Youtube research ingestor
Ingest a YouTube URL — lecture, interview, tutorial, talk — into a structured, local, source-graded research artifact: metadata, transcript (official subtitles first, local Whisper fallback), structured summary, an extracted-claims table, references/entities mentioned, a verification-needed list, optional technical artifacts (commands/code/config) and frame notes, and an audit. Local-first (yt-dlp + ffmpeg + whisper), treats YouTube as a SECONDARY source, never promotes an unsupported claim to fact, and never fabricates references, DOIs, timestamps, or quotations. Modes: transcript | summary | claims | scientific-review | technical-review | teaching-notes. Use when the user gives a YouTube URL and wants notes, a transcript, claim analysis, extracted references/tools/papers, or reproducible technical steps — NOT for copyrighted reproduction, paywall bypass, or processing patient-identifiable clinical material without lawful authorization.From its SKILL.md
npx -y skills add kicrazom/scriptorium --skill youtube-research-ingestorAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
15.7 KB, as published. Nobody here has run it
youtube-research-ingestor
Turn a single YouTube URL into a durable, self-describing research artifact: a small directory of markdown files that carry (a) where the content came from, (b) the transcript, (c) a structured summary, (d) an extracted-claims table separating opinion from evidence, (e) references and entities mentioned, (f) what still needs primary-source verification, and (g) an audit of how the data was acquired and how much to trust it.
This is an INGEST skill in the scriptorium lifecycle. It produces the raw, sourced unit; it does not adjudicate truth.
Scriptorium integration
This skill composes with the rest of the toolkit — don't reinvent their jobs:
- Claim grading → use
epistemic-statusfor each row ofclaims.md(speculative → working → corroborated → operational → canonical, + contradicted), with the strict asymmetric-risk mode for clinical/medical claims. TheVerification statuscolumn here is the input to that skill, not a substitute. - Durable note / provenance →
field-note-from-urlwrites the canonical provenance-frontmatter note; this skill'smetadata.md+summary.mdfeed it. - Verifying claims against primary literature →
literature-searchskill or theresearch-scoutagent (PubMed/Crossref/OpenAlex), never the video itself. - Output location → defaults below are universal; if a
profile.knowledge_base.pathis set, write the artifact under it instead.
Purpose
Use this skill to ingest, transcribe, summarize, and critically analyze YouTube videos for research, education, scientific writing, technical learning, and project documentation. It converts a YouTube URL into a structured local research artifact: metadata, transcript, structured summary, key claims, source-quality assessment, missing references, possible unsupported statements, and reusable Markdown notes.
This skill is for analytical work, not entertainment browsing.
Use this skill when
The user provides a YouTube URL and asks to:
- summarize a video, or extract its transcript
- analyze the scientific or technical claims
- convert a lecture/interview/tutorial into notes
- identify claims requiring verification
- extract references, tools, names, papers, software, or datasets mentioned
- compare video content with scientific literature or official documentation
- build project notes for an Obsidian/GitHub knowledge base
Do not use this skill when
- the user asks for full copyrighted video reproduction
- the user wants to bypass paywalls or access restrictions
- the video is private or inaccessible
- the user asks to download, redistribute, or rehost copyrighted material
- the video contains patient-identifiable clinical data and lawful authorization has not been confirmed
- the task is answered better from primary literature, official documentation, or uploaded source files
Core principle
Treat YouTube as a secondary source unless the video itself is the primary object of analysis. For scientific, medical, legal, regulatory, or technical claims:
- do not treat the speaker's statement as automatically true
- mark unsupported claims
- recommend verification against primary sources (PubMed, DOI-bearing articles, official documentation, standards, guidelines, regulatory documents)
Required input
At least one YouTube URL. Accepted forms:
https://www.youtube.com/watch?v=VIDEO_ID
https://youtu.be/VIDEO_ID
https://www.youtube.com/live/VIDEO_ID
https://www.youtube.com/shorts/VIDEO_ID
Optional user instructions:
language: pl|en|auto
output_dir: ./youtube-research
mode: summary | transcript | claims | scientific-review | technical-review | teaching-notes
verify_claims: yes|no
extract_references: yes|no
extract_commands: yes|no
extract_code: yes|no
extract_frames: yes|no
Expected local dependencies
Prefer local-first tools. Run this preflight gate FIRST, before Step 1. If a required tool is missing, stop and report what to install — do not start a run that will fail halfway:
# Required: yt-dlp (metadata + subtitles + download), ffmpeg (audio/frames), python3.
for t in yt-dlp ffmpeg python3; do
command -v "$t" >/dev/null || echo "MISSING (required): $t"
done
# Optional, only for the audio-transcription fallback (Step 3) and frame timing:
for t in whisper faster-whisper jq; do
command -v "$t" >/dev/null || echo "missing (optional): $t"
done
- Any
MISSING (required)line → abort and surface the install command below. whisper/faster-whisperare needed only when no subtitles exist (Step 3); if they are also missing and subtitles are unavailable, report transcript-acquisition failure rather than guessing content.
