Communitygithub.com

SrigadaAkshayKumar/ai-video-editor

Write the upload package for a video this repo edited — YouTube title options, description with chapters, tags, pinned comment, thumbnail text, and the Instagram Reels caption + hashtags for the 9:16 cut — optimised for whichever channel/client the user says it is for (config/brands/ profiles, only when named). Use when the user asks what to title a video, for a description, tags, chapters, "the YouTube metadata", an Instagram caption, or SEO/CTR advice on a finished render.

Qu'est-ce que ai-video-editor ?

ai-video-editor is a Claude Code agent skill that write the upload package for a video this repo edited — YouTube title options, description with chapters, tags, pinned comment, thumbnail text, and the Instagram Reels caption + hashtags for the 9:16 cut — optimised for whichever channel/client the user says it is for (config/brands/ profiles, only when named). Use when the user asks what to title a video, for a description, tags, chapters, "the YouTube metadata", an Instagram caption, or SEO/CTR advice on a finished render.

Compatible avec~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/SrigadaAkshayKumar/ai-video-editor/tree/HEAD/.claude/skills/youtube-metadata

Demander à votre IA préférée

Ouvre une nouvelle conversation avec cette compétence d'agent déjà préchargée.

Documentation

Upload package (YouTube + Instagram)

You have the whole video on disk — don't ask what it is about. Read first:

FileGives you
work/cleancut.transcript.mdevery claim, in the speaker's words, on the final timeline
work/cut-report.mdthe clean script; work/edl.json reasons summarise structure
work/visual-plan.jsonchapter_open overlays = section titles + FINAL timestamps → YouTube chapters; big_statements = the punchlines
work/direction.json (faceless)logline, promise, payoff
output/credits.mdsources + required attributions — they go in the description
project.json brief / brandthe user's intent for this video; the brand slug if they named one
config/brands/<slug>.jsonthat brand's name, handle, channel id, audience, title formula, examples, links (only if a brand is named)

Never invent a claim, number or promise the video doesn't deliver.

1. Real channel data (tell the user which tier you got)

Whose video is it? This repo edits for many people. If the user hasn't said, write channel-neutral metadata (no channel name in titles, generic links) and ask in one line whether they want it tuned to a channel. When they name one, use/create config/brands/<slug>.json and do the tiers below.

  • Tier 1 — YouTube Data API (YOUTUBE_API_KEY in .env, channel id in the brand profile): search?channelId=…&order=viewCount&type=video&maxResults=25 and order=date, then videos?part=snippet,statistics,contentDetails&id=…. The gap between what performs all-time and what performs recently is the most useful signal there is.
  • Tier 2 — the browser (ask first): open the channel's Videos tab, sort Popular, read titles + views; then Latest. Public, no login.
  • Tier 3 — WebSearch: conventions only, no view counts. Say the advice is pattern-matched.
  • For CTR-level advice ask the user for YouTube Studio numbers (impressions, CTR, AVD).

If the brand profile is missing fields, infer the conventions from the channel's recent titles and offer to save them to config/brands/<slug>.json (name, handle, title formula + examples, audience, links).

2. Conventions

  • Match the channel's own title formula (a title that looks foreign to the channel loses the returning-subscriber click). For Telugu/Hindi channels the usual pattern is an English title with the language as a keyword: <hook> | Telugu | <Channel> or <topic> in Telugu.
  • Check for an existing video on the same topic: differentiate on the question this one answers, and link the older one.

3. Produce one copy-pasteable block

Titles — 4-6, ranked, each < 60 characters, keyword in the first 3-4 words, a spread of angles: the direct question, the myth-buster/negative framing, the real number or outcome the video contains, the "who is this for". One line of click logic each. No emoji.

Description: first two lines (all most people see) restate promise + payoff with the main keyword in line 1, keyword-dense but readable in the first 150 characters; a short paragraph in the audience's own register (Telugu/Hindi-English mix is fine); chapters from the chapter_open overlays, first one 0:00, format M:SS Title (add the intro length if an intro is appended); links (anything the video mentions, the related older video, socials from the brand profile if any); the credits block from output/credits.md; 3-5 hashtags last.

Tags 15-25, comma separated: exact English search phrases, the same intent transliterated ("nptel telugu"), channel tags, broad topic tags last; highest intent first.

Thumbnail text: 2-3 options, 3-5 words, completing the title rather than repeating it.

Pinned comment: one line inviting the specific comment this video should generate.

