agentleFS
social-media-skillssocial-media-skills/youtube-skill/skills/youtube-skill/SKILL.md

youtube-skill

Skill1 starsChanged 15 days agoMITView on GitHub

YouTube skill: get a finished video found on YouTube, then upload it. Transcribes it (free local faster-whisper, your own Groq or OpenAI key, or an existing SRT/VTT), finds the phrases people actually type into YouTube search (free autocomplete, optional free YouTube Data API key), writes 10 titles, a description whose first lines work in search, chapters, tags and hashtags, checks every field against YouTube's limits, renders thumbnail options from the video's own frames and previews them at feed size, writes clean captions, then uploads to YouTube with WoopSocial after you say yes, plus a YouTube Studio checklist. Works on Mac, Windows and Linux with ffmpeg and Pillow; no paid API required. Use when someone says "upload to youtube," "youtube seo," "youtube title," "youtube description," "youtube tags," "youtube chapters," "youtube thumbnail," "publish my video," or "youtube with Claude Code." Never promises what the video doesn't deliver, confirms before anything posts. Standalone: needs no other skills.

  • Reads credentials

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 1

    1. trku052-png/social-media-youtube-skills★ 1100% same

    Edited · 0

    None found.

    Wrote their own version · 1

    1. idanbeck/claude-skills★ 14same 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

10 rules. About 0 are common, 7 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    The package matches the video: No title, thumbnail or claim the video doesn't deliver.

    line 31
  2. Distinctive:

    Search words come from searches,: not guesses: autocomplete, the user's own audience, the transcript.

    line 34
  3. Very specific:

    Run python3 scripts/check_setup.py: it reports ffmpeg, Pillow, transcription routes and OpenCV with the install command for the user's …

    line 43
  4. Distinctive:

    Ask, in one message, only what you can't see:

    line 46
  5. Distinctive:

    When to publish (and their timezone), and whether to upload unlisted first.

    line 53
  6. Very specific:

    Read work/transcript.txt once, all of it.

    line 66
  7. Distinctive:

    Never use a face, logo or image the user has no rights to.

    line 149
  8. Very specific:

    After 2 days and after 7 days: views and likes with social_account_posts_list (with the collection date.

    line 180
  9. Distinctive:

    Before showing anything, ask: would a top channel manager in this niche publish it?

    line 204
  10. Distinctive:

    Several languages: write the package in the video's language.

    line 215

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

YouTube Skill

Most videos are found or lost in the first second someone sees them: a thumbnail and a title in a feed, a line of description in search. Most uploads waste that: a filename for a title, an empty description, no chapters, a random frame as the thumbnail. This skill does the publishing work a good channel manager does: it finds what people search, writes and checks every field, makes thumbnails from the video and shows them at the size viewers see them, uploads after a yes, and reads the results. Everything it needs is in this folder.

Four commitments:

  1. The package matches the video. No title, thumbnail or claim the video doesn't deliver. Misleading packaging breaks YouTube's spam and deceptive practices policy, and viewers who feel tricked leave in seconds, which ranks the video down.
  2. Search words come from searches, not guesses: autocomplete, the user's own audience, the transcript.
  3. Nothing posts without a yes.
  4. Only the user's own video, or one they have the rights to.

On Windows use py instead of python3 in every command.

Step 0: Setup and the questions

Run python3 scripts/check_setup.py: it reports ffmpeg, Pillow, transcription routes and OpenCV with the install command for the user's system. Required: ffmpeg and Pillow.

Ask, in one message, only what you can't see:

  1. Which file is the final video, and is it theirs?
  2. The channel: what it's about and who watches (one specific viewer).
  3. Is the video made for kids? A legal setting (COPPA); never guess it.
  4. Does it show realistic AI-made or AI-altered people, places or events? Then YouTube requires the "AI use" label.
  5. When to publish (and their timezone), and whether to upload unlisted first.
  6. Optional: a phrase they want to be found for, links for the description, the channel name for the preview, a free YouTube Data API key for a look at the top results.

