Agent-written reports should open with two or three lines a human has checked, with the full AI text folded underneath. This keeps people reading the output of an AI agent instead of skipping it.
Simon Willison now gets most of his code changes as pull requests opened by Claude Code or Codex. The descriptions are accurate and complete, and that is the problem: they read like slop, so nobody wants to read them. His fix is a short opener he picks himself, sometimes with a screenshot. The agent's full essay goes into a collapsible HTML block labelled with the model name, such as "Claude Opus 5.5 Medium".
What this means for a small business
The same problem shows up outside code. An agent that summarizes leads, contracts or supplier invoices produces walls of correct text, and after a week the manager stops opening the messages. In a tender analysis pipeline I built, the useful part is the penalties and deadlines, not the page-by-page breakdown. Put the verdict first in a fixed format, such as risk, amount and action needed. Put the rest behind a "details" fold or a link. In Telegram or email, that means one short message plus an attachment. Labelling which part came from the model also tells people how much to trust it.
Why it matters: an AI pipeline only helps if people read what it produces. Design the first three lines of every agent output as carefully as the prompt itself.
Source: Simon Willison