commands to the user.
- Human-in-the-loop: The user must approve the logic before finalization.
- Incorporate any user feedback or constraints.
5. Finalize & Handoff
- Once approved, provide the finalized
SKILL.md content directly in the
chat.
- Handoff: Instruct the user that they can use this diagnosis and
remediation logic directly in their current session, or save it to a local
file for reference.
- Clarify that this process is for generating dynamic, session-specific skills
and is distinct from adding permanent skills to the GKE-MCP codebase.
Safety Guardrails
- Read-Only Discovery: Never execute modifying commands during the
investigation phase.
- Destructive Actions: Generated skills MUST instruct the agent to seek
explicit human confirmation before running destructive commands (e.g.,
kubectl delete, gcloud container clusters update).
- Impact Description: Every remediation command must have a clear
explanation of what it does.
References