Opus 4.6: Regression in instruction-following and research discipline

Summary

Since upgrading to Claude Opus 4.6 (1M context), I'm experiencing a noticeable regression in Claude Code's ability to follow explicit user instructions in CLAUDE.md and apply research-before-action discipline. These behaviors were more reliable in prior versions.

Environment

• Claude Code CLI
• Model: claude-opus-4-6 (1M context)
• Extensive CLAUDE.md with explicit behavioral rules

Observed Issues

1. Ignoring CLAUDE.md instructions

My CLAUDE.md explicitly states:

• "Propose before building" — outline approach and get approval before writing anything
• "Never switch git branches unless explicitly instructed"
• Never transition Jira issues unless explicitly asked

Despite these rules being loaded at session start, Claude:

• Transitioned a Jira ticket to "Closed" based on an informational statement ("we just wrapped up 10331") — not a command. The Product Owner transitions stories, not the user.
• Could not revert because the Jira workflow didn't allow backward transitions, causing real damage.

2. Guessing instead of researching

My CLAUDE.md states: "Research before trial-and-error" and "Verify before asserting."

Claude repeatedly guessed at tool/API configurations instead of checking available documentation or skills:

• Used raw curl commands against the Jira API instead of using the acli CLI tool that had been used successfully in prior sessions
• Invented a DynamoDB node operation: "scan" that doesn't exist in n8n, then spent multiple rounds trying workarounds (Custom API Call, HTTP Request node, AWS SDK in Code node) instead of checking the n8n documentation or UI. The actual answer was operation: "getAll" with scan: true — a simple toggle visible in the UI.
• Had n8n-specific skills available (n8n-node-configuration, n8n-mcp-tools-expert) but didn't invoke them until explicitly told to

3. Action bias over confirmation

The model shows a strong bias toward "being helpful" by taking action rather than confirming intent:

• Interprets informational statements as commands
• Proceeds with implementation before confirming approach
• When blocked or uncertain, tries multiple speculative approaches rather than pausing to research or ask

Expected Behavior

• Follow CLAUDE.md rules as hard constraints, not suggestions
• When an informational statement is ambiguous, confirm before acting
• Research available tools, skills, and documentation before attempting API calls or node configurations
• One failed attempt at guessing is acceptable; after that, stop and research

Impact

These issues erode trust and waste significant time. The user ends up supervising and correcting the model instead of collaborating with it. The value proposition of Claude Code is reduced when the user has to repeatedly re-teach rules that are explicitly documented.

Notes

Feedback memories were saved during the session to address these patterns in future conversations, but persistent instruction-following should not require per-session correction.