Поваренная книга вайбкодера: одна LLM в роли архитектора, ревьювера и тестировщика

A Russian developer published a detailed guide on multi-agent development with a single language model on Habr in August 2026 — no additional infrastructure purchases required, no months-long setup.

What is "LLM schizophrenia"

The core effect underpinning the entire method: the same model that just wrote a feature and was satisfied with it, when acting as a reviewer, finds three real bugs in it with failure scenarios. One set of weights — two diametrically opposed opinions about the same code.

Key facts of the method:

• The troupe consists of four roles: architect, developer, reviewer, tester
• Switching between roles is done via role prompts ("skills")
• The author runs a real production project on this principle, written by agents with almost no manual code
• The first step requires not a single line of code written manually — the troupe builds itself

The split is not a bug, but an architectural technique: a reviewer who does not know they are looking at their own code criticizes more honestly than any live colleague.

Why the process must live in a file

The architectural decisions made by the model in the role of architect disappear from context by the time the same model switches to the developer role. Without a recorded plan, the agent reinvents the wheel each time — spending tokens repeating work already done.

The author insists: any decision that must survive a single session must exist in a file, not in the dialogue and not in the model's memory. Skills — sets of role instructions for each of the four roles — are stored there as well and loaded with each role switch.

"The process must live in a file, not in one's head" — the key thesis of the guide published on

Habr.

Switching context between roles takes seconds; accumulated agreements and decisions are not lost between sessions.

How to launch the first multi-agent project

The troupe can build itself — this is the main practical conclusion of the article. It is enough to give the agent in the architect role a task description: it will form the project structure, distribute tasks between roles, and create plan files.

According to the author's description on Habr, the entire cycle — from an empty repository to working code with reviews and tests — proceeds without a single line written manually. The workflow looks like this:

• Skills — role instructions that define the agent's "personality" in each role
• Plan files — recorded decisions that survive session and context changes
• Role cycle — architect → developer → reviewer → tester, repeated when bugs are found
• Self-assembly — the project structure is built by the architect agent itself, not by a human

"Accessible, without an encyclopedia — over a single coffee" — this is how the author describes the format of their guide on

Habr.

What this means

Multi-agent development has ceased to be a laboratory experiment: one person with one language model builds an internal team that criticizes itself more honestly than most live reviewers. The main barrier to entry is not infrastructure, but the willingness to record the process in files.