v0.8.0「照着爆款切、坏片自己修」/ Clip like the viral one, repair itself

v0.8.0「照着爆款切、坏片自己修」/ "Clip like the viral one, repair itself"

这版做两件事:让 HotClip 照着你给的爆款学节奏,并且把自己质检出的问题当场修好。

🎯 参考视频驱动——丢一条对标爆款,照着它的节奏切

• pnpm cli highlights/clip <视频> --reference <对标爆款>:本地实测参考片的时长、语速(说话时段口径)、平均句长、镜头切换频率(TransNetV2)、开场钩子形态,生成风格画像
• 画像注入找爆点提示词,选段向对标节奏靠拢——是偏好不是硬约束,内容本身的爆点潜质永远优先,不为凑节奏选废片
• 参考片同样全程本地分析(转写带缓存),素材照旧不出电脑;分析失败明确告知并按无参考继续

🔧 出片自我修复——质检从「报告」升级为「修复循环」

• v0.7 的质检会告警,v0.8 对可自愈的告警当场修:首尾长静音/黑屏自动裁边(留呼吸垫,新边界 30ms 淡化)、响度偏离/真峰值超标只重编音频二遍归一(视频流复制零画质损失)
• 修完重跑一遍质检,告警严格变少才替换成片——机器不能把片修得更糟;修复动作与前后对照记进 clips.json 的 qa.repair,CLI/MCP 回执同步显示「🔧 已自动修复」
• 有意不碰的:片中静音/黑屏是内容取舍(不归机器管),半词告警保持仅提示(上游本就有词边界守卫)

🚫 平台违禁词 lint——发布前点名,零上传

• 120+ 条本地正则规则、8 大类:绝对化用语(广告法)/医疗功效宣称/夸大承诺/收益承诺/价格误导/站外导流/权威背书/迷信博彩
• 每条切片的标题、钩子、发布文案、字幕文本逐一扫描,命中词+类别+出处写进 qa.contentHits 并汇总成告警——哪条片的哪句话可能限流,发布前一目了然
• 纯本地规则零 API 零上传;是风险提示不是审查,最终取舍在你

💥 Hormozi 大字爆点字幕

• 新字幕样式「大字爆点」:特大加粗、厚描边硬阴影、60% 高度线、短块顶入、逐词卡拉OK点亮、英文自动全大写——海外带货/营销短视频的标配形态,配商品讲解模式正合适

🎞 抽帧接触表——VLM 一次看九帧

• 视觉爆点信号改为九帧拼一张带序号九宫格、一次调用批量打分:VLM 调用数 20→3,抽帧额度反而 20→27,长视频视觉研判更快更省
• 切片审阅台新增**「画面速览」**:候选片段内 3×3 抽帧一图排开,不点播放就能扫完整个片段的画面

底线不变

本地处理、素材不出电脑;真免费、无水印、不限时长;每个 AI 决定可见、可改、可推翻——现在机器自己修过什么,也逐条可审计。

---

Two things in this release: HotClip now learns pacing from a viral clip you hand it, and fixes the problems its own QA finds.

🎯 Reference-clip driven detection

• pnpm cli highlights/clip <video> --reference <viral-clip>: the reference clip's duration, speech rate (speaking-time based), average sentence length, shot-cut frequency (TransNetV2) and opening-hook shape are measured locally into a style profile
• The profile steers highlight selection toward the reference's rhythm — a preference, never a hard rule: genuine viral potential always wins
• The reference is analyzed fully on-device (cached transcription); footage still never leaves your machine

🔧 Self-repairing QA loop

• v0.7 reported problems; v0.8 fixes the fixable ones on the spot: leading/trailing silence & black frames are edge-trimmed (with breathing room and 30ms fades), loudness deviation / true-peak overshoot gets a second normalize pass with the video stream copied untouched
• The repaired clip is re-checked and only kept when warnings strictly decrease — the machine may never make a clip worse; every action lands in clips.json → qa.repair, and CLI/MCP receipts show a 🔧 line
• Deliberately untouched: mid-clip silence/black (that's a content call) and mid-word-cut warnings (upstream already has word-boundary guards)

🚫 Platform banned-words lint (zero upload)

• 120+ local regex rules across 8 categories: absolute claims, medical claims, exaggerated promises, income promises, price bait, off-platform funneling, authority endorsement, superstition/gambling
• Each clip's title, hook, post copy and caption text are scanned; hits (term + category + source) land in qa.contentHits with a summary warning — know which line may get you throttled before you publish

💥 Hormozi-style impact captions

• New caption preset: big bold chunks, thick outline + hard shadow at the 60% line, punch-in per chunk, word-by-word karaoke lighting, auto ALL-CAPS for Latin — the marketing-short standard, made for product mode

🎞 Contact-sheet frame scoring

• The vision signal now packs 9 numbered frames into one grid per VLM call: calls drop 20→3 while the frame budget grows 20→27
• The clip review modal gains a frame overview — a 3×3 strip of the clip, scannable without pressing play

Same baseline

Local processing, footage never leaves your machine; genuinely free, no watermark, no length caps; every AI decision is visible, editable, reversible — and now every machine-made repair is auditable too.