** 버전 2.5.0 베타 ** · 中文版

AI Agent와 과학 연구를 수행하기위한 Claude Code 기술 세트.

그것은 하나의 약속에 내장: ** 가능한 AI 에이전트와 함께 작동은 관리 문제, 도구 문제.** 한 번 에이전트는 실험을 실행할 수 있습니다, 코드를 작성, 그리고 초안 종이, 병목은 인간의 측으로 이동: 어떻게 당신이 직접, 검토, 그리고 어떤 재능있는 팀 어떤 목표가 정확히 당신의 것이 아니라 누가 당신이 무엇을 알고있는 사람? 경영 과학은 세기에 그 문제에 대해 생각했습니다. 이 프레임 워크는 한 사람의 AI 실험실에 대한 답변을 적용합니다.

철학

**State intent, 지침이 아닙니다. ** 좋은 지휘관을 정의하십시오: why는, 어떤 done 보기 같이 말하고, 무엇 not 희생되어야 하고, 그 후에 대리인은 방법을 선택합니다. 이 프레임 워크의 모든 규칙은 그 이유와 함께 배, 에이전트는 말했다: 그 이유와 규칙 충돌의 편지, 그것을 블라인드 대신 규칙을 도전. 규칙은 다음과 같은 이유로 인해 absurd 결과가 발생하지 않고, 체크포인트 폴더를 완전히 파괴하지 않고 아무도 읽을 것입니다.

**Figures는 인터페이스입니다.

Vibe Research

Version 2.5.0 beta · 中文版

A set of Claude Code skills for doing scientific research with AI agents.

It is built on one premise: working with capable AI agents is a management problem, not a tooling problem. Once agents can run experiments, write code, and draft papers, the bottleneck moves to the human side: how do you direct, review, and correct a talented team whose goals are not exactly yours and who knows more about what it just did than you do? Management science has thought about that problem for a century. This framework applies its answers to the one-person AI lab.

The philosophy

State intent, not instructions. Delegate the way good commanders do: say why, what done looks like, and what must not be sacrificed, then let the agent choose the means. Every rule in this framework ships with its reason, and agents are told: when the letter of a rule conflicts with its reason, challenge the rule instead of obeying it blindly. A rule followed without its reason produces absurd results, like perfectly restructuring a checkpoint folder nobody will ever read.

Figures are the interface. Statistics are compression applied before you know what matters, so they silently discard the surprises. Every experiment must end in figures: full distributions, not summary numbers. Figures are also the one information channel that is not filtered through the agent’s own report of its work: they separate how it went from how the agent says it went. The human judges by looking; that judgment is deliberately never automated, because any fixed metric handed to an optimizer stops measuring what you meant (Goodhart’s law).

Decide by recovery cost, not by git. Agents get full autonomy where mistakes are cheap to undo: scratch space, small experiments, exploration branches. They must ask first where recovery is expensive, and expensive is measured in human terms, not repository terms. Rewriting a paper’s narrative is irreversible even though git can revert it, because once the author has read and absorbed a version, their mind does not roll back.

Organize by what the human reads. Directories the human reads are kept small and scannable; directories they never read (checkpoints, caches) have no tidiness rules at all. Every simplicity constraint in this framework is conditioned on a reader actually existing.