将任何输入——GitHub仓库、URL、文章或原始文本——转化为 9：16 竖版旁白视频（Reels / Shorts / TikTok格式）。

与固定模板视频生成器不同，代理（Claude Code 或 Antigravity）可以 深度提取本身，编写叙事计划，然后将每个场景组装为 从目录中挑选模板并填充真实数据——一条受保护路径， 不是手写的HTML，所以安全区、品牌条和说明带总是正确的。 工具箱测量配音，对每个场景进行门槛，渲染并合成 最终mp4。

是什么让作品感觉专业

这些都是内置在流程中的（从几次9/10的拆解中得知的 回收巡回短片 — 参见skill/references/reference-video-teardown.md）：

• 单词同步卡拉OK字幕 — 每场景刻录，逐字填充 声音，带着数字词的重音。（lib/compose/subtitles.py）
• 真实的重现画面片段 — 真实的滚动屏幕截图 仓库（文件树→README→截图）或作者简介，框架在 浏览器显示真实网址。这是真实性，不是合成文本。 （lib/render/repo_footage.py）

any2video

Turn any input — a GitHub repo, a URL, an article, or raw text — into a 9:16 vertical narrated video (Reels / Shorts / TikTok format).

Unlike fixed-template video generators, an agent (Claude Code or Antigravity) does the deep extraction itself and writes the narrative plan, then assembles each scene by picking a template from a catalog and filling it with real data — a guarded path, not free-hand HTML, so the safe-zones / brand bars / caption band are always correct. The toolbox measures the voiceover, gates every scene, renders, and composes the final mp4.

What makes the output feel professional

These are baked into the pipeline (learned from a teardown of several 9/10 repo-tour Shorts — see skill/references/reference-video-teardown.md):

• Word-synced karaoke captions — burned per scene, filling word-by-word with the voice, number-words accented. (lib/compose/subtitles.py)
• Real repo-footage scenes — a genuine scrolling screen-capture of the live repo (file tree → README → screenshots) or an author profile, framed in a browser chrome bar with the real URL. An authenticity beat, not synthetic text. (lib/render/repo_footage.py)
• Beat-split reveals — list/bar items appear (and bars fill left→right) exactly as the narrator names them, not all at once.
• AV-sync by design — TTS is measured first; scenes are joined with a gap-hardcut so the picture never drifts ahead of the voice.
• Reads brand names right, shows them clean — the voice speaks tricky names phonetically (“README” → “ruýt my”, “any2video” → “en ni tu vi đeo”) while every on-screen surface (title, karaoke caption, post) keeps the exact spelling.
• Ducked background music — a bundled track sits ~−20 dB under the narration and swells in the gaps (on by default; random from the pool or pick one).
• Reads the actual source, not just the README — for a repo it clones and opens the entry point + core files, traces the real flow, and mines the non-obvious “weapon” and the honest caveat from the code. Every architecture/flow claim is cited to a file, not a marketing bullet, so the tour reflects what the repo does, not what its README claims.
• Template-driven scenes with 3-tier safe-zone enforcement for 9:16 — every scene is measured and gated before render (see below), so broken frames never ship.

How it works (pipeline)

input → 1. extract → 2. plan (+ narrative gate) → ✋ approve script → 3. TTS (measure)
      → 4. fill templates → 4.5 GATE every scene → ✋ approve stills → render → 5. compose → mp4

TTS runs before the visuals so every scene’s layout is sized to the real audio duration (no drift). See skill/SKILL.md for the full agent spec.

Two human checkpoints (optional, via Telegram) keep waste out of the loop:

1. Script — before TTS. You see each scene’s on-screen text **and the read-aloud