← ClaudeAtlas

learning-check-what-the-human-receivedlisted

Designs monitoring checks for an outreach engine that measure what the recipient actually got, after the last step between you and them, and gives every check a class, a way to go green and a quarantine. Use when adding, reviewing or trusting any automated check, alert or dashboard metric.
dostrovcev/outbound-agent-skills · ★ 0 · DevOps & Infrastructure · score 70
Install: claude install-skill dostrovcev/outbound-agent-skills
# Check what the human received A check asks its question after the last link between you and the recipient: the composer, the queue, the other table, their time zone. Name the corpus the check reads in a comment next to it. Every check also needs three things: a class, a way to go green, and a quarantine before it can wake anyone. ## Why A check that looks at the draft instead of the result is honestly green and completely useless. And because people trust green, a check on the wrong corpus breaks more, and more quietly, than no check at all. In our own engine the sameness check (how many near-identical texts a campaign sends) read the queue, where invitation notes live. The real messages had already left the queue. For one client it reported 2% duplicates. In the messages people actually received, duplicates were 48%. Checks also produce noise if nobody checks them. In one week our self-test grew from 10 checks to 48. On one day eight checks were red: two of them were false, and only one was a real breakage. The widest version of the same mistake: at one point we had 159 checks, and every one of them asked about our side of the pipe. One paying client had 524 invitations sent, a 17.6% accept rate and zero meetings in four weeks, and not a single light was red. ## How to apply 1. Write the question as a negative that can turn red: "a sender has no fuel", "a sent email has no paragraph breaks". 2. Trace the path from your record to the person. Put the check after the la