agentleFS
ekusiadadusekusiadadus/video-edit-harness/.agents/skills/youtube/SKILL.md

youtube

Skill0 starsChanged 41 days agoMITView on GitHub

Edit a local video for a regular YouTube upload with coherent long-form structure, natural pacing, scene-appropriate color, clear audio, captions and Final Cut Pro handoff. Use for /youtube or regular YouTube editing requests; route vertical Shorts to tiktok.

37 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use douyin, linkedin and 2 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • ballbasecn-starSkill

    douyin

    抖音视频解析、博主视频抓取/转录/运营分析/写入飞书。 触发词:解析抖音、抖音视频、抓取视频、博主视频、创作者视频、douyin、抖音解析、飞书多维表格 子命令: douyin:fetch - 抓取博主最新视频 → 本地转录 → AI运营分析 → 直写飞书多维表格 douyin:parse - 解析单个抖音短视频(链接/分享文本 → 元数据 + 转录)

    Also in 3 projects that use this

  • kkoppenhaverSkill

    linkedin

    Write LinkedIn posts in Keanan's voice. Use when drafting LinkedIn content, creating professional posts, or sharing builder-focused insights. Triggers include requests to "write a LinkedIn post", "draft something for LinkedIn", "help me post on LinkedIn", or any LinkedIn content creation. Generates upbeat, builder-centric content mixing practical how-tos with candid humor.

    Also in 3 projects that use this

  • OpenAISkill

    email

    Email sending integration guidance — Resend (native Vercel Marketplace) with React Email templates. Covers API setup, transactional emails, domain verification, and template patterns. Use when sending emails from a Vercel-deployed application.

    Also in 3 projects that use this

  • SentrySkill

    skill-creator

    Create new agent skills following the Agent Skills specification. Use when asked to "create a skill", "add a new skill", "write a skill", "make a skill", "build a skill", or scaffold a new skill with SKILL.md. Guides through requirements, writing, registration, and verification.

    Also in 4 projects that use this

The file

Files
SKILL.md

YouTube editing

In Claude Code, /youtube /path/to/video.mov [brief] supplies the local video and optional brief in $ARGUMENTS. Codex uses $youtube with the path and brief. The marketplace plugin exposes /video-editing:youtube; the standalone skill installed under ~/.claude/skills/youtube exposes the exact /youtube spelling. This is an editing invocation, not a shell command or permission to post.

Locate the Video Edit Harness checkout (v0.1.0-alpha.7) through VIDEO_EDIT_HARNESS_ROOT or the current project. If absent, follow the installation guide. Read its README.md, AGENTS.md and .agents/skills/video-editing/SKILL.md for the actual session commands. Run commands from that checkout. The standalone ZIP contains instructions, metadata and license; the execution harness is separate.

Run uv run video-harness doctor before editing. Check root, versions, skills, commands and required tools; resolve relevant findings. Persist source upload allow/deny through session cloud-policy, using the actual current authorization and provider scope. Unknown blocks cloud calls; existing deny survives resume. Reuse an unchanged allowed source without repeated permission questions.

Preserve the source and choose the scope

Inspect the actual local file's tracks, duration, frame rate, rotation and color metadata. Keep originals intact. Infer the audience, purpose and desired length from the brief and footage. Ordinary YouTube defaults to a regular video preserving its source aspect and resolution, not a 9:16 short. A 10–15-minute indoor talk is a useful existing use case, not a mandatory duration. If the user requests Shorts, read the checkout's .agents/skills/tiktok/SKILL.md and use that vertical workflow. For requested 16:9 reframing of non-16:9 footage, explicitly review padding/crop in FCP or a separate framing operation; the current session renderer preserves the source aspect.

Honor explicit no-upload/no-publication restrictions, including children's footage. Neither this command nor publication of the harness authorizes cloud transcription, posting, screenshots, or recording uploads. Cloud transcription is OpenAI then Azure only, after source-specific authorization; never run local Whisper/ASR. If upload is forbidden, use a sealed existing transcript when available. Otherwise continue local color/audio/framing previews and report that word-anchored story edits need a permitted transcript. Never invent words, alignment or review evidence.

Edit for sustained listening

For requested music or supporting visuals, apply the shared optional-pattern contract in .agents/skills/video-editing/SKILL.md. Favor sparse chapter, opening and ending placements for a talk when that matches the brief, and compare against no added music. Long-form YouTube does not imply continuous music or forced beat cuts.

