Why is breaking a complex request into smaller prompts generally more effective than sending one large prompt?

The principle — A single response has to satisfy every request in the prompt at once. The more distinct things it must do, the more each of them competes for the same limited attention and the same output.

Why the key is correct — Google's guidance is to manage complexity by breaking things down into simpler components, and its checklist names the specific failure: a prompt asking the model to perform several distinct cognitive actions in a single pass is likely trying to accomplish too much, and the requests should be broken into separate prompts. Separating them gives each step a full response, and Anthropic notes the further benefit that each step being its own call is what lets you inspect intermediate output, log it, or branch on it. The checkpoint between steps is where a wrong turn gets caught before it propagates.

Why the others are wrong — Input length is not the issue; task count is. Chained steps are sequential, since each consumes the previous output. And repeating the whole context in every step would undo the split rather than preserve coherence.

Remember this — One prompt, one job — and read the output before you spend it on the next step.

Sources — Google Cloud, Overview of prompting strategies.