Prompt Chaining Cheat Sheet
Splitting one impossible prompt into several reliable ones
When a single prompt tries to do four things it does all four adequately. Chaining trades latency and cost for accuracy you can measure per step.
Topologies
| Shape | Structure | Use when |
|---|---|---|
| Sequential | A → B → C | Each step genuinely needs the last one's output |
| Parallel | A and B → merge | Independent sub-questions; one round trip of latency |
| Conditional | Classify, then branch | Inputs of clearly different kinds |
| Map-reduce | Per-chunk, then combine | Input larger than the window |
| Iterative | Draft → critique → revise | Quality matters more than latency; cap the loop |
Where to split
- Where the output format changes Extraction and prose want different temperatures
- Where you want to validate You can only check a step whose output has a shape
- Where a cheaper model would do Classification rarely needs your most expensive model
- Where a human might need to approve A checkpoint must sit on a boundary
- NOT where it merely feels tidier Every split adds a call, a failure mode and a place to lose context
Errors compound
Five steps at 95% each is 77% end to end, and the last step cannot tell that its input was already wrong. Validate between steps rather than only at the end, and pass the ORIGINAL input alongside intermediate results so a later step can notice the drift.
Frequently asked questions
When is one prompt better than a chain?
Whenever the task genuinely is one task. Every split adds a call, a failure mode and a place to lose context, so a chain has to earn its complexity in measured accuracy.
How do I stop errors compounding?
Validate between steps rather than only at the end, and pass the ORIGINAL input alongside intermediate results so a later step can notice the drift.
Was this cheat sheet useful?
Comments
No comments yet — be the first.
Need a different cheat sheet?
Tell us what you would like to see and we will build it — free.
Request a cheat sheet
