zenstory-ai/video-recap-skills / skills/video-recap
video-recap
Verified把一段或多集视频做成带画框字幕的中文解说成片或纯原声高光短片。用户提供 .mp4 / .mov / .mkv / .webm,要求做影视解说、短剧二创、加旁白配音、跨集混剪或导出剪映草稿时使用;串联完成看片识台词、定主线与留原声、按台词停顿切镜头、写词检查、TTS 配音与混音加字幕。触发词:视频解说、影视解说、视频旁白、生成解说、视频 recap、video recap、voiceover、narration、recap。
scripts/briefing/builder.py
View on GitHub"""Build the agent narration brief from validated local evidence."""
import json
from pathlib import Path
from lib import CONFIG, load_background_research, log
from agent_text import _format_frame_facts
from briefing.context import (
_format_background_research,
_format_moderation_refusals,
_scene_number,
)
from briefing.timeline import (
_anchor_boundary_use,
_format_output_clip_list,
_format_sentence_entry_anchors_for_brief,
_load_cut_output_spans_for_brief,
_parse_target_seconds,
_remap_brief_evidence_to_output_timeline,
write_multi_source_output_speech_evidence,
)
def _quiet_windows(silence_periods):
min_quiet = CONFIG["quiet_window_min"]
return [
(s["start"], s["end"])
for s in silence_periods
if not s["has_speech"] and s["duration"] >= min_quiet
]
def _scene_asr_lines(asr_result, scene):
return [
f" [{seg['start']:.1f}-{seg['end']:.1f}] {seg['text']}"
for seg in asr_result
if seg["end"] > scene["start"] and seg["start"] < scene["end"] and seg["text"]
]
def _quiet_windows_for_scene(silence_periods, scene):
return [
(max(qs, scene["start"]), min(qe, scene["end"]))
for qs, qe in _quiet_windows(silence_periods)
if qe > scene["start"]
and qs < scene["end"]
and min(qe, scene["end"]) - max(qs, scene["start"]) >= 0.8
]
def _duration_label(seconds):
"""'45s' / '2min' / '1m42s' / '2h' / '2h01m' / '2h01m05s', rounded to whole seconds."""
whole = int(seconds + 0.5)
if whole < 60:
return f"{whole}s"
minutes, secs = divmod(whole, 60)
if minutes < 60:
return f"{minutes}min" if secs == 0 else f"{minutes}m{secs:02d}s"
hours, minutes = divmod(minutes, 60)
if minutes == 0 and secs == 0:
return f"{hours}h"
return f"{hours}h{minutes:02d}m" if secs == 0 else f"{hours}h{minutes:02d}m{secs:02d}s"
def _extract_manifest_tts_provider(manifest):
if not isinstance(manifest, dict):
return None
for candidate in (
manifest.get("tts_provider"),
(manifest.get("audio") or {}).get("tts_provider") if isinstance(manifest.get("audio"), dict) else None,
(manifest.get("settings") or {}).get("tts_provider") if isinstance(manifest.get("settings"), dict) else None,
):
if isinstance(candidate, str) and candidate.strip():
return candidate.strip().lower()
argv = manifest.get("argv")
if isinstance(argv, list):
for idx, token in enumerate(argv):
if token == "--tts-provider" and idx + 1 < len(argv) and isinstance(argv[idx + 1], str):
return argv[idx + 1].strip().lower()
if isinstance(token, str) and token.startswith("--tts-provider="):
return token.split("=", 1)[1].strip().lower()
return None
def _resolve_tts_provider_hint(work_dir, tts_provider=None):
explicit = tts_provider or CONFIG.get("_explicit_tts_provider")
if isinstance(explicit, str) and explicit.strip() and explicit.strip().lower() != "auto":
return explicit.strip().lower()
if work_dir is not None:
base = Path(work_dir)
candidates = [base / "recap_run_manifest.json"]
if base.parent.name == "sources":
candidates.append(base.parent.parent / "recap_run_manifest.json")
for manifest_path in candidates:
if manifest_path.is_file():
try:
manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
found = _extract_manifest_tts_provider(manifest)
if found:
return found
except (OSError, ValueError):
pass
if isinstance(explicit, str) and explicit.strip():
return explicit.strip().lower()
return str(CONFIG.get("tts_provider") or "auto").strip().lower()
def _narration_example_item(provider, *, end=11.2, text, emotion="紧张", include_policy=True):
item = {
"start": 0.2,
"end": end,
"narration": text,
"pause_after_ms": 250,
"overlaps_speech": True,
}
if provider in {"auto", "mimo-tts"}:
item["emotion"] = emotion
if include_policy:
item["source_entry_policy"] = "sentence_boundary"
return json.dumps(item, ensure_ascii=False)
_TAIL_CREDITS_MARKERS = ("演职员表", "制作人员", "制作组名单", "片尾字幕", "片尾滚动", "下集预告")
def _visual_subtitle_and_credits_callout(
scenes_analysis, brief_scenes, video_duration, *, cut_pass2=False
):
"""Surface hardcoded-caption and tail-credits signals already observed by VLM."""