For a request to hold attention, make the opening question or benefit concrete, keep the setup short enough to reach the first useful answer, then build chapters around meaningful changes in subject or emotion. Use source-grounded examples, purposeful B-roll, a wide-to-detail change or a modest punch-in where the explanation changes; retain longer holds when a demonstration needs inspection. A motivated J-cut or L-cut can carry an idea across a picture change if the source audio supports it. A sound hit, split screen, monochrome beat or color frame should clarify a comparison or reveal, with easing and quiet space around it; do not cover every cut. Check title and caption readability against faces and subject motion. For requested effects, follow the video_effects schema and exact render/evidence checks in the shared skill and docs/VIDEO_EFFECTS.ja.md, comparing the same source coverage to a quiet version. Do not infer viewer-retention gains or use an unverified trend as a mandate for faster cutting.

For a talk or explanation, retain context, reasons, transitions and a complete conclusion. Reduce repeated setup and unproductive pauses using actual word IDs and source times. Preserve negation, qualifications, names, word endings, breaths and emphasis. Do not apply a fixed silence threshold, force a short-form hook, remove every pause, or stretch a clip to a target length. Listen across each cut and check whether the result still makes sense without the picture. Use chapter spans to organize longer material; do not fabricate a source claim to improve the story.

Use the checkout's video-editing workflow to create the project/session, attach or obtain a permitted transcript, inspect word context, plan source-bound spans/omissions, render, inspect edit boundaries and revise. Actor records must identify the actual agent; agent selection is not human acceptance. Start with a review preview, then render the full reviewed plan using the checkout's approval/review gates. Record whether selection/review was by an agent or a human; never infer human acceptance. Match subtitles to this exact edit; inspect spelling, line breaks and timing. Deliver actual chapter timestamps derived from the final timeline when requested, not guessed source timestamps.

Finish the picture and audio

Choose a use case from the scene: indoor talk, daylight exterior, night, backlight, etc. Read docs/PRESET_CATALOG.md and compare styles on the same source interval. Start indoor explanations with natural or clean_editorial and readable facial brightness; retain natural skin and whites. The platform name does not determine a color grade. For color-only/local-only work, use preview PROJECT --styles ... without creating a transcript. Apple Log conversion must happen once; avoid stacking its Rec.709 transform with an FCP Camera LUT.

Measure speech and normalize according to the project. Check audible cuts, harshness, clipping, background noise and loudness across the full render. Do not synthesize a replacement speaker or add music by default. Treat loudness targets as project choices, not YouTube requirements. Offer an audio-only listening output alongside the video where the session supports it.

For a completed editorial workflow, deliver the local full MP4 and matching subtitles/review artifacts for the user to upload personally. For a local-only preview without a usable transcript, deliver that preview and its unresolved editorial scope instead. For FCP, package the reviewed session as a flat single-source FCPXML and distinguish separate LUT/subtitle/audio application from effects actually represented in XML. Verify full decode, duration, source/frame mapping and audio coverage. Keep technical checks, human viewing/listening, actual FCP import and YouTube playback as separate evidence. Do not label a preview as the full edit, claim checks that were not performed, or upload/publish without an explicit further instruction.

For structure alternatives, use session compare-candidates SESSION FIRST_RENDER SECOND_RENDER --mode structure with different sealed plans under the same brief/source pool/color/base audio/FPS/preview conditions. Include a natural version without retiming. Inspect ordered original source ranges and reasons, added/removed ranges including repeated uses, duration and retime operations. Play each full video independently; equal output times do not mean equal source moments. Selection exports are exact-video-SHA proposals, not adoption or human review. The default effects mode and same-plan timing mode retain their stricter boundaries. See docs/RETIME.ja.md; this capability is included in alpha.7.

Import a downloaded comparison proposal with session select-comparison SESSION --data-file comparison-selection.json --actor ACTOR --note REASON. The unadopted candidate restores the selected render's exact plan/project and retained transcript/context inputs, not the current plan with old effects. It validates video SHA, current brief, source and output time. Render the candidate before explicit adoption, then review the exact full output. Restoring older inputs does not restore human approval or publish a video. Legacy speech renders without input references require their actual embedded transcript; context is regenerated from that transcript and the same brief rather than inventing alignment. Included in alpha.7.

For motion_trail, read docs/VIDEO_EFFECTS.ja.md and propose an explicit session effects add operation on the actual mapping. Limit it to an observed movement within one shot; do not add it to natural edits. Render the unadopted candidate and compare against natural picture with the same audio. Inspect duplicated faces/hands, baked text, low-FPS stepping and phone readability; temporal_inputs proves frame/weight provenance, not aesthetic acceptance. It is full-frame temporal mixing, not a subject mask. This feature is included in alpha.7.

