dasein108/yt-mem-ai / skills/yt-agent
yt-agent
Verifiedyt-agent — summarize / highlights / Q&A / presentation / digest / review
One entry point for turning YouTube into artifacts. All data access goes through
the uvx yt-mem-ai <cmd> CLI (see [[yt]] for the full command
surface) — never touch the LanceDB store directly. Always invoke it exactly that
way: zero-install and cached, so nothing has to be on PATH. Do not go looking
for a yt-ai binary — the native plugins install no package. (Only inside a
source checkout may you use uv run yt-ai <cmd>.) Single-quote every URL —
YouTube URLs contain ?/&, which the shell globs on, so a bare URL fails:
uvx yt-mem-ai fetch 'https://www.youtube.com/watch?v=ID' --captions-only.
Everything is grounded in the transcript; highlight timestamps come from
uvx yt-mem-ai search, never invented. The analysis is done by this agent — no API
key, no OpenRouter.
Pick the scenario
- A — one video (a URL, a bare 11-char
video_id, or "this video" as a follow-up): produce asummarize/highlights/qa/presentationartifact. - B — process latest subscriptions ("catch up", "daily", "new uploads"): discover + fetch + analyze each + write a dated digest.
- C — subscriptions review ("review my subs", "themes lately", "what's been happening"): one cross-video synthesis over a date range.
- D — group (arbitrary set) ("process/review these videos <ids/urls>", "review channel ''", "review from to "): ingest a user-specified set, then per-video analysis + a group synthesis.
Core: analyze one video (used by A and B)
Given a video_id (and a URL if it may not be ingested yet):
- Ensure ingested (idempotent):
uvx yt-mem-ai show <video_id> --json.not foundand you have a URL →uvx yt-mem-ai fetch '<url>' --captions-only. If that printsno captions available: ..., fall back touvx yt-mem-ai fetch '<url>' --whisper(downloads audio + transcribes — slower, always yields a transcript).- Blocked instead?
Sign in to confirm you're not a bot→ runuvx yt-mem-ai config set YT_COOKIES_BROWSER chromeand retry;captions blocked ... IP rate-limited→ cookies won't help, see [[yt]]'s When YouTube blocks a fetch. - Already ingested → instant, no download (
is_seenskips it).
- Reuse if present: if the
show --jsonoutput has a non-nullsummary, reuse it — skip generation unless the user asked for a fresh artifact. - Anchor highlights: for each candidate highlight phrase, run
uvx yt-mem-ai search "<phrase>" --vector -k 3and use theMM:SSfrom a returned line whosevideo_idmatches. Never invent timestamps. - Produce (you, the model — no API):
summary_md(2–4 sentence exec summary- key bullets),
highlightsJSON[{"start_s": <seconds>, "label": "..."}](3–8, seconds from step 3),qaJSON[{"q": "...", "a": "..."}](3–6).
- key bullets),
- Persist:
uvx yt-mem-ai save-summary <video_id> "<summary_md>" --highlights '<json>' --qa '<json>'.
A — single-video artifacts
Run the core, then deliver the artifact the phrasing asked for:
- summarize (default) → executive summary + key bullets in chat (mention highlights / qa / presentation are available on request).
- highlights → each as
MM:SS — label, a deep linkhttps://www.youtube.com/watch?v=<id>&t=<start>s. - qa → the Q&A pairs.
- presentation → write a slide deck to
slides/<video_id>.md:----separated slides (renderable by reveal.js / Marp; no images).- Slide 1: title + channel + a one-line thesis.
- One slide per theme/section: a heading, 3–5 key-point bullets, and any
notable quote with its
MM:SStimestamp. - Final slide: takeaways + the watch link. Report the file path + the title slide in chat.
B — process latest subscriptions (daily digest)
uvx yt-mem-ai discover # new uploads → 'discovered'
uvx yt-mem-ai fetch-pending # download + transcribe + embed today's batch (skips failures)
Live streams are auto-detected and marked status=stream — fetch-pending skips
them (long + usually caption-less). List them with uvx yt-mem-ai list --status stream;
to transcribe one on demand, fetch it directly (uvx yt-mem-ai fetch '<url>', optionally
--whisper).