caption_hits = 0
for scene in brief_scenes or []:
if not isinstance(scene, dict):
continue
desc = str(scene.get("description") or "")
facts_obj = scene.get("frame_facts")
facts_text = (
" ".join(
str(item)
for items in facts_obj.values()
if isinstance(items, list)
for item in items
)
if isinstance(facts_obj, dict)
else ""
)
cleaned = (desc + " " + facts_text).replace("片尾字幕", "").replace("滚动字幕", "")
if "字幕" in cleaned:
caption_hits += 1
credit_scenes = []
if not cut_pass2 and float(video_duration or 0.0) >= 60.0:
cutoff = float(video_duration) * 0.55
for scene in scenes_analysis or []:
if not isinstance(scene, dict) or float(scene.get("start", 0.0)) < cutoff:
continue
desc = str(scene.get("description") or "")
if any(marker in desc for marker in _TAIL_CREDITS_MARKERS):
credit_scenes.append(scene)
bullets = []
if caption_hits >= 3:
bullets.extend([
"- **检测到源片多处画面带有内嵌硬字幕**:",
" - **专有名词校准**:ASR 易把人名、地名或设定词转写成同音字,定稿前请对照下方场景的画面文字 /【帧标签】与 `background_research.json` 核准字形。",
" - **烧录防叠字**:若成片烧录解说字幕,请确认是否开启底部遮罩(`MASK_SOURCE_SUBTITLES=1` 或实测 `--subtitle-y-top/--subtitle-y-bot`)或错开字幕位置;未全片不透明遮罩留白段时无需重复写 `original_subtitles.json`。",
" - **封面避字**:在 `promo_copy.json` 中挑选 `cover.timestamp` 时,优先选无底部硬字幕干扰的干净镜头帧。",
])
if len(credit_scenes) >= 3:
c_start = min(float(s["start"]) for s in credit_scenes)
c_end = max(float(s["end"]) for s in credit_scenes)
bullets.append(
f"- **检测到片尾演职员表 / 预告区间参考**:源片约 `{c_start:.1f}s–{c_end:.1f}s`(共 {len(credit_scenes)} 个场景),选段剪辑时注意剔除(除非其中含正片关键彩蛋)。"
)
if not bullets:
return []
return ["## 源片画面文字与片尾提示", "", *bullets, ""]
def _missing_research_callout(research, context_info=""):
if research or str(context_info or "").strip():
return []
return [
"## ⚠️ 缺少背景资料(建议先补齐再定稿)",
"",
"当前未找到 `background_research.json`,且 `--context` 为空。缺少角色真名、身份关系与前情利益冲突时,解说容易滑向复述画面表象。",
"定稿 `recap_story_plan.json` 或 `narration.json` 前,建议先核实人物关系与故事背景并写入 `background_research.json`(随后可用 `understand.py --brief-only` 刷新本简报)。",
"",
]
def _emotion_section(provider):
if provider == "index-tts":
return [
"## 旁白语气标签(`emotion` 字段)",
"",
