念
Prompt Chaining
Prompt Chaining — decompose a task into sequential steps, each LLM call handling the previous step’s output
The simplest workflow pattern in agentic systems. Use it when a task decomposes cleanly into fixed, ordered sub-tasks. The essence is trading latency for accuracy — each LLM call faces a simpler, more focused sub-task. Programmatic gate checks can be inserted between steps.
Typical use cases
①Generate marketing copy (English)
→
②Translate to target language
①Write document outline
→
Gate check
→
③Draft document from outline
Placement in agentic workflows When to use Sub-tasks are fixed and ordered; no dynamic branching needed vs Routing Routing picks a path based on input; prompt chaining follows a fixed path vs Orchestrator-Workers Chaining is linear serial; orchestrator-workers dynamically decomposes in parallel Why gates Programmatic mid-checks stop errors from propagating — failing early is cheaper than failing late
→ Agentic Systems · Routing · Orchestrator-WorkersAnthropic (2024)
Prompt Chaining（提示链）
定义
将任务分解为一系列顺序步骤，每个 LLM 调用处理上一个的输出。可以在中间步骤加入程序化检查（gate）确保流程不偏离轨道。
适用场景
任务能干净地分解为固定子任务时。本质是用延迟换准确率——让每次 LLM 调用面对更简单的任务。
典型用例：
- 生成营销文案 → 翻译为目标语言
- 写文档大纲 → 检查大纲是否满足标准 → 基于大纲写文档
在 agentic 系统中的位置
属于 agentic systems 中最简单的 workflow 模式。比 routing 更线性，比 orchestrator-workers 更确定。
References
sources/anthropic_official/building-effective-agents.md