Human Oversight Evidence Pack

What it does

This skill turns an agent transcript and a reviewer's own account of their oversight into a structured evidence pack: an oversight log, a decision trail, a reviewer attestation, and a sign-off block. It documents that a competent human reviewed AI-assisted work, in a form auditors, clients, and internal governance can read.

The problem it tackles

As AI-assisted output enters regulated and client-facing work, the artifact people increasingly need is not just the work product but proof that a human reviewed it. Auditors, clients, professional-liability insurers, and human-oversight expectations for AI all converge on the same question: who reviewed what, what they changed or questioned, and what they signed off on. This is structured review documentation, generalized from regulated industries to any field.

How it works

Give it the AI-assisted work product or transcript plus your account of what you actually checked. It reconstructs the decision trail from the transcript, maps your stated actions onto each decision point (reviewed, changed, questioned, accepted, or not reviewed), and drafts a first-person attestation for you to verify and adopt. Anything you did not confirm is surfaced as an open gap rather than assumed.

What you get

A clean pack: oversight summary, decision trail with a review status per item, your reviewer actions in your own words, a proposed attestation with an unsigned sign-off block, and an explicit open-gaps list. The bundled templates file gives reusable oversight-log, attestation, decision-trail, and sign-off formats.

One principle runs through the whole skill: it documents review that genuinely happened and never fabricates a review step, finding, name, date, or signature. The sign-off block stays blank until a real person signs it. It documents the review process, not the quality of the work, and it does not certify compliance with any law or standard. This is a documentation aid, not legal, audit, or compliance advice.