"当前 TTS 供应商为 `index-tts`:端点请求体固定为 `{\"voice\", \"text\"}`,不接收段级 `emotion`、`style`、`rate`、`pitch`,请勿在 `narration.json` 中填写这些字段(否则配音阶段会直接报错退出)。",
"",
]
if provider == "fish-audio":
return [
"## 旁白语气标签(`emotion` 字段)",
"",
"当前 TTS 供应商为 `fish-audio`:引擎不消费段级 `emotion` 标签,可省略该字段,主要靠选词节奏与 `pause_after_ms` 控制呼吸感。",
"",
]
return [
"## 旁白语气标签(`emotion` → MiMo TTS 语气指令)",
"",
"使用 MiMo TTS 时,每段 `emotion` 可写 1–2 个简短中文语气词(如 `\"紧张 深沉\"` 或 `\"无奈\"`):",
"- 基础情绪:开心 悲伤 愤怒 恐惧 惊讶 兴奋 委屈 平静 冷漠",
"- 复合情绪:怅然 欣慰 无奈 愧疚 释然 嫉妒 厌倦 忐忑 动情",
"- 整体语调:温柔 高冷 活泼 严肃 慵懒 俏皮 深沉 干练 凌厉",
"仅中性铺垫用「平静」,冲突推进优先选贴合戏剧张力的语气词;若后续改跑 `index-tts`,需去掉段级 `emotion` 字段。",
"",
]
def _floor_gates_and_emotion_section(provider="auto"):
return [
"",
"## 底线校验与质量提示(validate.py / shot_review.py)",
"",
"- **硬阻断(不过检不进 TTS / 合成)**:",
" - 句尾与原声切入完整性(`incomplete_sentence` / `interrupts_source_sentence`):旁白须以句末标点收尾;切入原声对白须对齐句末停顿锚点或静音窗,避免半句掐断。",
f" - 残镜闪帧与镜头切词(`flash_frame_sliver` / `shot_speech_conflict`):单段剪辑最短 ≥ {CONFIG['min_clip_duration']:.1f}s,`shot_review` 会拦截 ≤ {CONFIG['flash_sliver_max_seconds']:.2f}s 的残帧。",
" - 定稿字数与占位符(`over_budget` / `placeholder_leakage`):full 模式下超出窗口硬上限的段落会退回改稿,不静默删字;严禁残留模板占位词。",
"- **非阻断提示(写入报告供复核)**:",
" - 原声释放复核(`unsafe_source_release`):旁白收口落在粗粒度/对白区间内且后续 3s 无句末停顿锚点时提醒试听确认,避免在角色半句话中途恢复原声。",
f" - 开场与留白节奏(`weak_cold_open_silence` / `excessive_dead_air`):开场静默 > {CONFIG['max_cold_open_silence_seconds']:.1f}s 或连续 > {CONFIG['max_dead_air_seconds']:.1f}s 无旁白无对白时给出提示;若是有意留给动作或冷开场的呼吸空间可保留。",
" - 角色称呼与套话提醒(`generic_character_label` / `editorial_commentary_tone`):已有角色真名时提醒避免外貌泛称;提醒删掉空泛升华总结腔,结局由画面与原声自然落定即可。",
"",
"## 原声留白字幕 `original_subtitles.json`(可选)",
"",
"解说块之间留出的原声对白,可通过 `original_subtitles.json` 烧录成字幕(与解说字幕以 `「」` 区分):按成片 OUTPUT 时间轴写成 `[{\"start\": 秒, \"end\": 秒, \"text\": \"台词\"}]`。",
"- 只写留白里实际听得清且完整的原声台词;被旁白压低盖过或已剪掉的句子不要写。",
"- 订正 ASR 错别字与人名,单条尽量控制在约 20 字以内;不写该文件时留白区间不会自动烧录粗粒度 ASR。",
"",
*_emotion_section(provider),
]
def build_agent_brief(
scenes_analysis,
asr_result,
silence_periods,
video_duration,
work_dir,
style="纪录片",
*,
asr_evidence=None,
extra_context=None,
):
"""Write a compact Chinese brief that guides the agent to author recap artifacts."""
