Seven days of my mailbox: 1,919,892 tokens of raw MIME.
A plain-text mirror of the same week: 21,481.
89× fewer — tokenized in full, not sampled. A month of mail fits in one prompt.
https://t.co/ti0gJ9F1eC
@fabrice_mayrand inbox.txt – llms.txt for your inbox (measured: 89× fewer tokens)
Seven days of my mailbox: 1,919,892 tokens of raw MIME.
A plain-text mirror of the same week: 21,481.
89× fewer — tokenized in full, not sampled. A month of mail fits in one prompt.
https://t.co/ti0gJ9F1eC
The mirror is a text file an agent reads, so it's a prompt-injection surface by construction. Bodies are labelled untrusted, spam gets an index line but never a body, secrets redacted at sync.
Draft spec + reference impl + the tokenizer, so you can run it on your own mail.
Seven days of my mailbox: 1,919,892 tokens of raw MIME.
A plain-text mirror of the same week: 21,481.
89× fewer — tokenized in full, not sampled. A month of mail fits in one prompt.
https://t.co/ti0gJ9F1eC
Almost none of raw mail is information. One payment receipt: 60,000 characters of HTML tables, 600-char tracking URLs and quoted history, carrying ~350 chars a human would call "the email."
It also tokenizes badly: 2.23 chars/token vs 3.29 for clean text.
@LesterWikid @FreedomOvrFear1 @ianmiles@grok How dare you abusing to easiest way to do thing in order to get what you want. Get down in the trench getting durty for your informations with the rest of us Dave.