v0.13.0 整场看完、敢扔 / Watch it all, dare to drop
「整场看完、敢扔」——选段质量与人工掌控的一版。 / Watch it all, dare to drop.
质量门三档 / Three-tier quality gate
- AI 复评升级「零上下文终判」:假装没看过直播、只看这条片,判建议发 / 需人审 / 弃;确定性规则再抓一层硬伤(开头悬空的「所以/但是」、结尾截在逗号上、多人抢话密集)
- 判弃的折叠到列表底部、写明原因、可手动捞回;录播监听全托管只自动导出「建议发」档——AI 出片三四成是废片是行业现实,发废片拉垮的是整个账号
- Zero-context verdicts (publish / needs-review / drop) + a deterministic flaw catcher; dropped clips fold away with reasons and can be rescued; unattended mode only ships the publish tier
全场画面扫描 / Full-stream visual scan
- 约每 30 秒抽一帧把整场直播看完,画面事件(梗图/动作/翻车瞬间)带一句话描述直接进选段证据——补上纯文本选段看不见画面的盲区
- 本机 Ollama 免费;云端 qwen3-vl-flash 档整场约几毛钱,开关旁当面算清本场抽几帧、调用几次
- ~1 frame / 30s across the whole stream; on-screen events flow into selection evidence with descriptions; free locally, ~cents per stream on cloud, cost shown upfront
点题选段 / User brief
- 用自然语言告诉 AI「重点找什么 / 明确不要什么」(如「只要聊售后翻车的部分,不要抽奖和念弹幕」),填完即按点题重找
- Tell the AI what to hunt and what to skip, in plain words
主播口令打点 / Streamer clip commands
- 转写稿里识别「这段剪下来 / 剪个切片 / clip that」——主播亲口认证的爆点当最强证据(内容在口令之前,AI 会往前找)
- "Clip that" moments count as the strongest evidence there is
带货三段式 / Selling three-act stitch
- 「痛点 → 演示 → 价格」三要素散在直播各处时,自动拼成一条完整种草片;凑不齐不硬拼
- Pain point → demo → price, stitched only when the beats genuinely exist
文稿选段 / Transcript pick
- 逐句稿里点句成片(文字剪视频):搜台词、按说话人筛选、跨段任选,选中即成候选;手动拼接不受 AI 拼接段数护栏限制
- Build clips by clicking sentences; search lines, filter by speaker
连不上 AI 不再只甩报错 / No more dead-end LLM errors (#6)
- 选 Ollama 档提前提示要装什么;点开始先做配置预检——Ollama 没启动 / API Key 错 / 模型没拉取,当场拦下并给出下一步(附本机已装模型一键改选)
- Upfront Ollama hints, a preflight check that blocks doomed configs with next steps, and failure messages that say what to do
下载 / Downloads: macOS (Apple Silicon) .dmg · Windows x64 .exe/.zip