effective_rate = (
CONFIG["speech_rate"]
* CONFIG["speech_safety_margin"]
* CONFIG["narration_speed"]
)
breath_sec = CONFIG["breath_ms"] / 1000
target_pause_ms = CONFIG["breath_ms"]
edit_mode = CONFIG["edit_mode"]
target_duration = CONFIG["target_duration"] or "(未指定)"
provider = _resolve_tts_provider_hint(work_dir)
output_seconds = video_duration
cut_pass2 = edit_mode == "cut" and (Path(work_dir) / "edited_source.mp4").exists()
if cut_pass2:
spans = _load_cut_output_spans_for_brief(work_dir)
output_seconds = max(span["output_end"] for span in spans)
elif edit_mode == "cut":
target_seconds = _parse_target_seconds(CONFIG["target_duration"])
if target_seconds:
output_seconds = min(video_duration, target_seconds)
output_label = _duration_label(output_seconds)
source_label = _duration_label(video_duration)
lines = [
"# 解说创作简报(Agent Narration Brief)",
"",
"请根据当前工作目录中的画面与语音证据编写所需 JSON 产物;CLI 只负责时间与底线校验、TTS 配音及音画合成,不代写正文。",
"",
f"- 风格指引 (--style,原样自由文本): {style}",
"- 请直接吸收 `--style` 的语气与节奏要求结合本片证据落笔,不要把 `--style` 映射成固定枚举或套用模板腔。",
f"- 剪辑模式: {edit_mode}",
f"- 源视频时长: {video_duration:.1f}s",
f"- 目标成片时长 (cut 模式): {target_duration}",
f"- 推荐语速预算: 扣除 {breath_sec:.2f}s 气口后约 {effective_rate:.2f} 汉字/秒",
]
if asr_evidence is None:
asr_evidence = {"status": "MISSING_OR_STALE"}
research = load_background_research(work_dir)
if extra_context is not None:
displayed_context = extra_context.strip() or ("(已并入下方背景调研)" if research else "(无)")
else:
displayed_context = CONFIG["context_info"] or "(无)"
lines.extend(
[
"",
"## ASR 时间证据说明",
"",
f"- 状态: {asr_evidence['status']}",
"- 空文本窗口表示未检出文本(UNKNOWN_NOT_PROVEN_SILENCE),仅定位大致区间,不等同于已核实的静音。",
"- 时间粒度: 供应商粗粒度窗口;逐词对齐: NOT_PERFORMED;对白边界: NOT_VERIFIED。",
"- ASR 起止时间与安静窗口仅供参考,不能直接当作已验证的对白切点;保留或裁剪原声台词前请直接试听并核对画面。",
"",
f"- 段落默认气口: {target_pause_ms}ms",
f"- 补充上下文: {displayed_context}",
"",
"## 动笔前的创作规划",
"",
"先根据用户意图与现有产物判断创作控制模式:`CREATE` 从零构建叙事主线;`DIRECTED` 落实用户已指定的结构、镜头或台词;`REVISION` 只修改最新反馈点并冻结其余已确认内容。",
"- `CREATE` 主线清晰时直接写 1 个选定方案即可,仅在源片存在多种剪法取舍时比较 2–3 个候选假设;`DIRECTED` / `REVISION` 直接沿用既有主线。",
"- `REVISION` 先列「本次修改项」与「冻结项」:改文风/字幕更 `style_card.json`,改选镜/原声分工更 `visual_audio_board.json`,仅当主视角(POV)或核心节拍变化时更 `recap_story_plan.json`。",
"写 `clip_plan.json` 或 `narration.json` 前,先写好或更新 `recap_story_plan.json` 与 `visual_audio_board.json`:",
"- `recap_story_plan.json` 记录导演立意、POV、主线因果以及由局势/关系/筹码变化驱动的节拍(`hook | setup | turn | escalation | payoff`)。",
"- `visual_audio_board.json` 记录每拍选用的具体画面与反应、入出点依据、原声锚点、`audio_owner` 与 `narration_job`。",
"- 叙事与声画分工定妥后再写 `style_card.json`(只管口吻与句式节奏,不做预设分类表)。",
"- 成片画布视觉包装写入 `packaging_plan.json`(`portrait_cinema` 等预设);若用户需要平台标题、封面文案等外部宣发备忘,单独写入 `promo_copy.json`,切勿混用同名文件。",
"",
]
)
lines.extend(_format_moderation_refusals(scenes_analysis))
lines.extend(_missing_research_callout(research, CONFIG["context_info"]))
lines.extend(_format_background_research(research))
brief_scenes, brief_asr, brief_silence = (
scenes_analysis,
asr_result,
silence_periods,
)
if cut_pass2:
brief_scenes, brief_asr, brief_silence = (
_remap_brief_evidence_to_output_timeline(
work_dir, scenes_analysis, asr_result, silence_periods
