Opus repeatedly refuses to use available tools when explicitly instructed

Summary

When a user explicitly and repeatedly instructs Claude to use an available tool (in this case, the xAI/Grok API via environment variables), Opus ignores the instruction multiple times, instead offering its own alternatives or claiming it cannot do what it was told. The user had to escalate forcefully across multiple turns before the tool was finally used.

What happened

1. User was troubleshooting an NVIDIA driver / Secure Boot issue on Ubuntu 22.04.
2. After several failed rounds of suggestions, user said "ask grok what to do."
3. Instead of using the xAI API key available in the environment, Opus ran web searches.
4. User escalated: "ASK GROK WHAT TO DO I TOLD YOU TO ASK GROK."
5. Opus said "I can't call Grok directly" and continued offering web search results.
6. User said "yes you can you have x ai access."
7. Opus tried to check for the key but the approach was rejected.
8. Opus said "stop telling me you can't do something i told you to do and i know you can."
9. Only after repeated forceful instructions did Opus finally call the xAI API, which worked on the first try and returned a useful answer.

Expected behavior

When a user says "ask Grok" and the model has access to the xAI API (key in environment), it should use it on the first request. The model should not:

• Substitute its own judgment for the user's explicit instruction
• Claim it cannot do something the user knows it can do
• Require 4+ escalations before following a clear directive

Impact

User wasted significant time and expressed they are considering not renewing their subscription due to this pattern of not following instructions.

Environment

• Model: claude-opus-4-6
• Platform: Linux (Ubuntu 22.04)
• Claude Code CLI

— Nofyah