Skip to content
AssureSwarm Docs

Agent Operating Guide

/sox-from-video

Turn a recorded auditor walkthrough and its transcript into annotated SOX workpaper tabs, with matching video frames boxed by movable red Excel shapes over each tested field.

Takes a recorded auditor walkthrough, a Zoom, Teams, or Loom mp4, plus a transcript and produces the same SOX workpaper format the other evidence skills do: a Summary tab over per-sample-per-test detail tabs, each carrying the red-bordered narrative on top and annotated video frames below, with red-rectangle Excel shapes over each tested attribute (movable shapes, never burned-in pixels). The skill is self-contained: frame extraction is vendored, and vision is the model itself, no OCR.

It never reads the transcript or views a frame itself. A sox-walkthrough-parser turns the transcript into a list of review-moment timestamps, a vendored extractor cuts the matching frames, and then the same sox-evidence-boxer + sox-evidence-context + sox-evidence-reviewer fan-out as /sox-annotate-xlsx locates, reads, and verifies each attribute, one agent per frame, in parallel, while the skill auto-judges pass/fail from their JSON.

When to use

Section titled “When to use”

When the evidence for a control is a meeting recording with a transcript, and you want one annotated frame per attribute the auditor verified without scrubbing the recording by hand. If the evidence is already a screenshot deck in xlsx form, use /sox-annotate-xlsx; to plan a test before reviewing evidence, use /sox-testing.

Inputs

Section titled “Inputs”
Flag Required Notes
<video.mp4> Yes The recording. Other opencv-supported formats (.mov, .mkv) also work.
<transcript> Yes .vtt / .srt (real timestamps, preferred) or .txt / .docx (timestamps estimated from word position).
--workpaper <path> No Destination. Defaults to workpapers/<control-id>/workpaper.xlsx.
--no-review No Skip the on-by-default box-position review pass on every frame.
--no-context No Skip the on-by-default value-and-context scan on every frame.
--max-review-iters <N> No Boxer revisions the review loop may request per frame (default 2, ceiling 3).

The control ID, the samples in scope, and the test attributes are usually supplied by the calling /sox-testing invocation; otherwise the skill asks once.

Example

Section titled “Example”

/sox-from-video walkthrough.mp4 walkthrough.vtt dispatches the walkthrough parser to find each (sample, test) review moment, extracts the matching frames, boxes and reads each tested field via the leaf agents, auto-judges the results, and writes per-sample-per-test detail tabs with annotated frames plus updated Summary rows.

Good to know

Section titled “Good to know”
  • Transcript quality drives frame quality. A vague transcript with no field names yields noisy timestamps: the skill asks before extracting dozens of frames.
  • Estimated timestamps drift. .txt / .docx transcripts have no real timestamps, so moments are estimated (~150 wpm) and frames can catch the auditor mid-scroll; .vtt / .srt are exact.
  • Gaps are surfaced, not swallowed. If the parser finds no frame for a (sample, test), you decide before extraction rather than silently leaving cells n/a.
  • One detail tab per (sample, test). If the auditor revisits an attribute, its frames stack in transcript order on the same tab.
  • Residual drift is flagged. Boxes the reviewer couldn’t fully resolve after the retry cap are noted in the Summary reasoning cell for a human glance.
Section titled “Related”

Not audit or legal advice. Workpapers and assessments produced by these skills require review by qualified financial professionals before being relied on for SOX 404 compliance.