Vibe Filming · Codex + HolyCrab CLI

HOLYCRAB / AI ULTRA FESTIVAL

Vibe FilmingAI Video from Idea to Screen

AI 影片創作：從構想到成片

See what is possible. Learn the workflow. Make your first video.

Bruce Huang · AI Ultra Festival · Hong Kong · 13 September 2026

VIBE FILMING / AI ULTRA FESTIVAL

Why make an AI video?

把產品、人物與想法，變成觀眾看得懂的故事

01Show a product in use, not just on a slide.

02Reach customers through a presenter and their language.

03Test a creative idea before a larger production.

VIBE FILMING / AI ULTRA FESTIVAL

World-class ambition.
A workflow you can start.

先看高水準作品，再拆解成可操作的一小步

01Watch the story, performance and visual consistency.

02Look beyond a striking shot: assets, review and editing matter.

03Start small; build toward that level of craft.

Official Hunan TV trailer · 1080p · 30-second excerpt
Playback / 影片播放

CASE STUDY / AI LONG-FORM PRODUCTION

AI drama on television

《后西游记》 · 成本、周期与市场反应

RMB 900k+ / episodeAbout 40 minutes; director claim, not audited

~3 months / 15 episodes2 Sep interview; not all 30 episodes

Stock +44.03% / 2 sessions28 Aug–1 Sep; media index +4.93%

Mango Excellent 300413.SZ · daily · RMB

Event-window move; not proven drama-driven gains.

00–04分钟开场内，播放30秒高清节选，说明AI影像已进入制作与播出流程。来源：湖南卫视官方B站账号，https://www.bilibili.com/video/BV181tP6KENt/ ，标题《【后西游记】首部AIGC长剧〈后西游记〉上星黄金档》。完整原预告139.584秒，本课使用00:00–00:30，视频流直接复制1920×1080源编码，不做降分辨率或视频重压缩，音频转为AAC96k并做极短淡入淡出。保留官方台标。替代此前合作伙伴PPT来源的540p压缩节选。这是第三方行业案例，不暗示HolyCrab制作该剧，也不推断制作模型、成本或收视率。右侧数据取自Bruce在Codex任务《分析芒果TV AI影片商业数据》中完成的2026-09-09研究，本轮复核新京报导演专访和交易记录。这里是剧集，成本按单集而非整部电影：导演称每集九十多万元、每集约40分钟，非审计决算。导演9月2日称约三个月完成15集制作量，首批已播5集；平台另有首5集立项到上线约4个月的口径，两者范围和起点不同。不能称三个月完成全30集。公开资料未证明该剧使用Codex制作，已确认的是Bruce的研究在Codex完成。制作专访：https://www.bjnews.com.cn/detail/1788401162168184.html 不同周期口径：https://www.eeo.com.cn/2026/0908/1028446.shtml 市场：芒果超媒300413.SZ，8月28日收盘14.15，9月1日20.38，累计44.03%，同期中证传媒4.93%。K线按2026-08-20至09-08的14个交易日OHLC绘制，沿用原研究窗口，不是当前实时行情。8月31日首次涨停早于当晚首播，不能将全部涨幅因果归于该剧；市值增量不等于公司现金或项目利润。交易记录：https://finance.eastmoney.com/a/202609013861493595.html 历史行情：https://web.ifzq.gtimg.cn/appstock/app/fqkline/get?param=sz300413,day,2026-08-20,2026-09-08,320,qfq 搜狐交叉核对：https://q.stock.sohu.com/hisHq?code=cn_300413&start=20260820&end=20260908&stat=1&order=A&period=d Mobile handout: https://brucehuang.ai/talks/ai-ultra-video-workshop-2026-09/cli-review/handout/slide-04.html

