prompt-chaining

Define the end output first, then split the task into small steps where each has one responsibility. Lock an input/output schema for every handoff, validate after each step, and apply retry or stop rules when output fails validation.

Use prompt chaining when the task has a clear sequence, later steps need cleaned output from earlier steps, or you need step-specific prompts, tools, or models. If one prompt solves the task well, chaining adds unnecessary handoff cost.

Good boundaries include extracting facts, summarizing facts, drafting an answer, checking the answer against the facts, and formatting for the target system. Each step should have exactly one job and a well-defined handoff format.

Common failure modes include too many steps making the flow slow and fragile, vague handoff formats causing drift, later steps redoing earlier work, and missing validation letting errors compound. Copying full context into every step also spikes cost.

Avoid it when order does not matter, work can run in parallel, steps depend on dynamic branching rather than fixed flow, or the handoff cost exceeds the quality gain. In those cases use a single prompt or a different orchestration pattern.