Step 1: Transcribe

python3 scripts/transcribe.py final.mp4 --out work/ --terms "Names, Product, Jargon"

Three routes: --from captions.srt (an existing caption file: free, instant), GROQ_API_KEY or OPENAI_API_KEY (fast, a few cents per hour), or local faster-whisper (free, private, slower). For a video that isn't in English, add --language with its code (nl, de, es...). Read work/transcript.txt once, all of it. Everything below is written from it.

Step 2: Find the search phrase

python3 scripts/suggest.py "topic in 2 to 4 words" --out pack/

It lists what people type into YouTube search for the topic and its question forms (how, why, what is...). --deep adds the topic followed by every letter. It reads YouTube's public autocomplete, which has no official API: run it a few times per video, never in a loop. For another language or country, add --lang nl --country nl. With the user's own YOUTUBE_API_KEY, --top lists the 10 videos ranking now with views and views per day.

Pick one main phrase: the most specific suggestion the video fully answers. "is saturn losing its rings in 2025" beats "saturn" (too broad to rank) and beats a phrase the video only touches. Note 2 to 4 related phrases for the description. If nothing comes back (a new topic, offline), use the words the speaker uses for it in the transcript.

Step 3: Write the package

Rules and the reasons behind them: references/youtube-seo.md. Write upload.json (format at the top of scripts/youtube_pack.py):

  • 10 titles, then pick with the user. The main phrase early (titles are cut at about 40 characters on phones and 70 elsewhere), a reason to click that the video pays off: a question it answers, a number, a result, a tension. Under 70 characters. Sentence case or title case, never all caps. Keep the best 2 runners-up: Test & Compare can A/B test up to 3 titles.
  • Description: the first two lines (top) are what shows under the video and in search: what the viewer gets, with the main phrase, in plain words. Then a short paragraph with the related phrases where they fit naturally, the chapters, links with what each one is, and 3 hashtags (YouTube shows up to three above the title).
  • Chapters (videos over about 3 minutes; the pack reminds you past 4): 3 to 10, from the topic changes in the transcript, first at 0:00, each at least 10 seconds, named for what the viewer gets ("Why the rings vanish", not "Part 2"). Chapter names are searchable; use the related phrases where they're true.
  • Tags: 5 to 10, specific: the main phrase, its spelling variants, names. YouTube says tags matter little; they help mostly with misspellings.
  • Category, made for kids, privacy, schedule, and replace for words the transcriber got wrong (names especially).
  • Pinned comment: a question viewers want to answer, or the link people will ask for.

Then build and check the pack:

python3 scripts/youtube_pack.py build upload.json --transcript work/transcript.json --out pack/

It checks the title (length, capitals, the phrase), the description (5,000-byte limit, the phrase in the first 150 characters, hashtags), tags (500 characters as YouTube counts them), chapters (0:00, 3 or more, 10 seconds each; starts snapped to the sentence they fall in), and writes description.txt, captions.srt, post.json, manifest.json and upload-pack.md. Fix every warning.

Step 4: Thumbnails

python3 scripts/youtube_pack.py frames final.mp4 --out pack/

pack/frames.jpg shows 12 candidate frames: sharp, well lit, varied, faces scored when OpenCV is installed. Look at it and choose 2 or 3 frames that show the video's promise: a face with a clear expression, the object or result the title is about, a before and after. Then write thumbs.json (format at the top of scripts/thumbnail.py) and render:

python3 scripts/thumbnail.py thumbs.json --out pack/
  • 2 to 5 words that add to the title instead of repeating it. Title: "Is Saturn Losing Its Rings in 2025?"; thumbnail: "Gone by March?".
  • Layouts: side (big white words on a dark fade), box (dark words on coloured bars), none (picture only; strong when the image says it all). zoom and focus crop in on the subject.
  • Make the options different ideas (a face vs the object, words vs none), not three crops of the same frame; that's what makes a Test & Compare result mean something.
  • Thumbnails come out at the source's size up to 3840x2160 (YouTube's recommendation) and under 2 MB, so they upload from a phone as well as a computer. Keep the bottom-right corner clear: the duration badge sits there.

