Amid the tide of the digital era, Dewu is committed to driving the transformation of data warehouse development, especially achieving significant progress in the application of AI Coding tools. The Dewu team has greatly improved efficiency in repetitive tasks through Claude Code, a core tool. However, during practical application, the team has also identified several pain points that need to be addressed.
Firstly, the lack of "memory" in AI during the development process is a major issue. Claude Code tends to forget context information during long conversations, such as important field units, which may lead to serious errors in generated SQL, with data result differences up to 1,000 times. This happens because when the conversation content approaches the limit, the AI automatically compresses historical information, resulting in loss of context.
Secondly, the stability of compliance enforcement is unsatisfactory. During project pressures, the rate of manual compliance drops to 60% to 70%, while the AI's "memory" execution rate is only 70% to 80%. This indicates that relying on AI's memory for compliance is not reliable; what is truly needed is to embed these standards into the system for mandatory checks.
Finally, when handling large-scale requirements, the performance of AI becomes increasingly unstable. Complex development tasks often rapidly expand the AI's context, making it more prone to "forgetting," thus leading to frequent errors. To address these issues, the Dewu team proposed the concept of the "Harness" engineering, which involves writing code "guardrails" to ensure each execution follows the standards, thereby reducing human errors.
The core of the "Harness" engineering lies in transforming execution standards into system-level checking mechanisms, enhancing AI capabilities through hooks and automation tools. The Dewu team stated that the future goal is to eliminate uncertainties in the development process through these mechanisms, allowing AI to be more reliable in complex development environments.
Key Points:
🌟 **AI's "Memory Loss" Problem**: In long conversations, AI tends to forget important context information, leading to frequent errors in generated SQL.
🛡️ **Unstable Compliance Enforcement**: The manual compliance rate is low, and the AI's execution rate is also not ideal; a systematic solution is urgently needed.
🔧 **The Proposal of Harness Engineering**: By implementing automated checking mechanisms, transform execution standards into system-level mandatory checks to improve development stability.
