提示词也是技术债务
本文指出，精心调整的提示词（prompts）在大模型快速迭代中会悄然失效，比代码更难维护。代码至少稳定，而每个模型升级都可能让精心编写的提示词不再适用。作者建议多数人直接使用第三方AI编码工具（如Claude Code、Codex等），尽量保持默认配置，避免使用MCP和skills文件。如需编写AGENTS.md，应只包含具体事实，避免行为引导型提示，并随时准备删除。
相关报道
Gary Marcus scrutinizes recent claims from OpenAI and Anthropic, urging readers to examine the underlying details and math behind their headline-grabbing announcements rather than taking them at face value.
The article argues that prompts used for AI tools like large language models are a form of technical debt, as they require ongoing maintenance, are poorly documented, and can break silently when underlying models change, leading to unpredictable costs and system fragility.
Adding assumptions to a formal property makes it logically weaker—guaranteed to work in fewer cases. Engineers add assumptions when stronger properties are impossible, too costly, or unverifiable. These assumptions often involve factors outside the program, like the operating environment or external dependencies.