| name | to-notes |
| description | Converts a video URL, local video/audio file, or existing transcript/text into structured, concept-first lecture notes. |
| disable-model-invocation | true |
| license | Apache-2.0 |
| metadata | {"author":"Oleg Shulyakov","version":"1.1.0","source":"github.com/olegshulyakov/agent.md","catalog":"productivity","category":"notes","tags":["notes","transcript","video","lecture-notes"]} |
To Notes
Turns a video URL, local video/audio file, or existing transcript into concept-first lecture notes — organized by idea, not by transcript chronology. No speech-to-text: this skill assumes a transcript already exists or can be fetched as captions. Does not handle .pdf/.docx/.pptx/.xlsx/.msg/.epub documents.
Step 1 — Resolve the source to plain text
If the source is a file attached in this chat rather than a typed/local path, resolve it using this harness's convention for attached-file paths first.
python3 scripts/resolve_source.py <url-or-path>
The script handles all three source shapes:
- Video URL — fetches captions via
yt-dlp (never downloads the video). Prefers manual captions over auto-generated; prefers en, falling back to whichever single language track is available. Requires outbound network access — unavailable in some sandboxed cloud sessions.
- Local video/audio file — looks for a sibling
.vtt/.srt/.txt transcript next to it, matching either the exact stem (talk.vtt) or a language-coded variant (talk.en.vtt).
- Transcript/text file (
.vtt/.srt/.txt/.md) — reads it directly, stripping timestamps/cues if present.
If the input is pasted transcript text rather than a file or URL, skip the script and use the pasted text as-is.
The script prints LANGUAGE: <code> on the first line, a --- separator, then the plain transcript text.
Done when the script prints resolved transcript text, or exits with a specific error (no manual/auto captions found, no sibling transcript, unsupported file type, no network reaching the URL) — never fall back to transcribing audio yourself. If a URL fetch fails for lack of network access, ask the user to paste the transcript text or upload a caption file instead.
Step 2 — Draft concept-first notes
Read the resolved transcript and draft lecture notes organized by the source's actual ideas — group related points under topic headers, do not follow the transcript's timeline.
Save the notes as <kebab-case-title>.md:
- Local/CLI session: write it to the current working directory, unless the user names a different location.
- Hosted/cloud session: write it wherever this harness's convention places generated files for the user, and use its file-delivery mechanism (e.g., present/download) if one exists.
Use this frontmatter:
---
title: <title>
author: <speaker/author, if identifiable, else omit>
source: <original URL or file path>
---
If the resolved language (from Step 1's LANGUAGE: line) isn't en, add a language: <code> field. Redact sensitive information (API keys, passwords, personally identifiable information) before writing the notes.
Done when the file is written with complete frontmatter and headers reflect the source's real concepts rather than its chronology.