Install (Ubuntu/Kubuntu) only with user approval, never on a shared/clinical/regulated host:
sudo apt update && sudo apt install -y yt-dlp ffmpeg python3 python3-pip jq
python3 -m pip install --user faster-whisper # optional
Output structure
Default naming pattern (honor profile.knowledge_base.path if set, else ./youtube-research/):
youtube-research/<date>_<video_id>_<short_slug>/
├── metadata.json
├── metadata.md
├── transcript.raw.txt
├── transcript.cleaned.md
├── summary.md
├── claims.md
├── references-mentioned.md
├── verification-needed.md
├── technical-artifacts.md # technical-review only
├── frame-notes.md # if frames extracted
└── audit.md
Example: youtube-research/2026-06-21_dQw4w9WgXcQ_ai_regulatory_affairs/
Workflow
Step 1 — Identify the video and capture metadata
yt-dlp --print "%(id)s" "<URL>"
yt-dlp --dump-json --skip-download "<URL>" > metadata.json
yt-dlp --print "Title: %(title)s
Channel: %(channel)s
Uploader: %(uploader)s
Upload date: %(upload_date)s
Duration: %(duration_string)s
Webpage URL: %(webpage_url)s" "<URL>" > metadata.md
Step 2 — Try official subtitles/transcript first
yt-dlp --list-subs "<URL>"
yt-dlp --skip-download --write-subs --write-auto-subs \
--sub-langs "en,pl,en.*,pl.*" --convert-subs srt "<URL>"
Prefer manually created subtitles over automatic ones. If only VTT is available, keep it and clean later.
Step 3 — Local transcription fallback (only if subtitles unavailable)
Use only for legitimate research, teaching, accessibility, or private analysis.
yt-dlp -x --audio-format mp3 --audio-quality 0 "<URL>"
whisper audio.mp3 --language auto --task transcribe --output_format txt --output_dir .
Do not upload audio to third-party transcription services unless the user explicitly requests it and confirms no privacy/confidentiality concerns.
Step 4 — Clean the transcript
Create transcript.cleaned.md. Remove obvious artifacts (repeated timestamps, broken
speaker labels, mid-sentence subtitle line breaks, duplicate auto-caption phrases).
Do not silently rewrite technical meaning — preserve names, tools, model/drug
names, equations, regulatory references, and citations. If transcript quality is poor,
say so in audit.md.
Step 5 — Structured summary (summary.md)
# Summary
## Video metadata
- Title / Channel / URL / Upload date / Duration
- Transcript source / Transcript quality
## Executive summary
Brief 5–10 sentence summary.
## Main points
1. … 5.
## Practical implications
- Research / Clinical / Software-AI / Teaching / Regulation
## Limitations of the video
- …
Step 6 — Extract claims (claims.md)
# Claims extracted from the video
| ID | Claim | Type | Support in video | Verification status | Notes |
|-----|-------|------|------------------|---------------------|-------|
| C01 | | scientific / clinical / technical / regulatory / financial / historical / opinion / anecdotal / speculative / promotional | strong / moderate / weak | verified / unverified / contradicted / partially supported / needs primary source / outside scope | |
If verification was not performed, do not label claims as verified. Grade survivors
with the epistemic-status skill.
Step 7 — References and entities (references-mentioned.md)
Sections: Papers/books/reports · DOI/PMID/arXiv/Zenodo/GitHub links · People ·
Institutions · Software/tools/models · Standards/regulations/guidelines · Datasets.
Mark incomplete references explicitly, e.g.
Incomplete reference: "the Stanford paper about medical LLMs" — needs identification.
Step 8 — Technical artifact extraction (technical-artifacts.md)
For technical videos, capture these sections:
- Commands — fenced
bashblocks - Code snippets — fenced blocks in the shown language
- Configuration files — fenced
yaml/jsonblocks - Libraries / frameworks — bullet list
- Architecture notes — bullet list
- Reproducibility concerns — bullet list
Do not invent commands or code not present in the video. Label every artifact:
[verbatim from transcript] · [reconstructed] · [inferred] · [needs verification].
Step 9 — Optional frame extraction (content-change, not fixed interval)
Only when visual content matters (diagrams, slides, charts, UI demos, on-screen code,
device setup). Prefer transcript-first analysis; do not overuse. (Needs the video, so
download it here: yt-dlp -f "bv*+ba/b" -o video.mp4 "<URL>".)