)
)
lines.extend(
_visual_subtitle_and_credits_callout(
scenes_analysis, brief_scenes, video_duration, cut_pass2=cut_pass2
)
)
lines.extend(_format_sentence_entry_anchors_for_brief(work_dir, edit_mode))
sample_text = "林砚刚推开档案室的门,就看见桌上那份三年前的结案报告被人换掉了签名。"
if edit_mode == "cut":
cut_target_example = (
target_duration if target_duration != "(未指定)" else "30m"
)
if not cut_pass2:
lines.extend(
[
"## Cut 模式 — 第 1/2 步:完成故事与声画设计,仅写出 `clip_plan.json`",
"",
f"目标:从 {source_label} 源片剪出约 ~{output_label} 的解说成片(体量按 ~{output_label} CUT OUTPUT 规划,而非按 {source_label} 源片铺满)。",
"先确认观众承诺、POV 与戏剧问题并落盘 `recap_story_plan.json` + `visual_audio_board.json`,再挑选源片片段写入 `clip_plan.json`;CLI 渲染出 `edited_source.mp4` 后会进入第 2 步对成片写解说,本步不要写 `narration.json`。",
"选段要点(结合下方场景时间参考与 ASR):",
"- 开场直入冲突或异常,建立一条连贯的因果主线(`hook → setup → turn → escalation → payoff`),避免拼凑无因果关联的集锦。",
"- 保留真正推动局势翻转、亮底牌或关键情绪反应的镜头;删掉空泛过渡镜、重复寒暄、片头片尾演职员表、广告赞助及大面积水印片段。",
"- `reason` 用简明自然的中文写清该段承担的节拍作用与入出点依据(无需按固定段数拼竖线模板)。",
"- 单段剪辑时长随戏走(最短 ≥ 0.8s,不留 ≤ 0.20s 残帧),收尾段留出动作或表情落定的气口。",
"- 锁定切点前排查密集镜头切换:对原片自带切点,整段删掉无关碎镜并把有效镜头延展到动作/反应完整;对本次拼接制造的切点,挪动边界、恢复同源连续动作或合并相邻段,不要用转场遮掩坏接点。",
"- 保持原声台词完整:ASR 窗口边界仅供参考,落刀前请直接试听并核对画面;`cut.py` 也会结合 `silence_periods.json` 与 `speech_boundary_anchors.json` 做句界吸附与切词拦截。",
"",
"### clip_plan.json 格式(填写源片时间戳)",
"",
"```json",
"{",
f' "target_duration": {json.dumps(cut_target_example, ensure_ascii=False)},',
' "clips": [',
' {"start": 12.0, "end": 38.0, "reason": "b01 (hook):亮出结案报告签名被换的异常,入点卡推门定住,出点卡看清签名抬头"}',
" ]",
"}",
"```",
]
)
else:
lines.extend(_format_output_clip_list(work_dir))
example_row = _narration_example_item(
provider, end=11.2, text=sample_text, emotion="紧张"
)
lines.extend(
[
"## Cut 模式 — 第 2/2 步:按 OUTPUT 成片时间轴编写 `narration.json`",
"",
"先查看剪后故事板与上方保留片段映射,在 `visual_audio_board.json` 中补齐 OUTPUT 时间段并复核每拍的 `audio_owner` 与 `narration_job`。",
f"剪辑成片已渲染为 `edited_source.mp4` (~{output_label}),全片按 ~{output_label} CUT OUTPUT 时间轴(0 .. 总时长)落点,切勿填写源片时间戳。",
"",
"### narration.json 格式(OUTPUT 成片时间戳,0..总时长)",
"",
"每条记录对应一段连贯口播;交由画面、原声对白或静音承担的拍不写条目。示例文本仅供格式参考,请按本片证据撰写:",
"```json",
"[",
f" {example_row}",
"]",
"```",
]
)
else:
example_row = _narration_example_item(
provider, end=11.2, text=sample_text, emotion="平静"
)
lines.extend(
[
"## narration.json 格式要求",
"",
"写 `narration.json` 前先完成 `recap_story_plan.json` 与 `visual_audio_board.json`,按视听板决定哪些节拍需要旁白。",
"beat 对应关系记录在 `visual_audio_board.json`;`narration.json` 仍只承载时间、文本与朗读参数。",
"每条记录对应一段连贯口播;交由画面、原声对白或静音承担的拍不写条目。示例文本仅供格式参考,请按本片证据撰写:",
"```json",
"[",
f" {example_row}",
"]",
"```",
]
)
lines.extend(_floor_gates_and_emotion_section(provider))
if cut_pass2:
lines.extend(
[
"## 场景时间参考(OUTPUT 成片时间轴)",
"",
"以下仅列出保留在 `edited_source.mp4` 中的画面片段(已换算为 0 .. 总时长 OUTPUT 时间轴,剪除段落不列出)。写作时优先依据深层潜台词与 ASR 对白,画面参考仅用于核对切点,切勿改写成看图说话。",
"",
]
)
else:
lines.extend(
[
"## 场景时间参考",
"",
"写作时优先依据深层潜台词与 ASR 对白;画面参考仅用于核对切点,切勿改写成看图说话。",
"",
]
)
for scene in brief_scenes:
duration = scene["end"] - scene["start"]
quiets = _quiet_windows_for_scene(brief_silence, scene)
quiet_text = ", ".join(f"{s:.1f}-{e:.1f}s" for s, e in quiets) or "无"
