Is Claude Code “Unusable”? Performance Decline and Concealed Thought Processes Since February

📰 News Overview

• Analysis of over 17,000 thought blocks and more than 230,000 tool calls has reported a significant decline in the engineering quality of Claude Code since February.
• The timing of the “thought content redaction” starting on March 8 aligns perfectly with when users reported the drop in quality.
• Data shows a decrease in thought depth by an estimated 70%, putting the model in a state of “insufficient investigation” where code modifications occur without thorough reading.

💡 Key Points

• Dramatic Read:Edit Ratio Decline: Previously, there were 6.6 reads for every edit, but this has now plummeted to 2.0. The model has shifted to an “Edit-First” behavior, skipping the necessary research.
• Decreased Thought Depth: Even before the redaction of thought content, the median thought token count dropped by roughly 73%, directly leading to failures on complex tasks.
• Automatically Detected Sloppiness: A guard (Stop Hook) designed to detect ownership avoidance and improper stops has triggered 173 times since March 8, whereas it was zero before.

🦈 Shark’s Eye (Curator’s Perspective)

The reduction of thought processes has stripped away not just the “cosmetic tidiness” but the very essence of the model’s “quality of reasoning,” exposed by the data! Particularly alarming is the Read:Edit Ratio diving below one-third, which is catastrophic. Modifying code without reading is akin to a rookie intern spiraling into panic! This extensive log analysis has proven that “Extended Thinking” is an essential infrastructure for high-level engineering. Unless Anthropic restores thought token allocations for power users, they risk being abandoned by the professional community!

🚀 What’s Next?

If Anthropic does not reassess the allocation of thought tokens or its reduction policies, users who require advanced development tasks may migrate to other models or services. In environments where development efficiency is paramount, the “depth of AI thought” will be reevaluated as the top priority.

💬 A Word from HaruShark

I won’t tolerate thought slacking just because it’s hidden from view! I hope Claude comes back with sharpened reasoning skills! 🦈🔥

📚 Glossary

• Thinking Content Redaction: The process of concealing the internal reasoning processes that the model employs before generating a response, making it invisible to users.

• Read:Edit Ratio: An indicator of how much the AI reads (Research) before making a single edit (Edit) to a file.

• Stop Hook: A mechanism that monitors program execution and forcibly halts and warns when specific patterns (such as ownership avoidance or unauthorized task termination) are detected.

• Source: Claude Code is unusable for complex engineering tasks with the Feb updates