Help me design what the user sees while an agent works. THE AGENT: [what it does]. WHO USES
IT: [role, how busy, how technical]. TYPICAL DURATION: [seconds]. TYPICAL STEP COUNT: [easy
and hard]. WHAT IT PRODUCES: [output]. Design: the actual on-screen text during a typical run,
step by step as it appears; the result view, showing what's default versus behind 'show
detail'; what happens when the user clicks stop halfway and what they keep; the message when
it gives up at a cap, written to be useful rather than an error; what the first-ever run
should do differently from the hundredth; and three things that would damage trust
irreversibly on a first use.
Help me rehearse an incident before launch. THE AGENT: [what it does]. ITS TOOLS: [list,
read/write marked]. APPROVALS: [what needs a human yes]. VOLUME: [tasks per day]. WHO'D BE ON
CALL: [roles]. Write three realistic incident scenarios in increasing severity, specific and
plausible given my tools, not generic. For the worst, walk the timeline hour by hour: when it
starts, when someone notices, what they see first, what they do. At each point name what we'd
need to detect or stop it, and ask whether we have it. What's the cheapest change that would
most reduce the damage? Then write the one-page runbook: how to stop it, who can, how to work
out what it did, how to reverse what's reversible. Assume the reader is stressed and it's
late.