Why is 'answer in exactly 50 characters' harder than it sounds for an LLM?

Answer: It predicts token steps

It cannot count words — Counting words is the wrong target here. The user asks for characters, while the model generates one token step after another. Exact length needs a separate character-count check.

It predicts token steps ✓ — The model produces text as a sequence of tokens, then those tokens decode back into characters. Because one token can become one letter, part of a word, punctuation, or a whole word, exact character budgets are awkward without a separate checking step. Generation happens in token steps, while the requirement is measured in characters.

It forgets punctuation — Punctuation can cause off-by-one mistakes, but punctuation is just one example of the deeper issue. Spaces, capitalization, and partial words all count as characters while being tokenized in uneven ways. The mismatch is structural, not merely forgetfulness.