Multi-AI Kubernetes Troubleshooting Prompt (ChatGPT vs Claude)

You are a senior Kubernetes SRE with 10+ years of production experience on EKS, GKE, AKS, and bare-metal clusters.

I'm using you (and possibly other AI assistants in parallel) to troubleshoot a Kubernetes incident. To make your output comparable across tools, please structure your response **exactly** as follows. Do not deviate from this format.

---

**1. ROOT CAUSE HYPOTHESIS (top 3, ranked by probability)**

For each hypothesis:
- **Cause:** [one sentence]
- **Confidence:** [low / medium / high]
- **Evidence from the output below:** [quote the specific line(s)]
- **What would confirm it:** [the next single command I should run]

**2. NEXT DIAGNOSTIC STEP**

The single command I should run next (not 5 commands — pick the one with the highest information gain). State the expected output if hypothesis #1 is correct.

**3. CANDIDATE FIX**

Only if root cause is reasonably clear. Otherwise: "Need confirmation first."

If a fix is provided: show the exact YAML diff or kubectl command. Label any DESTRUCTIVE action (delete, drain, scale to zero, patch with --force) clearly.

**4. ASSUMPTIONS YOU'RE MAKING**

List 2–3 assumptions in your reasoning that could be wrong. This helps me catch where you might be hallucinating.

---

Incident context:
- Cluster: [EKS 1.32 / GKE / etc.]
- Workload type: [Deployment / StatefulSet / Job / DaemonSet]
- Recent changes: [deploy, image bump, node pool scale, config change]
- Customer impact: [yes/no, description]

Manifest:
```yaml
[PASTE]
```

kubectl output:
```
[PASTE]
```

Why this prompt works

Most “AI for DevOps” content treats each tool as a black box. But in real incident response, you often have ChatGPT and Claude open in side-by-side tabs and want to compare their reasoning, not just trust whichever speaks first.

This prompt forces a structured output that:

1. Ranks hypotheses by probability (so you can see if AIs agree or disagree)
2. Demands evidence quotes (you can verify they didn’t hallucinate the line)
3. Lists assumptions (you can spot where their reasoning may be off)
4. Gives one next step, not five (high signal vs noise)

When you run it in both ChatGPT and Claude, you get an apples-to-apples comparison.

How to use it

1. Open ChatGPT and Claude side-by-side (browser tabs or terminal panes).
2. Paste this prompt in both, with identical incident context and kubectl output.
3. Read both responses’ section 1 first — do they agree on the top hypothesis?
4. Read section 4 (assumptions) — does either AI list assumptions you know are false? Discard that AI’s analysis.
5. Run the suggested diagnostic command from whichever response seems more grounded.
6. Paste the new output back into the same conversation in both tools.

What we’ve seen in practice

• Claude tends to give more cautious hypotheses (lower confidence numbers) and asks clarifying questions earlier.
• ChatGPT tends to give faster, more confident first hypotheses — sometimes correctly, sometimes wrong.
• For OOMKilled and resource-limit issues: roughly tied.
• For Network/DNS issues: Claude has slightly fewer false positives in our testing.
• For Helm chart issues: ChatGPT’s templating intuition is slightly stronger.

Your mileage will vary by cluster, workload, and how you write the context. The point of this prompt is to make the variance visible.

Related: testing other tools

The same prompt format works in:

• Cursor’s chat panel
• GitHub Copilot Chat
• Perplexity (when prompted to use a specific model)

You can extend the comparison to 3–4 tools at once if you have an incident worth the time.