def start(task_id, params: VideoParams, stop_at="video"):
    # 1. 用 AI 生成剧本
    video_script = generate_script(task_id, params)
    # 2. 用 AI 提取视频搜索关键词
    video_terms = generate_terms(task_id, params, video_script)
    # 3. 把剧本变成语音（TTS）
    audio_file, audio_duration, sub_maker = generate_audio(task_id, params, video_script)
    # 4. 根据语音生成字幕
    subtitle_path = generate_subtitle(task_id, params, video_script, sub_maker, audio_file)
    # 5. 根据关键词从图库下载素材
    downloaded_videos = get_video_materials(task_id, params, video_terms, audio_duration)
    # 6-7. 拼接片段 + 叠加字幕配乐
    final_paths, combined_paths = generate_final_videos(task_id, params, downloaded_videos, audio_file, subtitle_path)
          

def generate_script(video_subject, language="", paragraph_number=1):
    prompt = f"""
    # Role: Video Script Generator
    ## Goals:
    Generate a script for a video, depending on
    the subject of the video.
    ## Constrains:
    1. return as a string with specified paragraphs.
    2. do not reference this prompt.
    3. get straight to the point.
    5. only return raw content of the script.
    8. respond in the same language as subject.
    # Initialization:
    - video subject: {video_subject}
    - number of paragraphs: {paragraph_number}
    """.strip()
          

def generate_terms(video_subject, video_script, amount=5):
    prompt = f"""
    # Role: Video Search Terms Generator
    ## Constrains:
    1. return as a json-array of strings.
    5. reply with english search terms only.
    ## Output Example:
    ["search term 1", "search term 2"]
    """.strip()
    search_terms = json.loads(response)
          

def tts(text, voice_name, voice_rate, voice_file, voice_volume=1.0):
    if is_azure_v2_voice(voice_name):
        return azure_tts_v2(text, voice_name, voice_file)
    elif is_siliconflow_voice(voice_name):
        return siliconflow_tts(text, model, full_voice, voice_rate, voice_file)
    elif is_gemini_voice(voice_name):
        return gemini_tts(text, voice, voice_rate, voice_file)
    return azure_tts_v1(text, voice_name, voice_rate, voice_file)
          

# 如果比例不同，等比缩放后加黑边
if clip_ratio > video_ratio:
    scale_factor = video_width / clip_w
else:
    scale_factor = video_height / clip_h
new_width = int(clip_w * scale_factor)
new_height = int(clip_h * scale_factor)
background = ColorClip(
    size=(video_width, video_height),
    color=(0, 0, 0)
).with_duration(clip_duration)
clip_resized = clip.resized(
    new_size=(new_width, new_height)
).with_position("center")
clip = CompositeVideoClip([background, clip_resized])
          

_clip = TextClip(
    text=wrapped_txt,
    font=font_path,
    font_size=params.font_size,
    color=params.text_fore_color,
    bg_color=resolve_subtitle_background_color(),
    stroke_color=params.stroke_color,
    stroke_width=params.stroke_width,
    text_align="center",
)
if params.subtitle_position == "bottom":
    _clip = _clip.with_position(
        ("center", video_height * 0.95 - _clip.h)
    )
          

if _enable_redis:
    task_manager = RedisTaskManager(
        max_concurrent_tasks=_max_concurrent_tasks,
        redis_url=redis_url
    )
else:
    task_manager = InMemoryTaskManager(
        max_concurrent_tasks=_max_concurrent_tasks
    )
task_manager.add_task(tm.start, task_id, params, stop_at)