This release focuses on tighter, evidence-grounded output and realistic speech pacing. Package: speaker-v8.skill (replaces speaker-v7.skill).
English
- Pause-aware pacing model — a deterministic budget (English ~110 wpm, Chinese ~165 characters/min) that reserves time for slide transitions and
[PAUSE]marks. Fixes the prior problem where a "15-minute" script ran to ~1,800 words and overran to ~30 minutes; a 15-minute talk now targets roughly 1,300–1,400 words. Chinese and English notes are tuned per slide. - Per-slide word budget —
SKILL.mdcomputes and records a per-slide budget; the timing table now reports words/characters, budget, and pauses with a TOTAL row. - Glossary toggle — a new on/off switch (default on) can skip the "Key Parameters And Methods" table end to end.
read_slides.py --mode compact— drops redundant raw OOXML dumps and non-visual geometry while keeping picture bounding boxes, producing much smaller intermediate JSON.visual_inventory.py --ocr-scope image-regions— OCR runs only on picture/media regions (text, tables, charts already come from XML), with automatic full-slide fallback. Includes byte-safe OCR decoding and symlink path resolution.vision_review.py --format compact— the shared review prompt and result schema are hoisted to the top level instead of repeated per slide; Markdown is now optional.- Concise final delivery — returns a summary plus file paths by default instead of pasting the full slide-by-slide script into chat. Reply
show notesfor the complete script.
中文
- 带停顿的配速模型 —— 确定性预算（英文约 110 词/分钟，中文约 165 字/分钟），为翻页过渡和
[PAUSE]停顿预留时间。修复了此前"15 分钟讲稿生成约 1800 词、却要半小时才念完"的问题；现在 15 分钟的演讲目标约 1300–1400 词。中英文讲稿都按页做了优化。 - 逐页字数预算 ——
SKILL.md会计算并记录每页预算；timing 表新增词数/字数、预算、停顿数，并带 TOTAL 合计行。 - 术语表开关 —— 新增 on/off 开关（默认 on），可全程跳过 "Key Parameters And Methods" 表。
read_slides.py --mode compact—— 丢弃冗余的原始 OOXML 转储和非视觉几何信息，同时保留图片 bounding box，显著减小中间 JSON 体积。visual_inventory.py --ocr-scope image-regions—— OCR 只对图片/媒体区域运行（文本、表格、图表已从 XML 获得），必要时自动回退整页 OCR。同时修复 OCR 字节安全解码与 symlink 路径解析。vision_review.py --format compact—— 把共用的审查提示词和结果 schema 提到顶层，不再每页重复；Markdown 改为可选输出。- 精简的最终交付 —— 默认只返回摘要 + 文件路径，而不是把逐页讲稿全部贴进对话。回复
show notes查看完整逐页讲稿。