span = f"{scene['start']:.1f}-{scene['end']:.1f}s"
lines.append(
f"### 成片 {span}(源场景 {_scene_number(scene['scene_id'])})"
if cut_pass2
else f"### 场景 {_scene_number(scene['scene_id'])}: {span}"
)
lines.append(f"- 时长: {duration:.1f}s")
if scene.get("depth_analysis"):
lines.append(f"- 深层分析: {scene['depth_analysis']}")
asr_lines = _scene_asr_lines(brief_asr, scene)
if asr_lines:
lines.append("- ASR 对白重叠:")
lines.extend(asr_lines[:8])
lines.append(f"- 安静窗口: {quiet_text}")
lines.append(
f"- 画面参考(仅供核对切点): {scene.get('description', '')}"
+ (
" [moderation_refused]"
if scene.get("analysis_status") == "moderation_refused"
else ""
)
)
facts = _format_frame_facts(scene)
if facts:
lines.append(facts.rstrip())
lines.append("")
brief_path = Path(work_dir) / "agent_narration_brief.md"
brief_path.write_text("\n".join(lines), encoding="utf-8")
log(f"已写入 Agent 解说写作 brief: {brief_path}")
return brief_path
def write_multi_source_clip_brief(work_dir, source_records, args):
"""Write the Pass-1 multi-source clip-planning brief using the shared brief context."""
work_dir = Path(work_dir)
research = load_background_research(work_dir)
lines = [
"# 多视频剪辑规划简报(Multi-source Clip Plan Brief)",
"",
"当前为多视频跨片源剪辑解说(通过 `--edit-mode cut` 跨源选段):先完成跨素材的故事主线与声画设计,写出 `clip_plan.json`;下一步剪出 `edited_source.mp4` 后,再按成片 OUTPUT 时间轴写 `narration.json`。",
"",
"## 创作决定",
"",
"先判断创作控制模式:`CREATE` 主线明确时写 1 个方案,存在多路线取舍时比较 2–3 个候选假设;`DIRECTED` 落实用户指定结构;`REVISION` 只改最新反馈点并冻结其余已确认内容。随后落盘 `recap_story_plan.json` 与 `visual_audio_board.json`:前者记录观众承诺、POV、戏剧问题与节拍推进;后者记录每拍选用的画面/反应、入出点、原声锚点、`audio_owner` 与 `narration_job`。",
"",
"多视频剪辑不是把每个片源各做一段小总结,所有选中片段都应推进同一条主线,并用 `source_id` 标明素材来源。",
"",
]
lines.extend(_missing_research_callout(research, getattr(args, "context", "")))
lines.extend(_format_background_research(research))
lines.extend(
[
"## 必须写入的格式",
"",
"```json",
'{"target_duration":"10m","clips":[{"source_id":"src_xxx","start":12.0,"end":38.0,"reason":"b01 (hook):亮出结案报告签名被换的异常,入点卡推门定住,出点卡看清签名抬头"}]}',
"```",
"",
"- 每个 clip 必须带 `source_id`,单段时长 ≥ 0.8s,不留 ≤ 0.20s 的残镜头闪帧。",
"- `reason` 用简明中文写清该段在主线中的作用与入出点依据,收尾拍留足情绪或动作落点。",
"- `start`/`end` 填写对应 `source_id` 源视频的时间(秒)。",
"- 已从源证据确认必保的问答、反应或兑现段,可写入 `clip_plan.json.required_evidence`(`nodes` 登记 `id`/`source` 绝对路径/`start`/`end` 源片秒/`track`/`content`,`before` 登记先后顺序)。",
"- 不同 `source_id` 的相同时间段不算重叠;同一 `source_id` 内避免重复或重叠选段。",
"- 选片前请查看下方各片源目录(`sources/<source_id>`)中的 `scenes.json`、ASR 与单源简报。",
]
)
if getattr(args, "target_duration", None):
lines.append(f"- 目标时长:`{args.target_duration}`。")
lines += ["", "## 片源列表(Sources)", ""]
for s in source_records:
swd = work_dir / s["source_work_dir"]
lines += [
f"### {s['source_id']} — {s['source_name']}",
f"- path: `{s['source_path']}`",
f"- work_dir: `{swd}`",
f"- material_id: `{s.get('material_id')}`",
]
path = work_dir / "agent_narration_brief.md"
path.write_text("\n".join(lines).rstrip() + "\n", encoding="utf-8")
return path
def write_multi_source_output_brief(
work_dir, source_records, validated_plan_path, *, tts_provider=None
):
"""Write the Pass-2 multi-source OUTPUT narration brief using shared context & gates."""
