Почему ты вайбкодишь неправильно ? Как я построил свой 600+ часов workflow с ИИ

Summary

The presenter argues that the simple “prompt, wait, inspect” approach to AI coding often fails because requirements are underspecified, context gets lost, code is not reliably checked, and the work is handled by a single agent.

They recommend a five-part workflow:

1. Ask clarifying questions first. Have the agent resolve ambiguities before it starts work.
2. Create a plan and specification. Record the agreed requirements so the agent can stay on track as the conversation grows.
3. Write tests before code. Tests turn the specification into explicit checks. For example, authentication tests might require a token for a correct password, a 401 error for an incorrect password, and a temporary lockout after repeated failures. The tests should initially fail; the agent then writes code and keeps running them until they pass.
4. Use an orchestrator. A stronger model can divide the plan into subtasks, assign them to worker agents, and check their results against validation criteria. This can reduce the amount of work done by the more capable model.
5. Isolate parallel work. Git worktrees provide separate copies of the project, allowing multiple agents to work in parallel without changing the main working copy.

An agent saying “everything works” is not proof. Tests provide a verifiable result rather than a polite assurance. The workflow is presented as especially useful for building a project from scratch or implementing a complex feature.

The video is a practical guide and tutorial. It suggests instructing an agent to ask questions before acting, produce a plan and specification, write and run tests before coding, and then implement until the tests pass. The presenter also recommends VibeForge, described as a multi-agent development app.

Main speaker/source: The video’s presenter from the AshyGray channel.