Communitygithub.com

review-ugc-render

UGC 影片發布前的強制審查關卡:使用 Whisper 轉寫音訊並比對核准腳本,阻擋因語音錯誤而發布有問題的影片。

review-ugc-render 是什麼?

這是一個用於 UGC 影片工作流程的強制性品質檢查工具,專門在 `set_final_render` 之前執行。它從完成的渲染影片中提取音訊,使用 Whisper(或 OpenAI API)進行語音轉寫,並將轉寫結果與使用者事先核准的腳本進行逐詞比對(word-diff)。如果發現任何詞彙被錯誤發音(例如將「human-vetted」說成「human witted」)、遺漏核准的短語,或者音訊完全靜音,此工具會以非零退出碼終止,阻止發布流程。它支援 `remix-ugc-*-from-sample` 和 `create-ugc-*-video-from-refs` 等食譜,並可透過環境變數 `OPENAI_API_KEY` 或本機 `whisper` CLI 運行。其設計目標是在自動化發布管線中提供一個可靠、確定性高的檢查點,避免因語音合成錯誤而導致錯誤字幕或品牌名稱被誤植。

相容平台~Claude Code✓Codex CLI~Cursor
npx skills add https://github.com/gooseworks-ai/goose-skills/tree/HEAD/skills/ads/packs/ugc-video-formats/review-ugc-render

在你喜歡的 AI 中提問

開啟一個已預先載入此 Agent Skill 的新對話。

說明文件

review-ugc-render 是做什麼的?

每個 UGC 影片食譜在發布前必須通過的品質控制關卡。這不是靠肉眼觀看 /watch,而是一個確定性的轉寫比對工具,若發現缺陷則以非零退出碼終止,讓食譜能夠硬性停止 set_final_render。

為什麼需要這個工具

Seedance 會原生生成音訊。但有時候它會唸錯單字——核准的台詞 human-vetted 可能被唸成 human witted;其他已知案例包括 Hume→Hune、Alitu→al-too。這個缺陷存在於渲染影片的音訊中,因此肉眼觀看 /watch(「對話跟腳本相符嗎?」)可能會漏掉,而後續的字幕處理則會將錯誤的詞彙一字不漏地寫進字幕。沒有人將實際說出的音訊與使用者核准的腳本進行比對。

這個工具正是用來做這件事,它會以確定性方式執行,並在比對失敗時拒絕發布。

何時執行

  • 強制性:在每個 remix-ugc-*-from-sample 和 create-ugc-*-video-from-refs 食譜的品質控制階段中,在主渲染完成之後且在 set_final_render 之前執行。
  • 每次修正或重新渲染後都要重新執行,直到通過為止。

合約

在渲染之前,將核准的台詞(逐字陳述,不含節奏註解)寫入 working/approved-script.txt。然後在渲染完成後,執行:

python3 <pack>/review-ugc-render/scripts/review_render.py \
  --video working/final.mp4 \
  --script-file working/approved-script.txt \
  --json working/review-verdict.json
  • exit 0 → 通過 — 繼續發布(set_final_render)。
  • exit 2 → 失敗 — 不要發布。閱讀報告,修正問題,重新執行。
  • exit 3 → 錯誤 — 檢查無法執行(詳見下文);修正環境問題,不要盲目發布。

語音轉寫後端(優先順序):OPENAI_API_KEY(支援 OPENAI_BASE_URL,因此當設定後會路由到 gooseworks Whisper 代理)→ 本機 whisper CLI。ffmpeg 必須在 PATH 中。

失敗的意義及修正方法

報告行根本原因修正方法
[high] said [witted] where script has [vetted]Seedance 在生成的音訊中唸錯單字重新生成一個新的種子。 如果是品牌/自創詞彙,請在 SPOKEN LINE 中用語音拼寫(例如 Ali-too,不要使用 (發音為...) 的括號——Seedance 會把括號內容也唸出來)。請參閱 create-video-seedance-2-fal 的失敗模式。
[medium] dropped [...] / 相似度過低Seedance 遺漏了核准的短語重新生成;如果只是結尾詞,可以使用 stitch_replacement.py 進行局部修復。
⚠ audio is effectively silent錯誤的渲染 / 後製中音軌遺失重新渲染或檢查混流;絕不要發布一個靜音版本。
ERROR: no transcription backend沒有 OPENAI_API_KEY 也沒有本機 whisper設定金鑰(OPENAI_BASE_URL)或安裝 whisper,然後重新執行。

--expect-music 僅供參考(如果缺少背景音樂則發出警告);它本身不會導致關卡失敗。

輸入參數

  • --video PATH(必要)— 渲染完成的主 MP4 檔案。
  • --script-file PATH 或 --script "text"— 核准的台詞腳本。兩者都省略僅適用於真正無腳本的片段(此時會跳過漂移檢查,關卡僅供參考)。
  • --min-ratio FLOAT(預設值 0.90)— 轉寫與腳本的詞彙相似度門檻。
  • --json PATH — 將機器判定結果寫入此檔案,供應用程式的審查面板使用。

