When we build a complex web application, the AI become error prone, we may introduce some simple Harness engineering scheme, like add project.md, agent.md, memory.md, progress.md, rules.md, etc, and create some meaningful prompt to let the AI Agent to follow a certain rules to let LLM understand the project better, like the current project status, errors, bugs has been made, what progress has been done, etc. And shall have a good logic to compress context to avoid context overflow.