work_dir = Path(work_dir)
plan = json.loads(Path(validated_plan_path).read_text(encoding="utf-8"))
source_by_id = {s["source_id"]: s for s in source_records}
speech_evidence = write_multi_source_output_speech_evidence(
work_dir, source_records, plan
)
research = load_background_research(work_dir)
provider = _resolve_tts_provider_hint(work_dir, tts_provider=tts_provider)
example_row = _narration_example_item(
provider,
end=4.2,
text="林砚推开档案室的门,发现结案报告早已被人换掉签名。",
emotion="紧张",
include_policy=False,
)
lines = [
"# 多视频成片解说简报(Multi-source Output Narration Brief)",
"",
"`edited_source.mp4` 已由多个源视频拼接渲染完成,请按剪后成片的 OUTPUT 时间轴编写 `narration.json`。",
"",
"## 更新创作决定",
"",
"先查看 `edited_source.mp4` 与下方成片保留片段映射,在 `visual_audio_board.json` 中补齐 OUTPUT 起止时间,并根据实际成片确认每拍的 `audio_owner`、原声锚点与 `narration_job`;若剪后顺序调整了主线推进,同步更新 `recap_story_plan.json`。",
"",
"beat 对应关系保留在 `visual_audio_board.json` 中;`narration.json` 只承载时间、文本与朗读参数。",
"",
]
lines.extend(_missing_research_callout(research))
lines.extend(_format_background_research(research))
lines.extend(
[
"## narration.json 格式",
"",
"```json",
f"[{example_row}]",
"```",
"",
"注意:`start`/`end` 是剪后成片时间,不是任一源视频时间。",
"",
"## 输出时间线写作规则",
"- `narration.json` 只使用 `edited_source.mp4` 的 OUTPUT 时间线。",
"- 先为每拍确定 `audio_owner` 与 `narration_job`;旁白与原声比例(如 7:3)不是配额,只在素材本身难以判断声画主次时作粗略参考。",
"- 对照下方保留片段映射与各片源 `work_dir` 核对人物、台词与前情,只讲背景、筹码与因果潜台词,避免复述画面表象。",
]
)
lines.extend(_floor_gates_and_emotion_section(provider))
lines.append("## 成片保留片段映射(output → source)")
for c in plan["clips"]:
src = source_by_id[c["source_id"]]
reason = f" — {c['reason']}" if c.get("reason") else ""
lines.append(
f"- output {c['output_start']:.3f}-{c['output_end']:.3f}s → "
f"{c['source_id']} `{src['source_path']}` "
f"source {c['source_start']:.3f}-{c['source_end']:.3f}s{reason}"
)
anchors = [
(row, use)
for row in speech_evidence["sentence_anchors"]
if (use := _anchor_boundary_use(row)) != "none"
]
if anchors:
lines += [
"",
"## 原声句末安全切入点",
"",
"`unverified` 点由粗粒度 ASR 窗口标点估算并吸附到短停顿:不会切断字词,但未必是整句结束(误差上限见 `±`),切入前请试听确认。",
]
for row, use in anchors:
if use == "verified":
label = row["confidence"]
elif "timing_bound_seconds" in row:
label = f"unverified ±{float(row['timing_bound_seconds']):.1f}s"
else:
label = "unverified"
lines.append(f"- {row['time']:.3f}s [{label}] ({row['source_id']})")
lines += ["", "## 片源目录(Source work dirs)"]
for s in source_records:
lines.append(f"- {s['source_id']}: `{work_dir / s['source_work_dir']}`")
path = work_dir / "agent_narration_brief.md"
path.write_text("\n".join(lines).rstrip() + "\n", encoding="utf-8")
return path
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |