Native Subtitle Quote Image: real video frames → 3:4 subtitle quote cards, an open Agent Skill (ChengYi)
从真实视频帧拼出字幕长图，发不发仍由人决定。 Stitch real video frames into subtitle cards, and leave publishing to a person.
程意（@ChengYi3629）开源的 Agent Skill：从本地视频或用户有权处理的 YouTube 链接出发，用字幕轨 / Whisper 建时间索引、选题选句、精确取帧，再把真实视频帧拼成 3:4 社交长图并逐张质检。两种模式不混用——原生字幕只保留画面像素、不 OCR 重绘；脚本字幕把已审核台词画到真实帧上并标明后期。仓库含 Codex 插件结构与开放 Skill 目录，可复制到其他兼容 Agent；作者推文称 2.0 后拿它做抖音，成图后发不发仍由人来。 ChengYi’s (@ChengYi3629) open Agent Skill: start from a local video or a YouTube link you have rights to, build a timestamped index from subtitle tracks / Whisper, pick topics and lines, grab exact frames, and stitch real frames into 3:4 social cards with per-image QA. Two modes never mix — native mode keeps pixel subtitles as-is (no OCR redraw); script mode draws reviewed lines onto real frames and labels them as post-added. Ships a Codex plugin layout plus an open skill directory that can be copied into other compatible agents; the author’s post says 2.0 is what they now use for Douyin, but publishing stays a human step.
为什么推荐Why we recommend it 公开仓库（2026-09-11 用 gh 核对约 616 star，MIT，最新 release v2.1.1 于 2026-08-24；topics agent-skills / codex / social-media / subtitles / video-processing）。开放 Agent Skills 目录格式，README 给的是 Codex 装法（$skill-installer 或复制到 ~/.codex/skills）；Grok Bot 侧需按官方 skills 文档把 skills/native-subtitle-quote-image 放进对应目录，本站未实测。推文（2026-09-09）称「升级到 2.0」「平均每个视频赞藏 1 千」为作者自称，不是本站核验数据；README Demo 图为脚本字幕模式，不代表原片自带字幕。示例图与第三方内容不在 MIT 授权内。 Public repo (~616 stars checked with gh on 2026-09-11, MIT, latest release v2.1.1 on 2026-08-24; topics agent-skills / codex / social-media / subtitles / video-processing). Open Agent Skills directory format; the README documents the Codex path ($skill-installer or copy into ~/.codex/skills). For Grok Bot, drop skills/native-subtitle-quote-image into the directory the official skills docs specify — not tested by this site. The post (2026-09-09) claiming an upgrade to 2.0 and “~1k likes+saves per video” is the author’s own claim, not a metric verified here; README demo images are script-subtitle mode, not proof the source video carried subtitles. Example images and third-party content are outside the MIT grant.
成图后发抖音 / 小红书 / X 前人审；素材使用权、原生→脚本模式切换、yt-dlp 读取浏览器 Cookie 都要用户明确同意，脚本台词必须已复核。 Human review before posting cards to Douyin / Xiaohongshu / X; rights to the footage, switching native→script mode, and letting yt-dlp read browser cookies all need explicit user consent, and script lines must be pre-reviewed.
先在可撤回、不外发的范围内试跑。本条未批准前：成图后发抖音 / 小红书 / X 前人审；素材使用权、原生→脚本模式切换、yt-dlp 读取浏览器 Cookie 都要用户明确同意，脚本台词必须已复核。。 Try it first in a reversible, non-outbound scope. Until approved: Human review before posting cards to Douyin / Xiaohongshu / X; rights to the footage, switching native→script mode, and letting yt-dlp read browser cookies all need explicit user consent, and script lines must be pre-reviewed..
这条用法已包含可复用 Skill。跑通安全范围后，立刻把步骤固化下来。 This use already includes a reusable Skill. After the safe pass, lock the steps in.
确认输出可靠后，再考虑定时或事件触发；例行默认仍应停在待审。 After output is reliable, consider a schedule or event trigger. Routines should still stop at review by default.