Then for each of the day's transcribed videos
(uvx yt-mem-ai list --status transcribed --since <DATE> --json), run the core
analysis. Compose digests/<DATE>.md:
- Top executive digest: cross-video themes and what's worth the user's time.
- One section per video:
## <title>+ link, the 2–4 sentence summary, top highlights (MM:SS — label), 2–3 Q&A.
Create digests/ if needed. Report the digest path + the executive digest.
Idempotent — re-running overwrites each summaries row and rewrites the file.
C — subscriptions review (cross-video themes)
Select the period's videos: uvx yt-mem-ai list --status summarized --since <DATE> --json
(analyze any still transcribed via the core first). Then write one essay to
reviews/<DATE>.md — no per-video sections:
- Common threads across the videos, contradictions / disagreements, emerging trends.
- Ground every claim in the videos; when citing a specific moment, name the video
title + a
MM:SSdeep link.
Report the review path + a short lede in chat.
D — group of videos (arbitrary set)
Process a user-specified set (not tied to today's subscriptions), then per-video analysis + a top-level synthesis.
- Resolve the set → ids/URLs:
- comma list (
id1,id2,https://youtu.be/id3) → parse directly; - channel (URL/@handle) →
uvx yt-mem-ai channel-list '<url>' --limit N [--from D] [--to D] --json; - date range over a channel → same with
--from/--to. Report the resolved count first; if it's large (> ~15), say so and confirm/cap before mass-ingesting (whisper is slow).
- comma list (
- Ingest each:
uvx yt-mem-ai fetch '<url>'(captions→whisper; streams auto-markedstatus=streamand skipped; continue past failures — note any skipped). - Per-video: run the core analysis (summary + search-anchored highlights +
Q&A,
presentation→slides/<id>.mdif asked), persisted viasave-summary, in each video's original language (FTS-anchor non-English). - Group synthesis →
groups/<label>.md(label = channel handle / date-range slug / timestamp): an executive synthesis (themes, standouts, what's worth watching) + one section per video (## <title>+ link, summary, top highlights asMM:SS — label, 2–3 Q&A). - Report the
groups/<label>.mdpath + the executive synthesis.
This is the daily-digest shape (B) over an arbitrary set. Use C instead for a themes-only essay with no per-video sections.
Conventions
- All data via the
uvx yt-mem-aiCLI ([[yt]] has the full surface). Never touch the store directly. - Grounded strictly in transcripts; highlight timestamps only from
uvx yt-mem-ai search. - Language / translation: transcripts may be in ANY language — the source
language is stored and returned by
show --jsonastranscript_lang. Default: produce each artifact in the video's OWN original language (Russian video → Russian summary) — no translation. Only translate when the user asks for a specific target language. If the user hasn't stated a preference and the batch mixes languages, ask once which output language they want, then apply it to the whole run (and treat that as their saved default). Anchoring caveat: vector search is language-sensitive — search with a phrase in the transcript's original language, then write the label in the chosen output language. (For heavily multilingual libraries, a multilingualYT_EMBEDDING_MODELimproves search — switch it from chat withuvx yt-mem-ai config set YT_EMBEDDING_MODEL paraphrase-multilingual-MiniLM-L12-v2, thenuvx yt-mem-ai reembedto migrate the library.uvx yt-mem-ai config set/listreconfigures any.envsetting — Webshare creds, cookies browser, backend — see [[yt]].) is_seenis status-based, so re-fetch is a no-op → follow-ups ("now highlight it", "make slides") and re-runs never re-download.- Dates are
YYYY-MM-DD. Always report what ran + the output paths (slides/,digests/,reviews/). - Related: [[yt]] (the full CLI surface — discover, recommend, compile, supercut, frame, status).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |