Skip to content

v0.9.0「越用越懂你、装完就能跑」/ Learns your taste, works out of the box

Choose a tag to compare

@github-actions github-actions released this 24 Jul 06:09

v0.9.0「越用越懂你、装完就能跑」/ "Learns your taste, works out of the box"

这版做两件事:让 HotClip 记住你每一次「要/不要」的选择,并且把首次安装到出片的路修平

🧠 审阅反馈回流——越用越懂你的口味

  • 挑爆点页点「导出」时,你采用了哪些候选、否决了哪些(标题/钩子/评分/片长/关键词)自动落进本地偏好档,最近 40 场滚动保留
  • 下次找爆点,提示词自动带上「用户采用过的(同类优先)/否决过的(同类少选)」样例——并明确要求 AI 总结共性而不是机械避开字面;内容质量永远是第一判据,偏好只是参考
  • 桌面审阅积累,CLI / MCP / 录播监听只读同享——全托管模式没有人的决策,不会污染偏好档
  • 偏好数据是一个本机 JSON 文件,不出电脑;删掉即重新开始

🩺 环境自检 doctor——一条命令把环境查明白

  • pnpm cli doctor:逐项检查 ffmpeg/ffprobe、九个 AI 模型的安装状态(实际体积/断点余量)、LLM 端点连通性、磁盘余量、转写缓存,每条问题都给能照做的修复建议
  • pnpm cli doctor --download:把默认管线要用的模型(转写/人脸/表情/镜头检测)现在一次预下载好,不必等首次出片时现场下

📦 模型下载断点续传——1GB 模型断流不再从零重来

  • 下载改为逐字节落盘 + Range 断点续传:断网、断流、中途退出,重来时接着上次的字节继续下,还能跨镜像接力(国内代理断了换源续传)
  • 续传起点严格校验,镜像响应对不上就归零重下,坏档案永远进不了模型目录;解压校验兜底
  • 转写模型体积标注修正(实际约 1GB)——进度条终于是真的

✍️ 发布文案角度模板——标题有角度、收尾有召唤

  • 发布文案生成引入 8 种钩子角度(悬念/冲突/数字/反常识/情绪/干货/身份共鸣/紧迫感)× 5 类行动号召(关注/评论/分享/收藏/商品),标题按角度选型、结尾带 CTA
  • 每份文案带 angle / ctaType 标签可审计;CTA 行一并过平台违禁词 lint,紧迫感角度的提示词里明确划掉会撞 lint 的话术

🎯 参考爆款入口补全——三端都能「照着爆款切」

  • MCP 工具新增 referencePath 参数:Agent 一句话就能带着对标视频驱动检测
  • 桌面挑爆点页新增「参考爆款」按钮:选一条对标视频重新检测,画像行(时长/语速/句长/镜头/钩子)可见,分析失败明确提示不静默

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——现在连你的口味偏好也存在本机、随时可删。


Two things in this release: HotClip now remembers every keep-or-reject call you make, and smooths the road from first install to first clip.

🧠 Review feedback loop — it learns your taste

  • When you hit Export on the highlight picker, which candidates you kept and which you rejected (title / hook / score / duration / keywords) lands in a local preference file, rolling over the latest 40 sessions
  • The next detection run injects "previously kept (pick more like these) / previously rejected (pick fewer like these)" examples — with an explicit instruction to generalize the pattern, not dodge identical titles; content quality still rules, preferences only advise
  • Collected on desktop, shared read-only by CLI / MCP / watch folder — fully-automated runs make no human decisions and never pollute the file
  • Your preferences are one local JSON file that never leaves your machine; delete it to start fresh

🩺 Doctor self-check

  • pnpm cli doctor inspects ffmpeg/ffprobe, all nine AI models (install state / on-disk size / resumable bytes), LLM endpoint reachability, free disk space and caches — every issue comes with an actionable fix
  • pnpm cli doctor --download pre-fetches the default-pipeline models so your first clip doesn't start with a download

📦 Resumable model downloads

  • Downloads now write every received byte to disk and resume with Range requests: dropped connections, network switches or app restarts continue from where they stopped — even across mirrors
  • Resume offsets are strictly validated; a lying mirror resets the file, and archive extraction is the final integrity gate
  • The ASR model's size label is corrected (~1GB real) — the progress bar is finally honest

✍️ Structured post-copy templates

  • Post copy now draws from 8 hook angles × 5 CTA types, with auditable angle / ctaType tags on every piece and CTA lines going through the banned-words lint too

🎯 Reference-clip entrance completed

  • MCP tools gain a referencePath param; the desktop highlight picker gets a "reference clip" button with a visible profile row and explicit failure reporting

Same baseline

Local processing, footage never leaves your machine; genuinely free, no watermark, no duration limits; every AI decision visible, editable, reversible — and now your taste profile lives on your machine too, deletable any time.