HOLYCRAB / PRODUCTION

A production system,
not just a prompt

從劇本、角色、場景到鏡頭，保留每一次製作決策

HolyCrab production workflow · actual system screenshot

HolyCrab H3 music video · 30-second excerpt
Playback / 影片播放

VIBE FILMING / CASE STUDY

Performance meets rhythm

表演、節奏與風格：HolyCrab 音樂影片案例

开场可选短放15–30秒，Bruce最终选示范。带声音播放前30秒，再用30秒拆解创作。观察两位表演者、黑白服装与红色腰带、统一的图形色彩，以及近景、中景和全身景别之间的剪辑。HolyCrab 帖文注明使用 MiniMax H3，但未公开全部制作步骤，也未声称整片由一次生成完成。可讲解表演、重复使用的视觉素材和随音乐剪辑的作用，不要编造原始提示词。完整视频61.416秒，见学员手册。来源：https://x.com/holycrabAI/status/2098092554093039735 。这是 HolyCrab 发布的模型示例，不应称为艺人官方委托作品。教学设计待补充：老师示范____；学员动作____；完成标准____。 Mobile handout: https://brucehuang.ai/talks/ai-ultra-video-workshop-2026-09/cli-review/handout/slide-06.html

Bruce introducing himself in Japanese · 34 seconds
Playback / 影片播放

VIBE FILMING / CASE STUDY

A real introduction.
In Japanese.

真實商務用途：Bruce 在日本客戶現場的自我介紹

VIBE FILMING / WORKSHOP DEMO

One person. One product.
A finished 30-second film.

先看成品：人物、商品與兩段銜接

Seedance 2.0 Fast 15s + 15s · English audio Advanced demo. Our first exercise is one simple 10s clip.

Open video →

VIBE FILMING / AI ULTRA FESTIVAL

What makes this repeatable?

看完成品，再看原始 Prompt 與參考照片

01Person asset: who appears. Product asset: what it looks like.

02Prompt: what happens, in what style, with what voice.

03Last frame: where the next shot begins.

VIBE FILMING / AI ULTRA FESTIVAL

Why use a coding agent?

你決定創意；Codex 負責執行與保留製作記錄

01Turn a brief into prompts and API requests.

02Check costs, track tasks and download the result.

03Reuse assets and continue a project without starting over.

VIBE FILMING / AI ULTRA FESTIVAL

Optional: see the canvas

同事示範 5–10 分鐘，學員只需觀看

01Prepared assets → connected generation nodes → result.

02A colleague shows one change in a prepared project.

03Then Bruce returns to the step-by-step workflow.

VIBE FILMING / AI ULTRA FESTIVAL

First understand.
Then make.

前半場解說，後半場循序動手

0100–30: why video, cases, assets and optional canvas.

0230–40: website first. 40–50: one video with Codex.

0350–55: optional frame continuation. 55–60: questions.

PROMPT 1 / 3 · 10 minutes

Make your first video on the web

第一步：先在網頁完成一條，不用安裝或配置 Key

Sign in to HolyCrab and check your credits.
Open text-to-video and paste this prompt.
Select 2.0 Fast / 480p / 10s / 16:9 + audio; review the quote, then generate.

Done: play it, hear the voice, and download it. Queued is not finished.

Scan to open prompt 1 — SCAN · COPY · RUN
完整步骤 + 可复制 Prompt

Create a 10-second, 16:9 realistic social video with native audio. One original virtual adult creator in a plain navy shirt sits at a light-oak desk. Soft daylight, warm neutral background. This is a fictional person, not a celebrity or a real person's likeness. A generic unbranded smartphone is already in the creator's hands, below the face. In one steady medium shot, the creator raises the phone slightly toward the camera and smiles. Keep the face, both hands and the entire phone visible. One person, one phone, one simple movement. No unfolding, screen interaction, scene changes, captions or music. Between seconds 1 and 7, say in natural, upbeat English: "One idea. One prompt. Let's make it real." Quiet room tone. Hold still with the phone clearly visible during the final two seconds. No additional speech. Mobile handout: https://brucehuang.ai/talks/ai-ultra-video-workshop-2026-09/cli-review/prompts/1.html

VIBE FILMING / AI ULTRA FESTIVAL

One playable result
is the first milestone

完成第一條，再增加難度

01Queued? Keep the task. Do not click Generate again.

