YouTube Research with Scribefy

You have four Scribefy tools. Three are free; one costs credits. The core discipline of this skill: spend free calls to make every paid call count.

The workflow

1. Search wide (free)

Run search_videos with 2–3 different phrasings of the question — YouTube search rewards different keyword angles ("PO3 trading strategy" vs "power of three ICT explained"). Collect ~5–10 candidates.

2. Vet before you spend (free)

Call get_video_metadata on the shortlist. Decide using:

• Duration — extraction cost scales with length (1 credit for short

videos up to 8 for very long ones). A focused 12-minute video usually beats a 3-hour podcast that mentions the topic once.

• Caption tracks — the metadata lists every track and marks it

✏ authored or ⚙ auto-generated. Authored captions give cleaner text. No caption tracks listed → do not attempt extraction (it will fail with NO_CAPTIONS); pick another candidate.

• Recency and channel — for fast-moving topics, check the upload date

before trusting the content.

3. Extract selectively (paid — usually 1–2 videos, not 10)

Call extract_transcript on the best 1–2 candidates first. Read them. Only extract more if the question is still open.

• Cached transcripts cost 0 credits — popular videos are often already

cached, and re-extracting the same video is always free. Never hesitate to re-pull a video you've extracted before.

• For non-English videos, pass the language code of the caption track you

saw in metadata. Extract in the original language and translate in-chat — that preserves nuance better than hunting for a translated track.

4. Synthesize with timestamps

Transcripts arrive as timestamped segments. When answering, cite moments as [mm:ss] next to claims so the user can jump straight to them in the video. For multi-video research, attribute each point to its video title.

Multi-video research pattern

For "what do people say about X" questions:

1. search_videos (2–3 phrasings) → candidate pool
2. get_video_metadata on each → build a quick table: title / channel /

duration / captions / date

1. Pick the 2–3 strongest, state the expected credit cost, extract
2. Synthesize a comparison: agreements, contradictions, unique points —

each cited with video + timestamp

To go deeper from one good video, walk the recommendation graph with get_related_videos (free) instead of re-searching.

Error handling

Cost etiquette

Before extracting anything long (cost ≥ 4), tell the user the credit cost from the metadata and confirm. Never bulk-extract a whole search result list — vet, rank, extract the minimum, expand only if needed.