A weekly report is produced by one prompt that reads six data files, summarises each, finds cross-file trends, and drafts an executive note. The final note is usually good but occasionally states a trend that contradicts the underlying files, and nobody notices until a reader queries it. Total cost may rise, but errors must become detectable. What is the strongest change?

The principle — An error you cannot see is worse than an error you can. Decomposition is usually taught as a way to improve output, but its more valuable property here is that it creates checkpoints where a wrong result is still cheap to catch.

Why the key is correct — This prompt performs four distinct cognitive actions in one pass — summarise, summarise across files, infer trends, and draft — which is the exact pattern Google's checklist calls trying to accomplish too much, with the stated remedy of breaking the requests into separate prompts. Google's chaining guidance describes what to build instead: make each step a prompt, with the output of one step becoming the input of the next. The consequence that matters for this constraint is the one Anthropic names — each step being a separate call is what lets you inspect intermediate outputs, log them, or branch at any point. Once the per-file summaries exist as their own artefacts, a trend claim can be checked against them, and the contradiction that currently reaches a reader is visible one stage earlier. The constraint explicitly permits higher total cost, which is what pays for the extra calls.

Why the others are wrong — Asking for visible reasoning in a single response produces an explanation generated alongside the answer, which can rationalise the same error rather than reveal it. Running the same overloaded prompt three times measures stability rather than correctness, and consistent agreement is exactly what a systematic error looks like; it also never tells you which of the four actions failed. A recap is a real prompt component and does reinforce constraints, but it leaves the four actions bundled and produces no intermediate artefact anyone can audit.

Remember this — Decomposition buys checkpoints. That is often worth more than the quality gain.

Sources — Google Cloud, Overview of prompting strategies; Anthropic, Prompting best practices.