測試

python3 tests/test_review_render.py    # 純邏輯判定;不需音訊或網路

涵蓋標準的 vetted→witted 唸錯、品牌名稱唸錯、遺漏結尾、無害的填充詞,以及無腳本的參考檢查路徑。

與 content-goose 審查引擎的關係

這是更完整的 coworkers/video/molecules/review/review-loop(18 軸評估標準)中已發布的單檔案、檢查點式的子集。在這裡我們強制執行唯一能在發布時捕捉音訊與腳本缺陷的軸線(review-transcript-integrity / brand_text_accuracy)。更深度的多軸審查仍保留在 content-goose 研究室中。

Individual skills in this repo

This repo contains 10 individual skills — each has its own dedicated page.

gooseworks-ai/goose-skills

Burned-in captions for a finished vertical video, three kinds. transcribe.py gets word timings from the video's own audio through the GooseWorks proxy (fal Whisper, bills the Ads agent, cents); captions.py burns one to three words at a time with Pillow + ffmpeg (no libass needed), either pinned to a split-screen seam (plate 25% above / 75% below) or at a fixed height, in a plate, outline or one-word serif style, with an optional red hook card; plates.py burns per-beat caption blocks (black, one union silhouette, placed in the emptiest band) for formats with no voice. The last caption (the CTA) holds to the final frame. Use as the last step of any video ad.

gooseworks-ai/goose-skills

Assemble a cartoon / animated / hand-crafted music-video ad from a config — a sung song carries the whole narrative while N per-bar i2v clips (one recurring animated character, one look pack) are each cut to their BAR window from librosa beat-tracking and hard-concatenated on the bar, VEED-whisper white bold-sans captions in the BOTTOM third (Alignment 2, above the logo bug, no pill) burned from the song's word timings re-spelled against the locked lyrics, a persistent brand logo bug held over the body (suppressed on the end card), and closed on a solid-brand-color PIL end card with the song still playing under it — never AI-rendered text. This is the FREE deterministic assembly stage (cut-to-bar + hard concat + logo bug + captions + end card + song mux); the song, character, keyframes, and clips come from create-music-elevenlabs / create-image-fal / create-video-fal. Use for the cartoon-music-video format.

gooseworks-ai/goose-skills

Assemble a cinematic live-action-style music-video ad from a config — an original sung anthem carries the whole narrative while N 35mm-film-look i2v clips are each cut to their lyric window and hard-concatenated on the beat as a 3-act arc, the anthem muxed at loudnorm I=-14, cinematic lower-third serif captions built from the song's OWN word timings (never Whisper) with the hook line landing on the chorus drop, and closed on a brand end card composited from the real asset — never AI-rendered text. This is the FREE deterministic assembly stage (cut-to-window + hard concat + anthem mux + captions + end card); the anthem, keyframes, and clips come from create-music-elevenlabs / create-image-gpt-image-fal / create-video-fal. Use for the cinematic-music-video format.

gooseworks-ai/goose-skills

Assemble an editorial-motion podcast-clip ad from a config — a real clipped podcast MP3 carries the narrative while N flat limited-palette editorial-illustration keyframes (one look pack) are animated NOT by generative i2v but by DETERMINISTIC ffmpeg ken-burns (zoompan) + hard cuts (no crossfades, which expose geometric drift), each beat snapped to its spoken line, the real audio muxed, Whisper-driven captions burned only mid-sentence, and closed on a PIL brand end card — never AI-rendered text. This is the FREE deterministic assembly stage (ffmpeg ken-burns + hard concat + audio mux + captions + end card); the real audio is clipped from source and the keyframes come from create-image-fal. Use for the editorial-motion-podcast format.

gooseworks-ai/goose-skills

Analyze the message match between your ads and landing pages. Checks if the promise in the ad copy carries through to the landing page headline, body, and CTA. Flags disconnects that kill conversion rates. Works with Google, Meta, and LinkedIn ads.

gooseworks-ai/goose-skills

Generate a single 4-15s vertical video clip with ByteDance Seedance 2.0 reference-to-video via fal.ai. Multi-image reference (avatar + product + setting), native lip-synced VO + ambient audio (generate-audio on by default), internal multi-cut handling within one render. Routes through the GooseWorks FAL proxy (bills the Ads agent). The default clip atom for AI-creator UGC ads built on the NB2 + Seedance architecture. Validated on beauty-by-earth/video-01.

gooseworks-ai/goose-skills

Render pixel-accurate Apple Notes (iPhone, light mode) screenshot mockups from a JSON note spec. Outputs HTML + PNG at the iPhone 16/15 Pro native 1180×2556. Supports paragraphs, images, checklists, dividers, autocorrect underline, smart quotes, and an optional iOS keyboard chrome overlay used by the parent video-ad molecule.

gooseworks-ai/goose-skills

Extract competitor and customer intelligence from any company's landing page HTML. Discovers tech stack, analytics tools, ad pixels, customer logos, SEO metadata, CTAs, hidden elements, and more. No API keys required.

gooseworks-ai/goose-skills

Find speakers, hosts, and guest profiles at conferences and events on Luma. Two modes - free direct scrape for hosts, or Apify-powered search for full guest profiles with LinkedIn/Twitter/bio.

gooseworks-ai/goose-skills

Find Twitter/X influencers to promote a product or brand. Use when asked to find influencers, discover Twitter accounts for partnerships, identify creators in a niche, or build an influencer outreach list.

相關技能