andyhuo520/vibe-knowledge / skills/ai-teardown
ai-teardown
Unverified「AI 拆机课」系列课工作室:输入一个主题(AI 产品、开源项目/书、人物历史、软件教程),产出由浅入深、每句话有出处的中文讲解视频系列 + 每节互动版 + 课程总页。流水线:收集资料 → 确认对象与课时 → 大纲 → 逐节旁白(MiMo 克隆音色 + 逐字对齐)→ scenes.json 画面编排(Notion 极简手绘引擎 / 马卡龙模板引擎,IconPark 图标与插画、Noto 动态 emoji、AntV X6 流程图、像素讲解小人、截图圈注、终端录屏)或 film.js 代码成片(cine 引擎:Three.js、粒子、着色器、真片段叠加、曲库配乐)→ Hyperframes / 无头 GPU 逐帧 1080p 渲染 → Artifact 发布 → 课程总页 + 导学视频。触发:做系列课 / 做课程视频 / 把某本书或某个开源项目做成课 / 科普讲解视频 / 教学视频 / 课程地图 / 导学视频 / 互动课件 / 「拆机课」「上手课」「实战课」「编排课」/ 王炸讲解片 / 用真实会话和成片教人做视频。不用于:单支宣传片(用 promo-video)、单张信息图、纯文章写作。
engine/cine/tools/finish.py
View on GitHub#!/usr/bin/env python3
"""讲解课出片:读镜头表里的段落与音效点 → 生成背景乐和音效(cine/audio/tools/bed.py)→ 混音(旁白压低背景乐)
→ 拼上冷开场画面 → 成片。
用法(用装了 numpy、scipy、playwright 的 Python 跑,见 references/10-cine.md「环境」):
finish.py <课号> [--port 8964] [--skip-bed]
输出:lessons/<课号>/build/final_mix.wav、lessons/<课号>/build/lesson<课号>.mp4
"""
import json
import subprocess
import sys
import wave
from pathlib import Path
import numpy as np
from scipy import signal
ROOT = Path(__file__).resolve().parent.parent.parent # 系列目录(cine/ 和 lessons/ 的上一级)
CINE = ROOT / "cine"
SR = 48000
def arg(name, default):
return sys.argv[sys.argv.index(name) + 1] if name in sys.argv else default
def read_wav(p):
with wave.open(str(p), "rb") as w:
sr, ch, n = w.getframerate(), w.getnchannels(), w.getnframes()
x = np.frombuffer(w.readframes(n), dtype=np.int16).astype(np.float64).reshape(-1, ch) / 32768.0
if sr != SR:
x = signal.resample_poly(x, SR, sr, axis=0)
return x if ch == 2 else np.repeat(x, 2, axis=1)
def write_wav(p, x):
y = np.clip(np.round(x * 32767.0), -32768, 32767).astype(np.int16)
with wave.open(str(p), "wb") as w:
w.setnchannels(2); w.setsampwidth(2); w.setframerate(SR); w.writeframes(y.tobytes())
def film_info(lesson, port):
from playwright.sync_api import sync_playwright
from browser import launch
with sync_playwright() as p:
b = launch(p)
pg = b.new_page(viewport={"width": 1920, "height": 1080})
pg.goto(f"http://127.0.0.1:{port}/cine/index.html?lesson={lesson}&capture")
pg.wait_for_function("window.__film && window.__film.ready", timeout=180000)
info = pg.evaluate("({from: __film.from, cold: __film.cold, coldSrc: __film.coldSrc, duration: __film.duration, sections: __film.sections, sfx: __film.sfx, clips: __film.clips || [], music: __film.music || []})")
b.close()
return info
def envelope(x, att=0.04, rel=0.4):
"""旁白的平滑包络(0..1),用来压低背景乐。"""
a = np.abs(x.mean(axis=1))
win = int(0.05 * SR)
rms = np.sqrt(np.convolve(a * a, np.ones(win) / win, mode="same"))
on = (rms > 10 ** (-42 / 20)).astype(np.float64)
ka, kr = np.exp(-1 / (att * SR)), np.exp(-1 / (rel * SR))
# 一阶跟随:起得快、落得慢(用 lfilter 分两段近似:先按释放平滑,再取与起音平滑的最大值)
up = signal.lfilter([1 - ka], [1, -ka], on)
down = signal.lfilter([1 - kr], [1, -kr], on)
return np.clip(np.maximum(up, down), 0, 1)
