Treat Your LLM Prompts as Interfaces, Not Notes

The article discusses the importance of treating LLM prompts as interfaces rather than informal notes. It introduces the Instruction Contract Specification (ICS) to define a structured approach to prompt design, improving maintainability and reducing token costs.

Details

The article argues that the problem with LLM systems is not the model itself, but the way instructions are written. Most LLM prompts are informal, with shared context assumed and maintained by the original author. This approach works for one-off experiments but fails in production systems where instructions are executed thousands of times by teams who weren't involved in the original decisions. The article introduces the Instruction Contract Specification (ICS), which applies the same discipline used for APIs, databases, and network protocols to the instruction layer. ICS defines five distinct layers - IMMUTABLE_CONTEXT, CAPABILITY_DECLARATION, SESSION_STATE, TASK_PAYLOAD, and OUTPUT_CONTRACT - each with their own rules and lifetimes. This structured approach helps address issues like context collapse, implicit constraints, and subjective evaluation. Mathematically, the ICS approach is proven to be more cost-effective than the naive approach, with up to 63% token reduction at scale. The project also includes an open-source toolchain for validating, linting, scaffolding, diffing, and reporting on ICS-based prompts.