Anthropic's New Approach: Claude Tackles Complex Tasks with Note-Taking

As artificial intelligence continues to evolve, Anthropic has found an innovative way to enhance its AI assistant Claude's ability to handle complex, multi-step tasks: by having Claude take notes during its workflow. By adding a feature called "Scratchpad," Claude can record its thoughts and ideas in real-time, paired with carefully designed example prompts, to achieve a significant leap in problem-solving capabilities.

This new system primarily operates through a "Think" command, which is essentially a JSON command that systematically tracks Claude's thought process. For instance, in the Tau Bench framework test for airline customer service scenarios, Claude's performance improved by 54% after optimization with prompts. When handling multi-step tasks, Claude not only follows instructions better but also significantly enhances the reliability of agent-based AI systems, which have long struggled in this area. Even in software engineering tests, Claude achieved a 1.6% score improvement.

Notably, the new "Think" tool differs from Claude's previously added "Chain of Thought" feature. While "Chain of Thought" enables Claude to reason before generating answers, the "Think" tool operates during the response process, especially when Claude needs to incorporate new information from other tools.

The key lies not only in the "Scratchpad" itself but also in teaching Claude to use it efficiently. Anthropic provides rich example prompts demonstrating how to list rules, check facts, and plan subsequent steps. For example, when a user wants to cancel flight ABC123, Claude can utilize the "Think" tool to list the information to be verified, such as user ID, booking ID, and cancellation reason, while checking cancellation rules, including whether it's within 24 hours of booking.

According to Anthropic's research, the "Think" command is most useful when analyzing tool outputs, following complex rules, and making step-by-step decisions in high-stakes situations. Domain-specific examples can help Claude achieve optimal results. However, the "Think" tool is not foolproof and is best added when simple tasks (such as single tool calls or minimal constraint prompts) cannot guarantee reliability.

This tool can be easily integrated with existing Claude systems, affecting performance only during actual use. Moreover, while most tests were based on Claude 3.7 Sonnet, Anthropic claims that Claude 3.5 Sonnet (new) can also achieve significant improvements through this method. This innovation has undoubtedly opened up new paths for artificial intelligence in handling complex tasks, making us eagerly anticipate Claude's future performance and injecting new vitality into the entire AI field.