02Completed? Watch the person and phone; listen to the line.

03Download it. Change only one thing on the next attempt.

VIBE FILMING / AI ULTRA FESTIVAL

One project.
One private chat.

不用先學 Terminal；等待 Agent 提示，再貼測試 Key

01Open a vibe-filming project in Codex.

02Create a limited test key in your HolyCrab account.

03Paste it into the same private Codex chat only when asked.

PROMPT 2 / 3 · 10 minutes · Codex-ready learners

Let Codex make one video

第二步：同一專案、同一私人聊天視窗

Open a vibe-filming project in Codex.
Paste the instructions; provide a new limited test key only when asked.
Approve the quote. Generate and save one A.mp4.

Done: a playable A.mp4 with audio, its prompt, task ID and actual cost.

Scan to open prompt 2 — SCAN · COPY · RUN
完整步骤 + 可复制 Prompt

Help me generate one video with HolyCrab. I am a beginner. Give me only one action to take at a time. 1. Check local command access and whether HolyCrab CLI is available. Follow the current official instructions at https://holycrab.ai/cli/ . If installed but not found, inspect the installation path and use the verified executable path. Do not ask me to edit PATH, disable security, or bypass company policies. Explain any permission I must approve in one sentence. If blocked, explain why and let me continue on the website. 2. Check for an existing usable local API key. If missing, pause and ask me to paste a newly created, limited test key into this same private Codex chat. I will paste it here; do not require Terminal. Do not echo it. Use the official configuration method; never put it in project files, prompts, logs, or Git. Do not claim the key has disappeared from chat history. 3. Read current model capabilities. Use Seedance 2.0 Fast, 10 seconds, 480p, 16:9, text-to-video with native audio and the prompt below. If unsupported, explain first; do not silently switch models. Show the actual quote, account balance and cumulative exercise cost. Wait for my explicit confirmation, then submit exactly once. Approval does not include automatic retries. Recover the existing task after a disconnect rather than submitting again. 4. Download output/A.mp4. Save the prompt and task ID. Check duration and audio, and tell me where to play it. Generate only this one clip: do not extract frames, continue, or assemble yet. Clearly say if the task is still queued. VIDEO PROMPT: Create a 10-second, 16:9 realistic social video with native audio. One original virtual adult creator in a plain navy shirt sits at a light-oak desk. Soft daylight, warm neutral background. This is a fictional person, not a celebrity or a real person's likeness. A generic unbranded smartphone is already in the creator's hands, below the face. In one steady medium shot, the creator raises the phone slightly toward the camera and smiles. Keep the face, both hands and the entire phone visible. One person, one phone, one simple movement. No unfolding, screen interaction, scene changes, captions or music. Between seconds 1 and 7, say in natural, upbeat English: "One idea. One prompt. Let's make it real." Quiet room tone. Hold still with the phone clearly visible during the final two seconds. No additional speech. Mobile handout: https://brucehuang.ai/talks/ai-ultra-video-workshop-2026-09/cli-review/prompts/2.html

VIBE FILMING / AI ULTRA FESTIVAL

Stuck? Diagnose.
Do not regenerate.

安裝、Key、排隊：一次只排查一個問題

01Command not found: ask the agent to check its install path.

02Invalid key: check the full key, limit and revocation status.

03No result: retrieve the existing task ID before retrying.

PROMPT 3 / 3 · 5-minute walkthrough · finish after rendering

Continue from the last frame

第三步：進階選做，先有成片才做續接

Open the completed A.mp4 and inspect its ending.
Extract the actual last frame; upload it and wait for ready.
Approve a fresh quote, generate B, then join and inspect.

Done: keep A, bridge.png, B and final.mp4. Watch and listen across the join.

