Vol. III · Single-Agent Patterns · III.1
Prompt Chaining
A sequence of LLM calls where each step's output is the next step's input, with programmatic gates between them. The first workflow to reach for.
5 min read
Doctrine
- Chain when the task decomposes into named stages with checkable intermediates.
- Put code between hops: schema validation, regex, policy, retrieval.
- Trade latency for accuracy only when the eval says the extra hop pays for itself.
When chaining wins
Outline → critique outline → draft → translate. Each hop has a job the next hop should not redo. Anthropic's examples (marketing copy then translation; outline then document) still hold.
LangChain began as this pattern. It remains correct. The mistake was treating every problem as a chain of hidden prompts.
Anti-patterns
- A chain of five calls that could be one with a better prompt.
- Passing unconstrained prose between stages with no schema.