Andrey Borodulin
Moscow, Russland
- Abgeschlossene Buchungen
- 6
Andrey Borodulin, in Moskau geboren, hat einen Abschluss in art history and cultural studies, begann als Kultur- und Musikjournalist, wurde jedoch durch die Revolution 2014 und den Krieg in der Ukraine in die Nachrichten…
Gesehen in Хабр, Barron's, Franceinfo +18 weitere
The Fifth Button in One Evening: Frontend Architecture for LLM Context
Berichtsstil, Informativ: Primarily a practical first-person technical account, with recommendations and personal judgments about tools and architecture.
Drawing on a frontend project built largely with LLM agents, Andrey Borodulin argues that architecture should minimize the context needed for each code change. He describes how an initially fast workflow produced duplicated buttons, oversized components, tangled effects, and inconsistent copies of server data. His proposed structure uses a small set of Feature-Sliced Design layers—app, features, and shared—with isolated feature slices and public APIs. He recommends generating typed API hooks from the backend’s OpenAPI contract, keeping server data in TanStack Query and UI state in local state or stores, and enforcing dependency rules with dependency-cruiser, strict TypeScript, knip, and custom CI checks. Tailwind is presented as helpful because styles stay beside markup and use a limited design-token vocabulary, while shared UI components and review help prevent duplicate controls. Since static checks cannot establish whether a page looks right, he advocates Playwright tests, screenshots, and agent-driven browser checks, while noting that visual review remains a weak point requiring human oversight.
LLMs Write Terrible Code No Worse Than Humans. Just Faster
Berichtsstil, Interpretierend: First-person technical analysis with strong recommendations and evaluative claims, supported by practical examples.
After building a pet project largely with LLM agents, the author found that rapid initial development gave way to the familiar slowdown caused by tangled architecture, duplicated logic, and regressions. He argues that architecture matters especially for AI-assisted coding because low coupling and bounded contexts reduce the code an agent must load into its limited context. He recommends applying SOLID and related principles, using DDD selectively for backend business logic, and enforcing layer and context boundaries with two separate Deptrac configurations. Written project guidance should state intent, while automated checks such as Deptrac, PHPStan, and tests enforce it. For larger changes, the author advocates a development pipeline that separates planning, implementation, and review, with an independent reviewer agent using a clean context. The central conclusion is that LLMs amplify engineering practices: machine-enforced rules and sound architecture help them produce maintainable changes, while poor foundations accelerate technical debt.
