## 1. Set the publication window and video type

Restate the window, long-video rule, item limit and fields. Classify all four records by publication time. Do not search for more videos. Save scope.md.

Expected: Four scope decisions.
Check: Exclude old Y04 and Short Y02.

## 2. Record identity and content coverage

Record selected ids, titles, publication/observation times, views, transcript and thumbnail status. Keep missing fields null; invent no URLs or transcripts. Save observations.csv.

Expected: Y01/Y03 observation rows.
Check: Y03 is not zero; neither video is described as fully read.

## 3. Check age and sample before using a baseline

Calculate Y01’s descriptive ratio to the 10-video mean of 1000. State age matching is absent. Make no significance or causal claim. Exclude missing Y03 from ratios and do not use Y02 to fill the baseline.

Expected: A qualified 2× ratio.
Check: Do not call it a proven hit or thumbnail-driven growth.

## 4. Separate facts from editorial judgment

Write the report with selected/excluded ids, statistics status, descriptive comparisons, research ideas and missing material. Label ideas as judgment; cite facts by id. Do not evaluate unread transcripts or thumbnails.

Expected: A complete report and reading queue.
Check: Facts, suggestions and coverage limits remain distinct.

## 5. Validate real reads and Notion deduplication separately

Read at most 10 videos from <public channel/window>. Report pagination and fields. After review, inspect <Notion data source> schema and video_id keys; if specifically authorized, append only new items and read back page ids. Check duplicates on unclear results.

Expected: A bounded live observation or blocker.
Check: Existing videos are not repeatedly added; unavailable fields are explicit.

## 6. Separate new videos from updated statistics

Draft a video-watch Skill with window, channels, type evidence, video_id, observation times, baseline version, content coverage and failure handling. Propose a Routine without creating it; log discoveries and statistics updates separately.

Expected: A versioned monitoring method.
Check: Reruns do not duplicate videos or hide access failures.