The advice to give the model more context is right up to a point, and then it stops being right. Padding a prompt with everything you have makes the important parts harder to find.
Relevance beats volume
A model reading twenty pages to answer a question about one paragraph will do worse than a model given that paragraph plus a sentence of framing. Pasting your entire knowledge base into every prompt is not thoroughness, it is noise.
What to include
- The material the answer must be based on, in full and in its original wording.
- Constraints that change the answer – audience, format, length, tone.
- What has already been tried or ruled out.
- How you will judge whether the answer is good.
What to cut
- Background that is interesting but does not change the answer.
- Your own summary of material you are also pasting in full.
- Instructions that contradict each other.
The contradiction problem
Long prompts accumulate contradictions. Be thorough sits three paragraphs above keep it under 100 words, and the model has to pick one. Read a long prompt back looking specifically for instructions that cannot both be satisfied. That single pass fixes more bad output than adding anything new.
When output gets worse as you add context
That is the signal to stop adding and start cutting. Strip the prompt back to the minimum that could possibly work, confirm the quality returns, then add back one piece at a time.