def limiter(x, ceiling=10 ** (-1.5 / 20), look=0.003): # 留 1.5 dB 余量,AAC 编码后不过 0 dBFS
"""前瞻峰值限幅:起音即时、释放 50 ms。"""
from scipy.ndimage import maximum_filter1d
peak = np.max(np.abs(x), axis=1)
n = int(look * SR)
target = np.minimum(1.0, ceiling / np.maximum(maximum_filter1d(peak, size=2 * n + 1), 1e-9))
k = np.exp(-1 / (0.05 * SR))
g = np.minimum(signal.lfilter([1 - k], [1, -k], target - 1) + 1, target)
return x * g[:, None]
def main():
lesson = sys.argv[1]
port = arg("--port", "8964")
L = ROOT / "lessons" / lesson
B = L / "build"
info = film_info(lesson, port)
cold, dur, src = float(info["cold"]), float(info["duration"]), info.get("coldSrc")
total = dur + 0.6
lj = {"duration": round(total, 3), "bpm": 100, "key": "F minor", "chords": ["Fm", "Db", "Ab", "Eb"],
"sections": [{"t": round(s["t"], 3), "t_to": round(min(s["t_to"], total), 3), "level": s["level"], "name": s.get("name", "")} for s in info["sections"]],
"cues": [{k: (round(v, 3) if isinstance(v, float) else v) for k, v in c.items()} for c in info["sfx"]]}
if src: # 片头拼现成片段:配乐也从同一段原声里截
lj["cold_open"] = {"from_wav": str(ROOT / src["wav"]), "src_t0": src["t0"], "src_t1": src["t0"] + cold, "at": 0.0}
(B / "lesson.json").write_text(json.dumps(lj, ensure_ascii=False, indent=1), encoding="utf-8")
print(f"lesson.json:{len(lj['sections'])} 段,{len(lj['cues'])} 个音效点,时长 {total:.2f}s")
if "--skip-bed" not in sys.argv:
subprocess.run([sys.executable, str(CINE / "audio/tools/bed.py"), str(B / "lesson.json"), str(B)], check=True)
n = int(round(total * SR))
fit = lambda x: np.pad(x, ((0, max(0, n - len(x))), (0, 0)))[:n]
bed, sfx = fit(read_wav(B / "bed.wav")), fit(read_wav(B / "sfx.wav"))
narr = np.zeros((n, 2))
nv = read_wav(B / "narration.wav")
i0 = int(round(cold * SR))
narr[i0:i0 + len(nv)] = nv[: n - i0]
env = envelope(narr)
duck = 1.0 - 0.55 * env # 说话时背景乐约 −7 dB
if src:
duck[: int(round((cold + 0.25) * SR))] = 1.0 # 拼接的冷开场原样
# 案例片段的原声:full = 原样放(这段旁白应当是空的、背景乐 level 0);duck = 压在旁白下面;mute = 不要
clipa = np.zeros((n, 2))
for c in info.get("clips", []):
mode = c.get("audio", "duck")
if mode == "mute":
continue
cpath = L / "assets" / "clips" / f"{c['id']}.mp4"
tmp = B / f"_clip_{c['id']}.wav"
r = subprocess.run(["ffmpeg", "-y", "-loglevel", "error", "-i", str(cpath), "-vn", "-ac", "2", "-ar", str(SR), str(tmp)])
if r.returncode or not tmp.exists():
continue
a = read_wav(tmp); tmp.unlink()
k = int(0.12 * SR)
if len(a) > 2 * k:
ramp = np.linspace(0, 1, k)[:, None]; a[:k] *= ramp; a[-k:] *= ramp[::-1]
g = 10 ** (c.get("gain", -2.0 if mode == "full" else -15.0) / 20)
j0 = int(round(c["at"] * SR)); j1 = min(n, j0 + len(a))
if j1 > j0:
clipa[j0:j1] += a[: j1 - j0] * g
# 整段配乐插入(例如导学开场的高光快剪用曲库里的一首):{src 绝对路径, from 源内起点, at, dur, gain dB, fade_in, fade_out};有旁白时同样被压低
for m in info.get("music", []):
tmp = B / "_music.wav"
subprocess.run(["ffmpeg", "-y", "-loglevel", "error", "-ss", f"{float(m.get('from', 0)):.3f}", "-t", f"{float(m['dur']):.3f}", "-i", str(m["src"]),
