Communitygithub.com

shorts

Makes YouTube Shorts in two modes: Clip (cut from a finished episode: candidates from the transcript and retention, a 9:16 rough cut from the original recording around the face, portrait graphics, music with clear rights, the Critic's own Short check) and Original (a Short recorded for itself: three ideas with a proof from the user's real work, the user talks freely, the Writer shortens, only the speaker is recorded, the crew brings the B-roll). Publishing: one card per Short, or the user's own standing approval for Shorts; at most two a day. Use when: \"make shorts\", \"make a short about …\", \"clip this episode\", \"short from episode N\", \"publish shorts without asking\", \"how are my shorts doing\". Not for long videos (skills/episode, skills/edit).

What is shorts?

shorts is a Claude Code agent skill that makes YouTube Shorts in two modes: Clip (cut from a finished episode: candidates from the transcript and retention, a 9:16 rough cut from the original recording around the face, portrait graphics, music with clear rights, the Critic's own Short check) and Original (a Short recorded for itself: three ideas with a proof from the user's real work, the user talks freely, the Writer shortens, only the speaker is recorded, the crew brings the B-roll). Publishing: one card per Short, or the user's own standing approval for Shorts; at most two a day. Use when: \"make shorts\", \"make a short about …\", \"clip this episode\", \"short from episode N\", \"publish shorts without asking\", \"how are my shorts doing\". Not for long videos (skills/episode, skills/edit).

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/marcopatzelt/crewcut/tree/HEAD/skills/shorts

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

What does shorts do?

A Short is its own small episode in portrait (1080x1920) with fewer steps. Clip comes from a finished episode and costs the user nothing but one yes. Original is recorded for itself, to test quickly what lands. Data suggest, the user decides: a strong idea that makes no sense to the user right now is dropped without discussion.

Which mode (Showrunner)

SituationDo
An episode is cutThe Scout looks for clip candidates on its own; the card comes with or right after the preview
An episode has been online for two days and retention data existThe Scout checks for a strong passage that is no Short yet; one card if so
"make a short about …"Original
"make shorts", nothing moreOne card: up to three clip candidates of the last episode and three original ideas
No video for a long time, none in workOriginal ideas, at most once a week, unless the user switched that off
An original idea the user already said in an episodeSay so: "You already said this in episode 12. A clip takes no recording."

Clip first when material exists. New Short: crewcut episode new "<title>" --short clip --from <episode> or --short original. The step strip then shows only its steps (clip 6, original 9).

Clip

  1. Candidates (Scout). crewcut cut candidates --episode <source> lists stand-alone passages with times. A candidate needs: a first sentence that stands without the video before it and holds a claim or a result; a proof in the picture (a number, a comparison, a result), not only a talking head; a real end, never a sentence cut off in the middle; 20 to 55 s from at most about twelve whole sentences (they may come from different places). With retention data, prefer passages where the curve rises. Without data, say that you worked from the text only. Check that the originals and a clean voice exist before you propose a candidate.
  2. One card, the best three: title, first sentence, length, a still each, strength 1 to 10 and effort (small, medium, large: is the footage there in portrait, or must screens be recorded again?). Button "Make these three".
  3. Rough cut (Editor). crewcut cut short with the pieces by start and end word (or fixed seconds when word times are off): back to the original recording, 9:16 around the face, the episode's clean voice, new word times. A piece never crosses a cut boundary. If the original is too small for a sharp portrait crop (1080p landscape), say so before building, not after.
  4. Portrait graphics. 2D scenes rendered again in portrait with the style pack. Screens and web pages recorded again in portrait (crewcut render capture), never the landscape picture cropped. 3D of the episode comes back as a card or a layer; new portrait 3D costs rent and needs a card. At most shorts.text_cards_max text cards; every picture understood within two seconds; symbols before text; calm motion (nothing that grows pixel by pixel).
  5. Music. Only music with clear rights and a source file (crewcut render music-bed): under the voice at shorts.music_db, ducking, a short hole before the punchline. Never the long video's music unless its licence says so.
  6. Check (Critic). crewcut cut short-check: the Short rule set (face in frame, no text over the face or the right edge where YouTube's buttons are, brightness, length and pace from the style pack, loudness with the limiter headroom, piece boundaries at the audio level). Look at the contact sheet; write the "please listen" list with time stamps for the user: nobody but the user hears the sound. 0 errors before the preview card.

