self-harness is an open-source Chinese-language tutorial on Harness Engineering, published by the Datawhale community as a numbered series of chapters with an accompanying practice project. Its stated aim is to help developers understand how to build a robust underlying runtime architecture for complex, long-running AI agents.
The series organises the field as a progression rather than a set of alternatives: from Prompt engineering, through Context engineering as dynamic information management, to harness engineering at the system level. That framing is the tutorial’s own organising claim, and it supplies the sequence the chapters follow.
It is written for AI application developers, people interested in large-model technology, and Python developers with basic programming ability. The outcomes it states for a reader are understanding how context engineering differs from prompt engineering, handling dynamic context management, designing an extensible AI skill system, and building a minimal system of the kind Claude Code represents.
Scope & Structure
The tutorial is split into a theory part and a practice part. The theory chapters introduce prompt engineering’s concepts, methods and limitations; context engineering’s concepts and methods; harness design for agents that must run a long time without going wrong; and the progression connecting the three. The practice part is built around miniMaster, a minimal harness implementation whose code ships in the same repository, and which the tutorial uses to show how harness theory lands in an actual agent system.
The series is read online at https://datawhalechina.github.io/self-harness/, and the chapters live as Markdown in the repository alongside the practice code.
Status
The project is marked Alpha. Its own notice states that the content is still being improved and may contain errors or omissions, and invites readers to file issues. The repository’s outstanding-work list names one item: an analysis of agent products built on mainstream harness systems, giving NanoBot and OpenHarness as its examples.
The work is licensed under Creative Commons Attribution–NonCommercial–ShareAlike 4.0 International.