compose

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Plan

规划

Use a subagent for exploratory reading, keeping the main thread's context for the composition itself.
Choose the skills that matter to the ask, scaled to it — one skill per outcome. Define the stages the ask needs, within this session or across sessions; a single skill can fire more than once, at different stages. For each chosen skill, conclude its useful-when: the condition in this conversation that will call for it.
使用 subagent 进行探索性阅读,同时为主线程保留用于最终编排的上下文。
根据请求选择真正需要的 skills,并按请求的规模进行调整——每项成果对应一个 skill。定义请求所需的各个阶段,这些阶段可以位于当前 session 内,也可以跨 session;单个 skill 可以在不同阶段多次触发。对于每个选定的 skill,明确其 useful-when:即当前对话中会触发它的条件。

Invoke

调用

Work the ask, invoking each skill as the conversation reaches its condition, in dependency order:
  1. Invoke it via the Skill tool. As its body loads, conclude its wrong-if — the conditions under which it refuses.
  2. After the output, silently check the wrong-if against what came out; if it trips, correct the output in the same turn.
A wrong-if trip, or a turn where no chosen skill's condition fits, amends the one held belief that failed — re-derive the chosen set only if needed. The loop ends when every stage the plan defined has had its skill run, or the user stops the composing.
着手处理请求,按照依赖顺序,在对话达到相应条件时逐一调用每个 skill:
  1. 通过 Skill tool 调用它。在其主体内容加载时,明确其 wrong-if,即它会拒绝执行的条件。
  2. 输出后,将 wrong-if 与实际结果进行静默核对;如果触发拒绝条件,则在同一轮中修正输出。
如果 wrong-if 被触发,或某一轮没有任何选定 skill 的条件匹配,则修正其中一项失败的既定判断——仅在必要时重新推导出所选 skill 集合。当规划定义的每个阶段都已运行对应 skill,或用户停止编排时,循环结束。

Result

结果

The worked ask is the result, and nothing about the skills that built it. Where the user stopped the composing, name the work still undone, in the ask's terms.
The ask's words admit two readings that would build different things → use clarify skill, then compose over what the pick settles.
已完成处理的请求就是最终结果,不应提及构建该结果的 skills。如果用户中途停止了编排,应使用请求本身的表述说明尚未完成的工作。
如果请求中的措辞存在两种会产生不同结果的解读 → 使用 clarify skill,然后根据所选定的解读继续编排。