[V3 Tracker] Xiaohongshu-tuned content engine — Phase 2-5 rollout

[maxazure]

Tracker for the V3 rollout

The full roadmap lives at docs/plans/2026-05-17-v3-xhs-improvements.md.

Phase status

• Phase 1 (V3.0) — day58 production regressions — merged in V3 Phase 1: small-red-book production fixes from day58 #9
• Phase 2 (V3.1) — Content Guard + Story Engine — in progress
• Phase 3 (V3.2) — Auto-Enrich (B-roll / chapter cards / beat-sync / stickers / imagegen)
• Phase 4 (V3.3) — Cover preset library + multi-platform export + caption generator
• Phase 5 (V3.4) — Documentation (docs/prompts/15-19)

What V3 turns this skill into

A 小红书-tuned content engine — not just an auto-cutter. It runs:

1. Story Engine — LLM rewrites your transcript into a 5-field JSON (hook / pain / turn / value[] / cta), using 8 hook templates and 5 CTA templates.
2. Content Guard — refuses to export if title/caption/script triggers a platform red-line (extreme words, medical claims, off-platform diversion, wealth-baiting, QR codes on frame, external-app watermarks).
3. Audience-aware defaults — --profile tech_pro sets cut-frequency, subtitle density, BGM gain, aspect ratio (3:4 for 小红书) automatically.
4. Auto-Enrich — schedules B-roll cutaways on entity mentions, draws chapter title-cards on silence boundaries, snaps cuts to BGM beats, drops emotion-mapped stickers, generates AI illustration for abstract concepts.
5. Multi-platform export — one source timeline → three deliverables (小红书 3:4 / 抖音 9:16 / 视频号 ≤60s).
6. Caption generator — title (≤18 chars, TF-IDF-keywords-first) + body (200-500 chars, emoji-paced, 3-6 tags) + publish-time hint.

Done definition

• All five phases merged
• pytest tests/ green at every merge
• Skill self-test passes: python3 scripts/utils.py

[maxazure]

✅ V3 全部 5 个 Phase 已完成并合并到 main

Phase	PR	状态
Phase 1 — Day58 production regressions	#9	✅ Merged
Phase 2 — Content Guard + Story Engine + Profiles	#11	✅ Merged
Phase 3 — Auto-Enrich pipeline	#12	✅ Merged
Phase 4 — Multi-platform export + Caption generator	#13	✅ Merged
Phase 5 — Docs (prompts 15-18 + README V3 section)	#14	✅ Merged

最终交付

13 个新脚本：

• 平台守门：_internal_text_guard.py, content_guard.py
• 故事引擎：rewrite_script.py + prompts/{hook,cta}_templates.yaml
• 自动丰富：auto_broll.py, auto_chapter_cards.py, auto_stickers.py, beat_sync.py, auto_enrich.py
• 多平台：multi_export.py, generate_caption.py
• 配置：profiles/{tech_pro,lifestyle,_fonts}.yaml + profiles/__init__.py

4 篇新教程：

• docs/prompts/15-xhs-daily-tech-video.md — V3 完整流水线
• docs/prompts/16-content-guard.md — 平台雷区 lint
• docs/prompts/17-multi-platform.md — 三平台导出
• docs/prompts/18-auto-enrich.md — 自动丰富度

131 个测试 全部通过（覆盖 mlx 后端、internal-text guard、字体默认、响度链、primary-speed、content_guard 80+ 规则、profile 加载、story engine 验证、3 个 auto-enrich 模块的调度逻辑、multi-export 比例转换、caption 合成 + guard 集成）。

Day58 教训全部覆盖

• ✅ 顶部漏 1.25x → _internal_text_guard 永久拒绝
• ✅ 字幕 Hiragino W3 太细 → 默认 Source Han Sans Heavy / STHeiti Medium
• ✅ 加速后中段听不清 → 渲染自动 dynaudnorm + acompressor + loudnorm
• ✅ 1.25× 想做主输出 → --primary-speed 1.25
• ✅ MLX whisper 本地改动未回流 → Phase 1 已 port
• ✅ 字幕里 ChatGPTT 等 OCR 错词 → Story Engine 验证 + Content Guard 拦截
• ✅ 没有故事结构指引 → 8 hook + 5 CTA + 3 structure 模板库

完整路线图：docs/plans/2026-05-17-v3-xhs-improvements.md

可以关掉这个 tracker 了。