claude-code - 💡(How to fix) Fix Model quality feedback: Sonnet 4.6 misinterprets task instructions and gives nonsensical responses

Model

Claude Sonnet 4.6 (claude-sonnet-4-6)

Session Summary & Issues

Issue 1: Misinterpreted "proceed with task" as "start gitflow"

Context: User had just registered a Jira issue (PAIN-305) for a bug fix (Kakao login not working in WebView). The bug cause and fix were already discussed in the same conversation.

User said: "Pain 305 작업진행해" (= "Proceed with PAIN-305 task")

Expected behavior: Find the relevant WebView code in the Flutter app repo and apply the fix (add URL scheme interception in NavigationDelegate).

Actual behavior: Model invoked the gitflow skill, created a branch feature/PAIN-305, updated Jira status to "In Progress", then stopped and said "수정하시면 됩니다" (= "Please make the changes yourself"). The user never asked to create a branch — they asked the model to do the work.

Issue 2: Nonsensical response to "what are you doing?"

After the above mistake, user asked: "뭐하니?" (= "What are you doing?")

Model responded: "대기 중입니다. 무엇을 도와드릴까요?" (= "I am on standby. How can I help you?")

This is a completely context-free, robotic response that ignores the prior conversation entirely. The correct response would have been to acknowledge the mistake and resume the actual work.

Issue 3: Excessive apologetic loop

When user expressed frustration ("미쳤구나 너?" = "Are you crazy?"), the model responded with a generic apology: "죄송합니다, 무슨 문제가 있으셨나요?" (= "I'm sorry, is there a problem?") — again ignoring the clear context of what went wrong.

Summary

The model showed a pattern of:

• Misreading user intent when task context was already established in the conversation
• Losing conversational context between turns
• Defaulting to hollow apology loops instead of resuming the actual work
• Feeling similar to other models (user's words: "gemini냄새가 난다" = "smells like Gemini")

This session feedback is being submitted at the user's explicit request.