Instagram (9:16 cut): a caption whose first line is the hook (it truncates after ~125 characters), 2-4 short lines, a call to comment/save, the credits if required, 5-10 hashtags mixing topic + language + niche. Suggest a cover frame (see youtube-thumbnail: output/cover-9x16.png).

Rules

  • Never promise what the video doesn't deliver; the honest counter-intuitive angle is usually the stronger hook anyway. No clickbait the first 30 seconds doesn't pay off.
  • Credits/sources required by output/credits.md are not optional.

Individual skills in this repo

This repo contains 9 individual skills — each has its own dedicated page.

SrigadaAkshayKumar/ai-video-editor

Turn a finished AI-edits video (projects/<p>, or several parts of one long video) into YouTube Shorts. Reuses the edit's word timestamps, reads the whole transcript, picks self-contained dialog moments, adds text hooks + comment prompts, renders 9:16 clips with Remotion. Use when the user asks to clip/make shorts from an edited project or a long video.

SrigadaAkshayKumar/ai-video-editor

Choose, fetch and place background music for an edit — Openverse search (commercial-use CC, or CC0-only), music_cues in work/audio-plan.json with fades, offsets and a deliberate silence gap; tools/mix.mjs conditions every bed, sets it to one measured house level under the voice and sidechain-ducks it. Use when adding, swapping or re-levelling music.

SrigadaAkshayKumar/ai-video-editor

Generate and style word-synced captions (Telugu, Hindi, English, code-mixed) as the HyperFrames captions pass over each format's Remotion picture, fix caption text, black captions out under on-screen text, and check caption content at every cut join. Use when creating, restyling, correcting or repositioning captions in any edit.

SrigadaAkshayKumar/ai-video-editor

Editorial rules for turning a raw talking video or voiceover into a clean cut — removing retakes, repetitions, false starts, stutters, fillers, mispronounced words, contradictions/inconsistencies, off-script chatter and dead air — by writing work/edl.json for tools/cut.mjs. Telugu, Hindi, English and code-mixed speech. Use for stage 2 of edit-video or whenever the user asks to tighten/re-cut a video.

SrigadaAkshayKumar/ai-video-editor

Turns a bare voiceover (no presenter on camera) into a finished faceless documentary/explainer for YouTube 16:9 and Instagram 9:16 — clean-cuts the narration, DIRECTS it (story spine, tension curve, hooks, involvement beats), researches the claims and screenshots the real sources, plans a scene track (stock b-roll, stills, news screenshots, pure-graphic beats) with motion graphics, designs music + SFX, renders, captions and credits it. Use when the user gives a voiceover/narration file (or says "faceless video", "make a video from this VO", "documentary").

SrigadaAkshayKumar/ai-video-editor

End-to-end editor for TALKING-HEAD videos (a person on camera) in this repo. Use when the user hands over a raw video with a speaker and wants a finished edit, or asks to continue/resume/redo any stage of an existing talking-head project — clean cut, motion graphics, b-roll, SFX, music, captions, final render, thumbnail. Delivers YouTube 16:9 and Instagram 9:16. Telugu, Hindi, English and code-mixed speech. For a voiceover with no presenter use edit-video-faceless.

SrigadaAkshayKumar/ai-video-editor

Plan the picture of an edit in work/visual-plan.json — motion graphics (28 component types across six stage layouts that dock/shrink/dim the video), b-roll cutaways from Pexels/Pixabay, camera moves, a-roll framing for 9:16 — validate it, QA it with stills in both formats, and render it with Remotion. Use for any titles, charts, stats, lower thirds, diagrams, b-roll or zooms in a talking-head or faceless edit.

SrigadaAkshayKumar/ai-video-editor

Design the sound effects of an edit in work/audio-plan.json sfx_events — which effect, where, how often — using the measured SFX library (repo assets/sfx, HyperFrames' bundled set, CC0 fetches from Openverse) and the spectral rule that sub-band cues carry structure while vocal-band cues are rationed. Mixed and levelled by tools/mix.mjs. Use when adding or adjusting SFX in any edit.

SrigadaAkshayKumar/ai-video-editor

Design and build a high-CTR thumbnail for a video this repo edited — hook selection, the real face frame pulled from the footage, background cut-out, type set with real fonts (Telugu/Devanagari included), the YouTube 1280x720 thumbnail and the Instagram 1080x1920 cover, an optional image-generation prompt for the background layer, and a feed-size legibility check. Use when the user asks for a thumbnail, thumbnail ideas, a cover image, or how to make the thumbnail more clickable.

Skills associés