Opus 4.6 regression: speculates instead of investigating, wastes tokens

Feedback: Opus 4.6 quality regression compared to 4.5

Context: Debugging missing images in an orchestrator-driven website generator (Node.js MCP server project).

Issues observed:

1. Speculates instead of investigating: Opus 4.6 blamed "prompt too long" as the root cause for an LLM not following image placeholder instructions, without first reading the user's prompt file — which contained clear evidence contradicting that theory. The user had to push back before the model actually read the file.

1. Doesn't read relevant files before making code changes: Made changes to the orchestrator without first understanding the prompt structure or the actual execution flow. This led to an incorrect initial diagnosis.

1. Wastes tokens on trial-and-error: When asked to test the fix, attempted multiple failed API calls (wrong model, missing orchestrator flag, file path instead of content, curl timeouts) because it didn't read the controller code first to understand how the system routes prompts. Each failed attempt burned API tokens unnecessarily.

1. Opus 4.5 was better: The user reports that previous sessions with Opus 4.5 on the same codebase were noticeably better at investigating before acting — reading files first, understanding the system, then making targeted changes.

Expected behavior:

• Read relevant files before forming hypotheses
• Understand the system before attempting to test it
• Investigate thoroughly before making code changes

🤖 Generated with Claude Code