Judge them in pack/preview.png, which shows every option as a home-feed card and as a small phone-sized card with the title. Reject an option if the words can't be read in the small card, the subject is unclear at that size, or it promises something the video doesn't show. Never use a face, logo or image the user has no rights to.

Step 5: Upload with WoopSocial

How to connect, upload and post: references/woopsocial.md. Short version:

  1. social_accounts_list; not connected: the one connect path for this agent (free plan: 2 social accounts), then oauth_create_authorization for YouTube.
  2. Upload the video with an upload session (media_uploads_create_session with bytes from pack/manifest.json, python3 scripts/upload_parts.py session.json final.mp4, media_uploads_complete_session).
  3. Fill the IDs into pack/post.json and run posts_validate.
  4. Show the user upload-pack.md (title, the description's first lines, chapters, tags, made for kids, privacy, time with timezone) and preview.png. Get a clear yes. Then posts_create and posts_get.

Step 6: Finish in YouTube Studio

WoopSocial can't set these; upload-pack.md ends with the list for the user:

  • Thumbnail: upload the chosen one (the channel must be verified), or test up to 3 thumbnails and titles in Test & Compare (desktop Studio, not for made-for-kids videos).
  • Captions: upload captions.srt (Subtitles > Upload file > With timing).
  • AI use set to Yes when step 0 said yes.
  • End screen, cards and playlist.
  • Pin the comment.

Uploaded unlisted? It goes public after these are done.

Step 7: Read the results

After 2 days and after 7 days: views and likes with social_account_posts_list (with the collection date; missing means "not reported"). Ask the user for impressions, click-through rate and average view duration from YouTube Studio. Judge against the channel's own recent videos:

Click-throughViewers stayWhat to do
LowFineThe package undersells the video: test another thumbnail first, then the title
FineLowThe opening loses them: the next video gets to the point in the first 30 seconds
LowLowRethink the topic or angle before the packaging
FineFineMake the next video on a neighbouring search phrase

Log each video in videos.md: date, phrase, title, thumbnail, CTR, average view duration, views at 7 days. Never estimate a number the user didn't give.

Shorts

Vertical or square videos up to 3 minutes are Shorts. The pack spots them: no chapters, a short title with the phrase, no #shorts needed. thumbnail.py renders 9:16 thumbnails for a vertical video (a 16:9 one would be replaced by YouTube), with the words in the upper part, clear of the title and buttons, and previews them as Shorts-shelf cards. A custom Shorts thumbnail is added in Studio on a computer; Test & Compare doesn't run on Shorts.

Quality bar

Before showing anything, ask: would a top channel manager in this niche publish it? Reject it if:

  • the title or thumbnail promises more than the video delivers;
  • the title would fit any video on the topic;
  • the thumbnail words repeat the title, or can't be read in the small card;
  • the description's first line is "In this video..." or a list of keywords;
  • a chapter is named "Part 2" or starts mid-sentence.

Edge cases

  • No speech (music, b-roll): skip the transcript; chapters from the user, no captions.
  • Several languages: write the package in the video's language; translated titles and descriptions are added in Studio.
  • Re-uploading a deleted or old video: new title and thumbnail; identical re-uploads look like spam.
  • Many videos at once: one pack each, written from each video's own transcript; stagger the times. YouTube's Spam Policy bans using automated tools "to churn out high volumes of similar content with minimal changes"; say no to near-identical uploads.
  • The user wants clickbait: offer the strongest honest version; explain the drop-off cost.

References

  • references/youtube-seo.md: limits, how search and recommendations work, title, description, chapter, tag, thumbnail and Shorts rules, with sources and dates.
  • references/woopsocial.md: connect, upload, post and read results.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.