For experimental tracked_background, read docs/TRACKING.ja.md. Use a full-subject observed box on the retained pre-effects stage, not a torso-only pose box; pose requires manual masks for every selected frame. session track-effect --effect tracked_background prepares local GrabCut masks as an unadopted candidate. Keep the manifest/PNG/correction hashes and inspect every contour, hair, hands, clothing and occlusion on the exact rendered comparison. Manual full-size binary PNGs replace individual masks; never hide loss or fall back to a rectangle. Require the SHA-bound subject_masks visual check. Baked mix/video_only only; no depth or semantic matting claim. This alpha.7 feature is experimental: an eight-frame real-footage trial kept wall/ceiling in the foreground and the agent rejected the visual result. No human review is recorded. Require full contour review and manual masks before use; do not claim universal person matting.

For development compound effects, read docs/MOTION_TEMPLATES.ja.md and inspect motion-templates. Use session motion-template on an actual render with explicit frame-aligned interval, observed purpose, parameters and reduced_motion choice. The beat_focus recipe combines zoom/saturation; reveal_callout sequences zoom and a supplied keyword. Reduced variants limit animation, not all visual change. No beat, depth or native Motion/TikTok analysis is implied. Recipes remain unadopted and retain source-render/mapping SHA provenance; compare a natural baseline and review the exact final video. Do not bypass natural/off, composition or text collision checks. This development feature is not in alpha.7.

For development relative-depth image layers, read docs/DEPTH_LAYERS.ja.md. Prepare or infer fields on the full render’s retained visual-graded.mp4, preserving raw fields and interval-shared normalization; use depth correct for observed frame corrections. If flicker warrants it, propose opt-in depth stabilize with declared source-frame cuts; preserve the original, inspect recorded coverage/resets and compare actual moving contours. Low confidence resets history. Propose session depth-layer with a registered same-canvas RGBA asset and permitted visual-asset policy, compare a natural baseline, and retain the candidate through selection/adoption. Require exact-final-SHA depth_contours visual review of hair, hands, all performers, occlusions and flicker. Unrelated audio/effect revisions retain depth; changed grading must regenerate fields. Natural reset removes depth. Baked mix/video_only handoff preserves redacted field provenance; models and original images remain external. Relative depth is not a person mask, metric distance or guaranteed temporal consistency. A technical pass is not human acceptance or a finished video. This development feature is not in alpha.7.

For development portrait caption wrapping, read docs/CAPTION_LAYOUT.ja.md. Use an explicit ja/en layout and observed protected terms with tiktok-export --caption-layout; preserve reviewed SRT timing and manual breaks. The actual font/stroke ink must fit three measured lines; revise oversized cues rather than inventing timing or splitting terms. Inspect missing glyphs, reading time, subjects and platform UI on the exact portrait render. Portable delivery retains visible-line/runtime proof without input paths or unused protected terms. This is not semantic transcript regrouping, complete JLREQ compliance or an official platform safe zone; it is not in alpha.7.

For development editable FCP visual cues, read docs/FCP_OVERLAY_PLACEMENT.ja.md. Image/video/title exports retain static placement and actual source dimensions through a shared preview pixel model; preserve its pending FCP GUI calibration status. Inspect geometry against the exact preview, and do not claim playback parity from DTD or XML comparison alone. Return import accepts unchanged placement and rejects unknown spatial edits/locations/automation. Audio ducking/normalization parity remains a separate unfinished stage. This is not in alpha.7.

For development audio-cue renders, read docs/AUDIO_GAIN_EVIDENCE.ja.md. Retain measured sample gain arrays and the actual speech/environment and mixed WAV inputs; cache restoration must validate their fingerprints. For an explicitly requested experimental editable handoff, set fcp_handoff: editable and fcp_audio_automation: measured; bounded dB keyframes and separate dialogue PCM retain pre-normalization gains. Apply a final common normalization scalar only after all-sample proof against retained pre-AAC WAVs. Return import allows supported visual cue revisions while preserving all generated audio; reject changed audio automation/ranges/dialogue. Preserve pending FCP GUI calibration; inspect original-audio muting, loop source clocks and interpolation in FCP. Final normalization/playback parity remains pending. Local evidence references can contain private paths. Review the exact MP4 by listening; never infer human approval from gain data. This is not in alpha.7.

