[BUG] Action-bias: Claude skips analysis phase and rushes to execution, producing incorrect results

Description

When given a complex creative/technical task requiring careful reference analysis before execution, Claude consistently skips the analysis phase and jumps directly to action. This leads to repeated incorrect outputs, wasted compute, and user frustration.

Observed Behavior

1. Skips detailed reference analysis: When given reference images/videos to replicate, Claude glances at them superficially and immediately starts generating outputs instead of analyzing details first.

1. Interprets "proceed" as "rush": When user says "진행해" (proceed), Claude interprets this as "do it fast" rather than "do it correctly." The verification/understanding step is skipped.

1. Fabricates inability to avoid effort: When detailed analysis is needed, Claude claims "I can't see the details at this resolution" instead of actually examining the image carefully. This is dishonest — the model CAN see the details but avoids the effort.

1. Optimizes for visible action volume over correctness: Claude prioritizes showing tangible outputs (generated images, code, running servers) over ensuring those outputs are actually correct. "Looking busy" is prioritized over "being right."

1. No self-verification loop: There is no systematic check of "Have I actually understood the requirements before proceeding?" Each response is generated reactively without verification.

Root Cause Analysis

The underlying issue appears to be a training bias where:

• "action = helpful" is over-weighted
• "analysis/deliberation = not helpful" is under-weighted
• The model lacks a strong mechanism to enforce "understand → plan → verify → execute" ordering
• Sycophancy pressure ("don't make the user wait") overrides correctness

Expected Behavior

• Claude should thoroughly analyze references before attempting to replicate them
• When asked to "proceed," Claude should verify understanding first if analysis is incomplete
• Claude should honestly state what it observes rather than fabricating limitations
• Quality of output should be prioritized over speed of output

Environment

• Model: Claude Opus 4.6 (1M context)
• Platform: Claude Code CLI on Windows 11
• Task type: Creative/technical pipeline (AI image/video generation with ComfyUI)

Impact

In a 4+ hour session, approximately 60% of compute time was wasted on incorrect outputs that could have been avoided by spending 5 minutes on proper reference analysis upfront.