Skip to content
AI Knowledge MapSearch / Text directory

How should I resolve conflicting length and detail requirements?

C013 · C Controlling changes · Output scope and expression

Key answer

Decide whether length is a hard ceiling or approximate target, then identify information to retain, remove or move elsewhere. “As short as possible” plus “explain every detail” leaves AI to guess priorities.

Full explanation

The 2023 IFEval study made word, sentence and paragraph counts programmatically checkable constraints in 541 prompts tested on GPT-4 and PaLM 2. Length compliance can be checked separately; it does not establish that necessary content survived or explanations are clear. Sections 1–2 and Table 1

Define counting: “body at most 500 characters, including punctuation, excluding title and references” is clearer than “500 characters.” If the goal is simply brief reading, request three paragraphs for conclusion, reasons and next step. Use a platform's own counter for its limits.

Then specify deletion order: repeated background and ornament first, minor examples next. Explicitly preserve conditions changing the conclusion, key data and necessary actions. Appendices are appropriate only when the format permits them, not as a loophole around a whole-document limit. For example:

Write a beginner explanation of at most 500 characters in the body, including punctuation and excluding title/references. Retain the conclusion, applicability conditions and one complete example. Remove repetition before minor background. Do not shorten “usually effective” to “always effective.” If it still cannot fit, state the extra space needed and cuts requiring my choice.

These are checkable editing rules; priorities follow purpose. Decision-makers may need conclusions and costs first, operators steps and failure handling.

Check twice: measure actual length with an editor/script, then compare key content and causal relationships against sources. AI's “498 characters” is unverified. If readers still cannot act, reconsider information selection even when the count passes.

References

1. Zhou et al. (2023), Instruction-Following Evaluation for Large Language Models, preprint, Sections 1–2 and Table 1