Skip to content
Eldar Shakhvaliev
Discuss a project

How I Work With AI

Working with AI agents, to me, isn’t code generation on demand — it’s an engineering discipline: agents act under an explicit specification, results get checked, and architectural decisions and responsibility stay with the engineer. A short note on principles, without project-specific detail.

A specification, not a chat

A non-trivial task doesn’t start with a chat message — it starts with a written document: what to build, what the constraints are, how to verify the result. The specification is a shared language with the agent, and a brief you can return to a month later, once the chat context is long gone.

Verification is built into the process

Between “the plan is ready” and “we started building” sits a separate step: auditing the plan itself, not just the output. For high-uncertainty work — research or a code review — several agents work in parallel and independently, then adversarially check each other’s conclusions; anything not confirmed by all of them gets flagged separately. Tests capture the invariants before implementation, not written after the fact as a checkbox.

What stays with me

The agent handles volume — generation, exploring options, routine diagnostics. Architectural decisions, fact-checking, security, and final responsibility for the result stay with me. If a task was built on a wrong assumption, that’s what I work from when I look into it — I don’t bend the conclusion to fit a nicer-looking answer. Work carries over between sessions: context and decisions are written down, not held in a conversation’s memory.