"-vn", "-ac", "2", "-ar", str(SR), str(tmp)], check=True)
a = read_wav(tmp); tmp.unlink()
ki, ko = int(float(m.get("fade_in", 0.03)) * SR), int(float(m.get("fade_out", 1.2)) * SR)
if ki: a[:ki] *= np.linspace(0, 1, ki)[:, None]
if ko and len(a) > ko: a[-ko:] *= np.linspace(1, 0, ko)[:, None]
j0 = int(round(float(m["at"]) * SR)); j1 = min(n, j0 + len(a))
if j1 > j0:
if m.get("gain_curve"): # 分段音量:[[片中秒, dB], …] 线性插值,比如冷开场满音量、旁白段压低
pts = np.array(m["gain_curve"], dtype=np.float64)
tt = np.arange(j0, j1) / SR
g = 10 ** (np.interp(tt, pts[:, 0], pts[:, 1]) / 20)
else:
g = np.full(j1 - j0, 10 ** (float(m.get("gain", -4.0)) / 20))
clipa[j0:j1] += a[: j1 - j0] * (g * duck[j0:j1])[:, None]
mix = narr * 1.0 + bed * duck[:, None] + sfx * 0.9 + clipa
# 响度:按有旁白的部分对齐到约 −16 dBFS RMS
speech = env > 0.5
rms = np.sqrt(np.mean(mix[speech] ** 2)) if speech.any() else np.sqrt(np.mean(mix ** 2))
mix *= 10 ** (-16 / 20) / max(rms, 1e-9)
mix = limiter(mix)
write_wav(B / "final_mix.wav", mix)
print(f"混音 → {B / 'final_mix.wav'} 峰值 {20 * np.log10(np.max(np.abs(mix)) + 1e-9):.1f} dBFS")
if "--audio-only" in sys.argv:
return
# 画面:有 COLD_SRC 时先拼那段现成画面,再接本课渲染;然后把案例片段逐段叠到各自的框里
fps = 60
out = B / f"lesson{lesson}.mp4"
ins, fc = [], []
if src:
a0, na = round(src["t0"] * fps), round(cold * fps)
ins += ["-i", str(ROOT / src["video"]), "-i", str(B / "picture.mp4")]
fc.append(f"[0:v]trim=start_frame={a0}:end_frame={a0 + na},setpts=PTS-STARTPTS[ca];[1:v]setpts=PTS-STARTPTS[cb];[ca][cb]concat=n=2:v=1:a=0[base]")
k0 = 2
else:
ins += ["-i", str(B / "picture.mp4")]
fc.append("[0:v]setpts=PTS-STARTPTS[base]")
k0 = 1
last = "base"
for i, c in enumerate(info.get("clips", [])):
ins += ["-i", str(L / "assets" / "clips" / f"{c['id']}.mp4")]
k, at, d = k0 + i, float(c["at"]), float(c["dur"])
fd = float(c.get("fade", 0.15))
fades = f",fade=t=in:st={at:.4f}:d={fd:.3f}:alpha=1,fade=t=out:st={at + d - fd:.4f}:d={fd:.3f}:alpha=1" if fd > 0 else ""
fc.append(f"[{k}:v]setpts=PTS-STARTPTS+{at:.4f}/TB,format=yuva420p{fades}[c{i}]")
fc.append(f"[{last}][c{i}]overlay={int(c['x'])}:{int(c['y'])}:enable='between(t,{at:.4f},{at + d:.4f})':eof_action=pass[v{i}]")
last = f"v{i}"
ins += ["-i", str(B / "final_mix.wav")]
subprocess.run(["ffmpeg", "-y", "-loglevel", "error", *ins, "-filter_complex", ";".join(fc), "-map", f"[{last}]", "-map", f"{len(ins) // 2 - 1}:a",
"-c:v", "libx264", "-preset", "slow", "-crf", "20", "-pix_fmt", "yuv420p", "-r", str(fps), "-c:a", "aac", "-b:a", "192k", "-shortest", "-movflags", "+faststart", str(out)], check=True)
print(f"成片 → {out}")
if __name__ == "__main__":
main()
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: 194k | 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: 194k | 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: 194k | Updated | Claude CodeCodex | Verified | Install |