AI agents need more than rules: how to transfer engineering expertise
The author argues that declarative instructions like long prompts, generated documentation, or collections of skills are insufficient for AI coding agents to produce code that passes review. Instead, they propose a six-rule process where agents follow the same decision path as a developer, decompose tasks to reference solutions, use few-shot examples, and gradually turn verified tests and accepted decisions into reusable skills. The approach emphasizes iterative learning and human oversight rather than autonomous agent operation.
The article on Habr describes the evolution of AI coding assistance from detailed prompts to documentation and skills, but the author found that these alone didn't yield code ready for review. The problem is that competence cannot be fully transferred declaratively; much of it emerges during work. The author proposes six rules: agent must follow the same decision path as a developer, decompose tasks to the level of a reference solution, provide local project knowledge, use the verified reference as a few-shot for subsequent code, make verified tests a standalone acceptance stage, and turn accepted decisions into reusable skills. He distinguishes this from vibe coding and autonomous modes like Ralph loop, emphasizing human as final arbiter. He gives examples such as fixing N+1 issues through skills and using first implementations as templates for subsequent code.
- Abbreviations
- LLM = Large Language Model — большая языковая модель
- CRUD = Create, Read, Update, Delete — создание, чтение, обновление, удаление
- DDD = Domain-Driven Design — предметно-ориентированное проектирование
- TDD = Test-Driven Development — разработка через тестирование
- ORM = Object-Relational Mapping — объектно-реляционное отображение
Source: Habr — хаб ИИ —
original