youtube-summary
Skill0 starsChanged 5 months agoMITView on GitHub
Use when the user shares one or more YouTube links to save/summarize, or asks about previously saved YouTube videos (what was in video X, which video covered Y, find the one about Z). Covers ingesting links into the knowledge base and answering questions from the saved cards.
- Reads credentials
4 other projects use a version of this.
Adopted by
Most recently taken up by
- phuocphn/galleytheir own version, first seen Oct 20, 2026
- parkbarg/claude-skillstheir own version, first seen Oct 3, 2026
- ParthGanatra/agent-skillstheir own version, first seen Aug 5, 2026
- dandacompany/claude-skillstheir own version, first seen Oct 20, 2025
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- ParthGanatra/agent-skills★ 6same name
- dandacompany/claude-skills★ 2same name
- phuocphn/galley★ 1same name
- parkbarg/claude-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 2 are distinctive, and 4 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 23
Default engine openrouter.
- Very specific: line 32
If something landed in <top>/_inbox (no taxonomy match), say so and offer to add a subfolder to taxonomy.yaml.
- Very specific: line 35
Run the ingest command with --engine self → each video answers {"status":"need_card","prompt_file":...}.
- Very specific: line 36
Read prompt_file (SYSTEM + USER sections), write the card strictly to that JSON schema, including the category field, into a temp file.
- Distinctive: line 40
Never answer from memory. Follow every step:
- Distinctive: line 45
Verify: every claim you make is present in a card you read.
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
youtube-summary
Single entry point for the YouTube knowledge base. The pipeline is deterministic (scripts in this repo) — you pick the engine and read results; you never write cards or edit the index by hand.
REPO (run everything from here, it has its own venv with all deps):
cd ~/.claude/skills/youtube-summary then use .venv/bin/python.
(This is the install path from the README; adjust if the repo lives elsewhere.)
When to use
- The user shares YouTube link(s) → INGEST them.
- The user asks about saved videos → QUERY the knowledge base.
INGEST — links → knowledge base
Pass ALL the user's URLs in one command (one venv start, one git commit):
cd ~/.claude/skills/youtube-summary && .venv/bin/python -m scripts.yt_core "<url1>" "<url2>" --config config.yaml- Default engine
openrouter. Add--engine local(private, via ollama — distill AND classify stay offline) or--engine self(you summarize, see below) only if asked. - Force a folder with
--category top/subonly if the user specifies it; otherwise it auto-classifies. Tops come fromtaxonomy.yamlin the KB repo (default: skills, reviews, startup, random). - Output: one JSON line per video, then always a final
{"git": ..., "videos": N}line (videos= how many were saved/refiled). Statuses:ok— saved;title,tldr,file,top/subare in the line — report from them, no need to read the card.exists— already in the KB; nothing re-run. Tell the user; add--forceto re-process, or pass--category top/subto move it (cheap re-file, no LLM).refiled— an existing card was moved to the requested category.need_card— engine=self only, see protocol below.failed— reportreasonto the user (this includes a typo'd--category: the pipeline refuses rather than guessing).
gitin the final line:pushedis full success;committed= saved locally, push disabled (--no-push);clean= nothing changed (e.g.--forceproduced identical content) — fine, not an error;push_failed= committed locally but not pushed — tell the user;failed= the commit itself failed;skipped= nothing was saved so no commit was attempted;disabled= the KB folder is a plain directory (default no-git setup) — normal, but if the user asks about syncing/versioning, offer togit initit.- If something landed in
<top>/_inbox(no taxonomy match), say so and offer to add a subfolder totaxonomy.yaml.
engine=self protocol (you are the summarizer)
- Run the ingest command with
--engine self→ each video answers{"status":"need_card","prompt_file":...}. - Read
prompt_file(SYSTEM + USER sections), write the card strictly to that JSON schema, including thecategoryfield, into a temp file. - Rerun with
--card-file <path>and that video's SINGLE url — the pipeline continues (classify → write → index → commit). With several videos, do one rerun per video.
QUERY — questions → grounded answer
Never answer from memory. Follow every step:
- Retrieve:
cd ~/.claude/skills/youtube-summary && .venv/bin/python -m scripts.kb_query "<question>" --config config.yaml- Hybrid search: vector + keyword over the cards; returned
file/media_dirpaths are absolute. - "что я сохранял недавно / на этой неделе" → use
--recent N(newest first) instead of a question;--top skillsfilters by category.
- Hybrid search: vector + keyword over the cards; returned
- Read the actual card files (the
filepaths returned) — the full card, not just the snippet. - Verify every claim you make is present in a card you read. Not found → say so, don't invent.
- Answer concisely: video title, timecoded deep-links from
takeaways, and the screenshot (media_dir) when the question is visual.
Output rules
- "Только важное / без воды" → use
tldr+ key takeaways, not everything. - "Не повторяй прежнее" → don't resend a card already shown earlier in this conversation.
- "Подробнее про это видео" → read its full card and expand only from there.
Maintenance
- New machine / empty search results but cards exist → rebuild the vector index without any LLM calls:
.venv/bin/python -m scripts.reindex --config config.yaml
Config & secrets
config.yaml (gitignored, per-machine) sets kb_repo (empty = knowledge-base/ folder next to the skill, no git; a git clone path = auto commit+push), distill engine/models and card language, caption language priority, Whisper model, vector store, embedder. OPENROUTER_API_KEY comes from env_file or the environment. See README.md for setup.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

