Asking Claude to write exactly 50 words is a reliable way to get 80 words, a polite acknowledgment when you complain, and then 90 words. It's one of the more disorienting failures in an otherwise capable system, and the reason is architectural.
The Token Problem
Claude doesn't process text as words. It processes tokens - fragments of text that roughly map to syllables or word parts. "Unprecedented" might be 3-4 tokens. "I" is one token. A 50-word target and a 50-token target are completely different things, and the model operates in tokens, not words.
More importantly, Claude generates text one token at a time, predicting what comes next based on everything written so far. There is no parallel counter tracking the word total. By the time the model is approaching 50 words, it's already committed to a sentence structure that might carry it to 70. Backtracking isn't in the design.
This same limitation explains why LLMs struggle to count letters in a word, reliably identify the third sentence of a paragraph, or produce exactly N items in a list. Counting requires a separate operation from language prediction - and that operation isn't built in.
Prompting Approaches That Actually Work
Structural constraints beat raw word count targets:
- Sentence limits: "Write this in exactly 3 sentences" is more reliable than "50 words." Sentence completion is a natural stopping point the model already understands.
- Bullet structure with caps: "5 bullet points, one sentence each" works better than an overall count because each item is its own constraint.
- Approximation over precision: "About 50 words" produces better results than "exactly 50 words." The model handles approximation more gracefully than strict precision.
- Explicit in-progress counting: Telling Claude to count words as it writes - "count your words as you go and stop at 50" - improves accuracy because it forces mid-generation evaluation. It's not foolproof, but it helps.
- Two-pass editing: Generate a first draft, then ask Claude to count the actual word total, then ask for a targeted trim or expansion. Iterative editing beats trying to nail the count in one shot.
The honest baseline: current language models weren't designed for word-precise output. The workarounds reduce the gap but don't close it. For workflows where exact word counts matter - SEO meta descriptions, paid ad copy, legal character limits - plan to verify and edit manually. That's not a unique Claude failure. It's a property of how all these models currently work.