Default — scene-change detection. Capture one frame each time the picture changes
enough (a new slide, a chart swap, a UI step) instead of blindly sampling every N
seconds. scene scores frame-to-frame difference on 0..1; emit a frame when it exceeds
a threshold, and showinfo logs each kept frame's timestamp so you can fill the table.
mkdir -p frames
# THRESH knob: 0.3–0.4 = slide decks / talks; lower = more sensitive (subtle changes),
# higher = only hard cuts. Tune per video; start at 0.4.
ffmpeg -hide_banner -i video.mp4 \
-vf "select='gt(scene,0.4)',showinfo" -vsync vfr -frame_pts 1 \
frames/scene_%04d.jpg 2> frames/scene-times.log
# scene-times.log lines carry pts_time:<seconds> → map each frame to its timestamp.
Fallbacks (use only when scene-detect is wrong for the material):
# Talking-head with no slides, or you just want periodic checkpoints:
ffmpeg -i video.mp4 -vf "fps=1/30" frames/frame_%04d.jpg # every 30 s
# Known moment of interest:
ffmpeg -ss 00:05:30 -i video.mp4 -frames:v 1 frames/at_0530.jpg # single timestamp
frame-notes.md: table of Timestamp | Frame file | What is visible | Relevance | Needs verification.
Pull timestamps from scene-times.log (pts_time). If the threshold yields too many or
too few frames, re-run with a different value rather than keeping noise.
Step 10 — Audit (audit.md)
Record: data acquisition (URL, metadata yes/no, transcript source, audio/video/frames downloaded), quality assessment (transcript quality, speaker attribution, technical-term reliability, language, major risks), copyright & privacy (no full copyrighted content reproduced; no patient-identifiable data without authorization), scientific reliability (YouTube = secondary; unsupported claims not promoted), and open issues.
Verification policy
For medicine, science, law, finance, regulatory affairs, software dependencies, or
current events: verify externally before producing final conclusions; cite DOI, PMID,
official documentation, standards, or regulatory documents. If verification was not
done, write Verification not performed in this run.
For medical content: never infer clinical recommendations solely from YouTube; distinguish education from clinical guidance; flag outdated statements, missing guideline references, and stated/visible conflicts of interest.
Output modes
| Mode | Produces |
|---|---|
transcript | metadata.md, transcript.raw.txt, transcript.cleaned.md, audit.md |
summary | + summary.md |
claims | metadata.md, transcript.cleaned.md, claims.md, verification-needed.md, audit.md |
scientific-review | summary + claims + references-mentioned + verification-needed + audit |
technical-review | summary + technical-artifacts + claims + audit |
teaching-notes | summary + teaching-notes.md + questions.md + audit |
scientific-review also flags: unsupported biomedical claims, overgeneralizations,
conflicts with guidelines, missing comparator/methodology data, claims needing
DOI/PMID. technical-review also flags: reproducibility gaps, missing versions/hardware,
hidden dependencies, unclear commands, security and licensing concerns. teaching-notes
structures learning objectives, key concepts, common misconceptions, discussion
questions, and a practical exercise.
Style requirements
Concise structured Markdown; tables for claims/references/tools/entities; mark
uncertainty explicitly with labels: [verified] [unverified] [needs source]
[inferred] [speculative] [contradicted] [low transcript quality]
[AI Model-Based Prediction]. Do not fabricate references, DOIs, names,
timestamps, or quotations.
Quotation policy
Avoid long verbatim quotations. Short excerpts only when necessary for analysis; include a timestamp if available; prefer paraphrase for summaries. Do not reproduce substantial copyrighted transcript sections.
Safety and privacy constraints
- Do not process patient-identifiable clinical material without confirmed lawful authorization.
- Do not upload confidential material to third-party APIs.
- Do not store sensitive transcripts outside the project directory.
- For GitHub commits, commit only summaries, user-authored notes, metadata, claim tables, and verification notes — not full transcripts unless the source license explicitly permits it.
Minimal command template
URL="<PASTE_URL_HERE>"
OUTDIR="youtube-research/$(date +%F)_$(yt-dlp --print '%(id)s' "$URL")"
mkdir -p "$OUTDIR" && cd "$OUTDIR"
yt-dlp --dump-json --skip-download "$URL" > metadata.json
yt-dlp --print "Title: %(title)s
Channel: %(channel)s
Upload date: %(upload_date)s
Duration: %(duration_string)s
URL: %(webpage_url)s" "$URL" > metadata.md
yt-dlp --skip-download --write-subs --write-auto-subs \
--sub-langs "en,pl,en.*,pl.*" --convert-subs srt "$URL"
Done criteria
A run is complete only when: metadata was captured or its failure documented; transcript was acquired or its failure documented; transcript quality was assessed; the selected mode's output files were created; unsupported claims were separated from verified facts; limitations were stated; and no copyrighted or confidential material was inappropriately reproduced.
Recommended final response to user
On success, list the created files, then give one-line Key result and Main limitation. On transcript-acquisition failure, state the reason and offer: provide another URL · provide downloaded subtitles · allow local audio transcription · use manual notes.