This guide, written in Chinese by the AI-programming blogger Xiangyu (翔宇), is an overview page that links ten of the author’s in-depth tutorials, most of them about Claude Code. Its premise is in its title: rather than teaching how to operate a tool, it teaches methods for making an agent do its work by your rules. Tools turn over every half year, the author writes — last year Cursor, this year mainly Claude Code and Codex — but the methods for controlling an agent carry over, so someone who has learned the methodology only has to adapt to a new tool’s interface.
The author argues that most people stop at knowing how to use a tool, and on complex projects find that agents drift off course, forget things, overstep their authority or make unexpected judgments at key steps. In his view the problem is not the tool but the absence of rules set for the agent. The guide answers one question — how to make an agent work reliably in your project — and ends with a self-assessment checklist and a table recommending which tutorials to read for which pain point.
Key Points
- The author distinguishes three concepts he says are often confused. Prompt engineering concerns how to write instructions for a single exchange. Context engineering, which he attributes to Shopify CEO Tobi Lutke, concerns feeding the agent the right information at the right time. Harness Engineering he describes as the overarching framework he has summarised from his own practice: managing the agent like a new employee through rule files, context strategy, permission boundaries and extension mechanisms. In his framing the three are nested rather than alternatives, with the first two as sub-dimensions of harness engineering.
- He gives harness engineering a four-layer architecture that works together: a rules layer (project memory such as CLAUDE.md), an information layer (context engineering), an instruction layer (prompt engineering) and a boundary layer (runtime control through permissions and sandboxes). Its core challenge, he argues, is constraining an executor that makes its own judgments, so what one writes are criteria for judgment rather than execution steps.
- He sums up context engineering as feeding the right information, at the right time, in the right format, and sorts injection methods by how long the information stays valid: permanent information in a project memory file, session-level information in the system prompt or first message, on-demand information through tool calls and retrieval, and real-time information from execution results. He warns against equating context engineering with RAG, which he treats as only one of its techniques.
- He presents project memory files —
CLAUDE.mdfor Claude Code, AGENTS.md for OpenAI Codex,.cursorrulesfor Cursor — as the persistent carrier of rules for an agent that is otherwise stateless, highlights CLAUDE.md’s global, project and directory levels, and advises writing behavioural constraints and decision preferences rather than encyclopaedic knowledge or operating steps, iterating on the file whenever the agent makes an unwanted judgment, and putting hard red lines near the top because agents attend more to the beginning and end. - For prompts in agent programming he recommends explicit constraint boundaries, decomposition into independently verifiable steps and predefined criteria for done, and suggests that showing an example file works better than describing a desired style.
- He argues that specifying a thinking framework (first principles, inversion, MECE, 5-Why and the like), per session or fixed in the project memory file, matters less for making the agent smarter than for making its analysis predictable and reproducible.
- Runtime control, in his account, has three dimensions: managing the context window (including automatic compaction), choosing a thinking mode by the length of the task’s reasoning chain, and configuring permissions on the principle of least privilege, loosened gradually as trust grows but kept conservative in production. Extension mechanisms add Agent Hooks as runtime checkpoints for automated safety and quality checks, and plugins as a standard way to give the agent new capabilities.
- He predicts that agent programming methodology will become a basic skill for every technical person, like version control and automated testing.
Context
The guide is a practitioner’s synthesis, which also points readers to the author’s course, rather than a report of measurements; much of its detail concerns Claude Code’s features as the author describes them, while its claim is that the methods themselves are tool-independent. Its own use of “harness engineering” as an umbrella over context and prompt engineering is the author’s framing.