For development visual retime, read docs/RETIME.ja.md: default --timeline-settings migrate maps compatible explicit cues, pulse effects and fixed guides through exact frame membership into an unadopted candidate. Preserve the sealed migration evidence and inspect the actual rendered mapping. Phase maps retain old-frame zoom/saturation values and keyword-title fade/rise during holds and omissions. A title beginning at a surviving visible old frame must still pass collision checks. BGM plays normally on the new clock; its declared beat offset follows the music. Video overlays use version-2 phase maps selecting actual-output RGBA samples sealed by the original render, rather than reconstructing input clocks. Preserve the layer fingerprint and source/trim/placement digest. Changed opacity/placement/trim, layer tamper or preview canvas mismatch requires a fresh full original render; do not upscale the layer. Old timestamp-only maps need a fresh original render. Private layer references stay in the local session; baked MP4 retains the picture. Sealed version-2 video layers preserve loop position and baked fade. SFX content, including loops, follows the continuous picture map with pitch-preserving audio retime; use baked mix/video_only handoff. Wipe/trail content uses sealed original-stage samples through content_map when the original render retained them. Keep original event/plan/asset/order and picture-setting bindings, original and remapped layer hashes, and exact frame selections. Changed grade, overlays or effect parameters requires a fresh original candidate. Old renders without samples allow only complete one-to-one translation; speed changes need a fresh full render. All temporal stages keep their captured order when omissions collapse start times. Stale/omitted/unaligned intervals, ambiguous beat anchors, insufficient media/fades, tracked artifacts and previously retimed mapping must reject without silently dropping settings. Depth requires regeneration. Explicit --timeline-settings clear removes old settings with a report; speech-setting migration remains unsupported. Compare and review the exact full MP4; migration is not perceptual approval. This development feature is not in alpha.7.

For development semantic caption grouping, read docs/CAPTION_GROUPS.ja.md. Use session caption-source on a sealed speech render, then session caption-groups --spec-file to propose exact source word-ID/occurrence groups. Interpret requests such as keeping negation, names and units together using actual words; preserve complete ordered coverage and record reasons. Set ja/en, exact protected phrases and editor-selected reading thresholds; never use example IDs or invent timing/spelling. Unknown/missing/reordered refs, stale spelling and groups crossing edit junctions reject. Caption times follow actual cut-word times, retimed word frames or J/L audio samples. Review reading/overlap/punctuation diagnostics, the retained SRT and exact resulting MP4; diagnostics and agent grouping are not human approval. Regular YouTube uses separate SRT; burn captions into vertical MP4 through the existing export and review route when requested. Preserve caption-evidence.json, its SRT SHA and sealed delivery copy. Old renders without word evidence require a fresh render. This feature does not provide full semantic understanding, JLREQ or platform-UI compliance and is not in alpha.7.

In a post-alpha.7 checkout exposing SFX audio_retime, visual session migration preserves effect audio, including loops, through the continuous picture ramp. Original cue fades are applied before pitch-preserving stretch; freezes insert silence and resume source PCM. BGM remains normal speed and ducking is calculated against the new speech clock. Keep source/trim/fade/mapping SHA bindings and content-retime evidence. Changing the original fade or trim requires a fresh original candidate; do not remove these bindings to force reuse. Looped SFX binds its original decoded period and seam taper in version-2 metadata, repeating on the original clock before stretch; do not restart loops on the new clock. Version-1 nonloop metadata remains readable. Composed SFX retime is not supported. Use baked mix/video_only handoff; editable FCP must reject normal-rate substitution. Synthetic sample/pitch tests do not replace full human listening or device playback review.

In a post-alpha.7 checkout with video_loop_periods, video loop cues repeat only the selected source trim after one output-CFR conform. Keep the actual period-frame count, source/period SHA and sealed RGBA layer. Sealed version-2 phase replay preserves loop position and baked video fade/alpha; never replace old loop content with a newly started loop. Visual cue fades apply alpha. In development first-retime migration, image/static title fades select sealed version-2 RGBA samples from the original output clock, preserving alpha during holds and omissions. Keep source/image or generated title raster SHA, text, placement, opacity and fade bindings. Old static renders without samples allow only complete one-to-one translation; otherwise render a fresh original. Changing loop, trim, placement, opacity or a nonzero video fade requires a fresh original layer. Inspect entry, seam, exit and any repeated gesture. Editable FCP must reject unrepresented loops/visual fades; use baked mix/video_only. Technical synthetic checks do not approve aesthetic repetition or real-device playback.

For development comparison-page strength/off adjustments, follow the shared video-editing skill and docs/COMPARISON_ADJUSTMENTS.ja.md; review the newly rendered full result for clear speech and readable demonstrations. Page undo resets staged proposals, not adopted edits. This is not in alpha.7.

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.