Trace a wrong letter count back to the tokenizer

The model counted wrong, and it sounded completely sure.

BeginnerRead 6 minDoing: 20 to 30 minutes with a chat window open7 steps

What you'll have at the end

A token-by-token breakdown that explains the miscount, plus a rewritten prompt that gets the count right

Before you start

You need

Access to a chat model you can send plain text prompts to, and a free tokenizer viewer open in another tab to see how it splits a word into pieces.

Not covered

This does not cover getting a model to do exact arithmetic on large numbers, or building your own tokenizer from scratch.

Leans on

Budget your tokens before you send them

for counting how many tokens (not letters) your prompt uses so you can control cost and context limits, a different kind of counting than this one

Checked 15 Aug 2026