The lesson
Turn a surprising result from your own benchmark into a headline, then use the data to argue a broader position your product happens to serve.
The essay opens with a provocative title and a TL;DR stating that all ten Claude setups finished in the bottom eleven of the 25 combinations, then explains the eval method and shows a bar chart of scores. Question headings walk the reader through the investigation: 'How did Claude score in AgentMail Evals?', 'Did Claude fail to find the hidden emails?', and 'Which jobs caused the gap?', each answered with a table that narrows the gap down to replying to mail flagged as spam. A highlighted line, 'Claude's default is no, and everyone else's is yes,' sets up the thesis that a default is not a policy, and 'How do you set email security rules for an AI agent?' closes by pointing to AgentMail Lists and Drafts.
Don't copy the jab at a named model or the spam-reply finding, which depend on AgentMail's own eval. Keep the structure: lead with a surprising result, rule out explanations one heading at a time, name the real cause, then move to a stance on what the reader should do.
Tell us what you want to make. We’ll talk through how to put these ideas to work for your brand.