SKILL.md
- name:
- auto-reply-is-not-a-reply
- description:
- Use this skill when classifying inbound responses to outbound — out-of-office, vacation and parental-leave auto-responders, delivery notices, and other machine-generated mail that looks like engagement. Produces a triage decision that keeps the sequence alive, keeps the metrics clean, and reschedules the next touch to when the human is actually back.
An auto-reply is not a reply
An out-of-office arrives in the same inbox, in the same thread, as a real answer. Filed as one, it corrupts three things at once. This is how to file it.
The play
- Classify machine mail as its own outcome, alongside interested, question, referral, not now, not interested and unsubscribe. A triage vocabulary without this bucket does not skip the auto-responder — it files it under the nearest wrong label.
- Read the return date if the message states one; never infer one. "Back on the 20th" is a date. "Away for a while" is not, and a guessed date is a send at the wrong time dressed as intelligence.
- Reschedule instead of advancing. The touch was never spent: nobody read the message. Move the next one to one working day after they return — landing on their first morning, behind four hundred emails, wastes the sequence you just protected.
- Record nothing. Not a reply, not a positive, not a stage change. This is the step everyone forgets, and it is the expensive one.
- Keep the account in the sequence. Automatic mail must never mark an account worked, closed or exhausted.
- File delivery notices separately from auto-responders. "Temporarily deferred, will retry for 47 hours" is the mail system, not the person: no date, no reschedule, no bounce. A hard failure is a different outcome with a different consequence — suppression.
What good looks like
- The tell: it is not the vacation message that hurts, it is the metric it writes. One auto-responder counted as a reply inflates the reply rate of whichever template earned it, and every downstream decision — which sequence to scale, which segment to keep — is then made on a number a robot wrote. On low volume it is the difference between 2% and 4%.
- The mediocre version treats any inbound as engagement: the follow-up stops, the account moves to "in conversation", and someone drafts a reply to a robot. Six weeks later that account is in the "already worked, no answer" pile, having never once been read by a human.
- You know the output is good when auto-responders are visible and boring: a named outcome you can count, a next touch dated after the return, and no movement in your reply rate when a wave of holiday mail arrives.
Rules
- MUST give machine mail its own outcome, never the nearest human label.
- MUST leave the sequence active and the touch unspent.
- MUST exclude it from reply, positive and engagement metrics.
- MUST use a stated return date, and only a stated one.
- NEVER draft or send an answer to an auto-responder.
- NEVER mark an account worked, replied or closed on machine mail.
- NEVER treat a temporary delivery notice as a bounce — retry-language means the address may be fine, and suppressing on it kills a good contact permanently over a server hiccup.