Original

  1. Ideas (Scout). Sources: Shorts of the channels the user named, by views against the channel's median; what the user really did (recent episodes, tests, what they told in the chat). Never invent an experience. Forms that carry (Nate Herk's analysis of 50 Shorts, a US channel with over a million subscribers; a direction, not a promise): the Short is the proof ("Claude built this"), first choice; first-person story with a wow, second; someone else's story, possible; tool presentation and news, avoid; street polls, avoid. One card with three ideas: first sentence, the proof that will be in the picture, the form, the evidence from the data. No proof, no idea.
  2. The user talks first, then the Writer shortens. Ask in one message for two to three minutes of free talk (the microphone in the composer), along four questions: What did you do? What did you think before? What went wrong or surprised you? Which number do you have? Keep the user's words and order them: I, event, number (the first sentence holds the whole conflict); a time anchor or a number in a series; what I thought; the turn, with the proof in the picture (a demo runs within five seconds); what I do differently now; a loose end. Length from shorts.words (default 110 to 170 words, 30 to 50 s).
  3. Check and humanize. python3 skills/script/scripts/script_check.py <file> --short until 0 errors, then the Humanizer pass (skills/script). The script card's button is "This really happened": the user confirms the story is true (--data '{"button": "This really happened"}').
  4. Record: only the speaker, portrait or landscape with room to crop, from the teleprompter or from key points. Nothing else: no screen, no hands, no second angle.
  5. B-roll (Editor): the real proof from the user's work (a portrait screen recording, a render, a number as a scene), 2D scenes in the style pack, 3D only if the pack has it (with the usual estimate). Nothing staged, no invented quote. Layout from shorts.layout: "full", or "proof-top" (proof on top, face below, the end on the face).
  6. Check as in Clip step 6.

Title, thumbnail, link to the long video

  • Title with the usual title checks, shorter. The first sentence and the title say the same.
  • Thumbnail (Designer): portrait with a free lower edge.
  • Link: the package gets related_url (the long video) for clips; the upload adds it to the description and as a comment. "Related video" and pinning have no API: tell the user in one line to set them in YouTube Studio.

Publishing

  • One publish card per Short, unless the user switched on the standing approval for Shorts: the quiet second button on a Short's publish card, or crewcut shorts standing on (as a crew member this proposes the card; only the user's yes switches it on). It covers only Shorts, only after short-check found 0 errors, never with someone else's material or a person ("third_party": true in the package), and not the first Short after the style pack changed. Every Short goes up private, is read back, and only then becomes public or scheduled. Off: "ask me again for Shorts" → crewcut shorts standing off. With the standing approval the API is the right way: no hand on the file.
  • When: crewcut shorts slot --episode <short> gives the next free time (at most shorts.per_day_max a day at shorts.publish_times, never two Shorts with the same core message on one day, pairs one after the other). "All at once" is one sentence from the user.

Results

After 48 hours: crewcut shorts results --episode <short> adds a row to studio/shorts.tsv (mode, form, length, captions, first sentence, views, average view percentage, subscribers). From ten Shorts on, the Scout may propose a rule from the table as a card ("Your Shorts under 40 s keep 20 % more viewers. Make 40 s the limit?"). Before ten, say nothing about patterns.

Honest limits

  • Built and unit-tested; no Short has been made end to end with Crewcut yet.
  • A fixed crop needs the speaker near the middle; a crop that follows a moving speaker is planned.
  • Retention data need the YouTube Analytics API (connected with YouTube).

Sources

  • Nate Herk (https://www.youtube.com/@nateherk): Shorts as proof; the form ranking above is our summary of an analysis of his 50 Shorts.
  • Reel storytelling of Farza, Chris Raroque, Roy Lee, Riley Brown, Varun Mayya and Codie Sanchez (links in README): first person, the turn, uneven real numbers; the user talks, the agent shortens.

Individual skills in this repo

This repo contains 3 individual skills — each has its own dedicated page.

edit

Builds a video version from the user's recording: the speaker cut, then one edit plan with all inserts (2D scenes, 3D scenes, recordings, keywords), sound and music, built with one command (crewcut render). Use for: edit the video, build the cut, speaker cut, rough cut, insert, b-roll, 2D scene, 3D insert, counter, keyword, music bed, sound, assemble, render the preview, new plan version, short from the long video.

script

Writes the spoken script of an episode in the user's own voice (voice.md), in teleprompter form, after title and proof are fixed: opening that delivers the title's promise, small stories with rising stakes, the critical side, a verdict in its own form, the channel's closing line. skills/script/scripts/script_check.py checks form, sentences, opening, ending, numbers against the proof, and length. Delivered through the web teleprompter (crewcut prompter). Use when: \"write the script\", \"teleprompter text\", \"rewrite the script\", \"shorten the script\", \"make it sound natural\", \"script for the episode\". Not for titles or thumbnails. Short scripts: the `--short` mode of the check and skills/shorts.

title

Writes and checks the YouTube title and the thumbnail text of an episode, before the script. Delivers five candidates with a hook and a reason each, one recommendation, and three variants for the A/B test; skills/title/scripts/title_check.py checks every candidate against the style pack. Use when: \"title\", \"video title\", \"YouTube title\", \"what should the video be called\", \"check this title\", \"shorten the title\", \"better title\", \"thumbnail text\". Not for the blog's search title and not for the thumbnail image.

Related Skills