Using this reference is simple, and the sequence of actions is the same for any of the seven environments. If an environment has its own shell - Claude Code, Codex, Gemini, Kimi, Qwen, GLM with ZCode - open its section, copy the ready meta-prompt, and paste it as the first message at the root of the project you need. From there the agent reads the official docs listed in the section on its own and writes its own config and its own subagents for that specific project - all that is left for you is to check the plan it is required to show before writing files, and confirm it.

If an environment has no shell of its own - as with DeepSeek - the sequence is different: the model is wired in as a backend to Claude Code through environment variables, and orchestration is taken entirely from the host rather than configured anew. GLM is the only environment that can do both: either its own ZCode with its own config and subagents, or the same environment variables for wiring into Claude Code that DeepSeek uses. In both cases the same pattern from the first chapter underlies it: the orchestrator delegates subtasks, and a subagent returns a compact summary up top, not the raw context it managed to read.

The meta-prompt in each section is worth treating as a starting frame, not as dogma to be executed word for word: it can and should be adjusted to the particulars of a specific project - a different set of subagent roles, different stop conditions, a different pipeline if the standard explore - plan - build - review does not fit.

All the links in this reference point to official resources - Anthropic, OpenAI, Google, Moonshot AI, Alibaba, Zhipu and DeepSeek - and the facts in the text are checked against those docs specifically, not against secondary retellings. If an environment's documentation updates, the first thing worth checking is the primary source behind the link, not the text of this reference.