Scan to open prompt 3 — SCAN · COPY · RUN
完整步骤 + 可复制 Prompt

Continue in the same project. This is optional advanced work. Do not submit anything without a playable A.mp4. Reuse output/A.mp4 from Stage 2, or ask for the downloaded Stage 1 video. Do not regenerate A just to start this stage. 1. Open A and inspect the face, hands and phone at the ending. After I accept it, extract its actual last frame to assets/bridge.png and show it to me. If an earlier frame is needed, trim A to exactly that frame and record why; never silently replace the ending. 2. This exercise defaults to an original virtual person. Upload the frame with the current official CLI and wait for ready. Record its source task. If A contains a real person, keep the derived frame in that person's original verified group. Stop if verification fails; never bypass it with an ordinary upload. 3. Read the current Seedance 2.0 Fast first-frame parameters. Use the bridge as firstFrameAssetId for a 10-second, 480p B with native audio. Use adaptive aspect ratio as required by the current frames mode; do not mix in a reference-image array. Use the prompt below. Show the quote, balance and cumulative exercise cost. Wait for explicit confirmation and submit once. No automatic retry or top-up. 4. Download output/B.mp4 and check that it starts from the supplied frame. Assemble A then B locally into output/final.mp4, retaining originals, audio and both task IDs. Do not overwrite A or B. Let me watch and listen to two seconds on either side of the join: person, phone, movement and volume. A frame helps visual continuity but cannot guarantee the same voice. Report missing native audio as a problem, not a completed silent video. VIDEO PROMPT: Continue the supplied first frame as a 10-second video with native audio. Preserve the same virtual creator, face, navy shirt, hands, unbranded phone, desk, lighting and phone orientation. The phone is already raised; do not repeat that action. Seconds 0–3: hold the opening composition. Seconds 3–8: one slow, straight camera push toward the phone, keeping the entire phone and supporting fingers in view. Seconds 8–10: hold still. Do not change the phone, unfold it, cut to another angle, or add captions or music. Use a warm English off-screen voiceover between seconds 1 and 7: "Now take it one step further. Keep creating." The visible mouth remains relaxed and closed. Quiet room tone. Keep the final two seconds free of speech. Match the previous clip's vocal character if possible; the visual frame does not guarantee the same voice. Mobile handout: https://brucehuang.ai/talks/ai-ultra-video-workshop-2026-09/cli-review/prompts/3.html

VIBE FILMING / AI ULTRA FESTIVAL

Keep the work.
Build on it.

保留影片、Prompt、素材與任務 ID

01Save A, the bridge frame, B and the assembled film.

02Check picture and voice across the join.

03Website and API use the same credit pricing; quote each next task.

HOLYCRAB.AI @ AIULTRAFESTIVAL

Keep creating.
We are here to help.

有问题带着步骤和错误信息来问

Join the HolyCrab workshop WhatsApp group.
Share your stage, question and a redacted error if needed.
Continue with your existing credits and saved project.

Website and API use the same credit pricing.

Vibe FilmingAI Video from Idea to Screen

Why make an AI video?

World-class ambition.A workflow you can start.

AI drama on television

A production system,not just a prompt

Performance meets rhythm

A real introduction.In Japanese.

One person. One product.A finished 30-second film.

What makes this repeatable?

Why use a coding agent?

Optional: see the canvas

First understand.Then make.

Make your first video on the web

One playable resultis the first milestone

One project.One private chat.

Let Codex make one video

Stuck? Diagnose.Do not regenerate.

Continue from the last frame

Keep the work.Build on it.

Keep creating.We are here to help.

World-class ambition.
A workflow you can start.

A production system,
not just a prompt

A real introduction.
In Japanese.

One person. One product.
A finished 30-second film.

First understand.
Then make.

One playable result
is the first milestone

One project.
One private chat.

Stuck? Diagnose.
Do not regenerate.

Keep the work.
Build on it.

Keep creating.
We are here to help.