DEV Community: 张洲诚（Zack.ZHANG）

I Generated 50 Sales Reports with One CLI Command — Here's How

张洲诚（Zack.ZHANG） — Mon, 20 Jul 2026 10:20:58 +0000

Mid-year review season. You know the drill.

My team owed leadership a polished Q2 deck, plus every regional sales lead needed their own weekly report. Total count: about 50 reports. Same structure, different data.

The first deck I did the honest way — PowerPoint, three full days. Not because I can't use templates, but because the "raw data → insight → visual story" pipeline is brutally long. Each chart means copying numbers from Excel, fixing formatting, writing analysis. One slide averages 40 minutes. Twenty slides = two working days gone.

Then I looked at the remaining 49. That math doesn't work.

The GUI tool problem

I tried the obvious solutions:

Gamma ($20/mo): Fast AI generation, but no batch mode — each deck requires manual input and manual export
Beautiful.ai ($12/mo): Great design, but can't read my raw Excel data
WPS AI (~$4/mo): Can read data, inconsistent quality, no batch

The common issue: GUI tools don't support scripting. One deck saves time. Fifty decks means doing it fifty times by hand.

The CLI approach

I found Bailian CLI on GitHub — a command-line tool that calls large models directly from the terminal, reads files as context, and can be looped in a script. Here's what my workflow ended up looking like.

Setup

pip install dashscope-cli
bl configure --api-key YOUR_API_KEY

Free API key from the Bailian console — new users get complimentary credits.

Step 1: Generate an outline

The worst part of building a deck isn't layout — it's "what goes on slide one."

bl chat --model qwen-plus \
  --prompt "Generate a PPT outline for Q2 sales performance review, covering:
1. Performance overview (revenue, profit, YoY/QoQ)
2. Regional comparison (East/South/North/Southwest)
3. Growth attribution (which product lines drove growth)
4. Risks and anomalies
5. Next quarter action plan
Include 2-3 key points and visualization suggestions per module."

30 seconds → structured outline with chart recommendations (stacked bars for revenue composition, heatmaps for growth distribution). Previously this took 30 minutes on a whiteboard.

Step 2: Data-driven content

The key differentiator — feeding in the actual data file:

bl chat --model qwen-max \
  --file sales_q2.xlsx \
  --prompt "Based on this Q2 sales data, generate executive-facing PPT content:
1. Extract 3 core conclusions (with data support)
2. Identify 2-3 anomalies or notable trends
3. Recommend next-quarter actions
4. Write out each slide's title and bullet points
Emphasize YoY growth trends and regional differences."

It reads the actual numbers, analyzes them, and outputs specific conclusions — not template filler. For example: "East China Q2 revenue grew 23% YoY, driven by enterprise SaaS, while North China saw -5% negative growth — renewal rate decline needs attention."

GUI tools can't do this. Gamma and Beautiful.ai only work with text you type in; they won't read your source files.

One complete executive deck: ~1 hour (including manual tweaks and layout). Down from three days.

Step 3: Batch it

Same structure, different data per region. Textbook scripting scenario:

for region in east south north southwest central northeast northwest; do
  bl chat --model qwen-plus \
    --file "data/${region}_q2.xlsx" \
    --prompt "Based on ${region} region Q2 sales data, generate weekly report content:
1. Regional performance summary
2. Top 5 customer contribution analysis
3. YoY and QoQ changes
4. Items requiring HQ support
Output as structured PPT page content." \
    --output "output/${region}_weekly_report.md"
done

Seven regions, 8–10 minutes total. Structured Markdown output → batch import into deck template.

Approach	Time for 50 reports	Data accuracy	Repeatability
Fully manual	~25 hours (30 min each)	Copy-paste errors likely	Redo from scratch
GUI tools	~12 hours (15 min each)	Manual data entry	Semi-automated
CLI script	~10 min + 1 hr layout	Reads source data directly	Fully repeatable

Next quarter? Swap the data files, run the same script. Zero rewrite.

Beyond decks

The same pattern works for anything "given materials → structured document":

# Business plan
bl chat --model qwen-max \
  --prompt "Draft a business plan outline for an enterprise analytics SaaS..."

# Data analysis report
bl chat --model qwen-max \
  --file user_behavior_june.csv \
  --prompt "Write a product data report: DAU, retention, conversion, anomalies..."

# Excel formulas
bl chat --model qwen-plus \
  --prompt "Write an XLOOKUP formula that..."

Meeting minutes, contract drafts, quarterly summaries — same logic.

Cost

Tool	Monthly cost	Batch support	File input	Best for
Gamma	$20/mo	No	No	One-off polished decks
Beautiful.ai	$12/mo	No	No	Design-first
WPS AI	~$4/mo	Limited	Partial	WPS ecosystem
Bailian CLI	Pay-per-token	Native scripting	Yes	Batch + data-driven

Using qwen-plus: one deck ~$0.05–0.07, 50 regional reports ~$2–4, daily misc ~$1–2/month.

Honest limits

CLI has a learning curve — not as plug-and-play as GUI tools
Output is content structure; final .pptx layout still needs human work
Quality depends on prompt quality
Complex visualizations still need manual handling
If you only make one deck occasionally and care about design polish, Gamma/Beautiful.ai are better

The CLI's edge: batch processing + data-driven generation + programmability.

If mid-year reporting season is crushing you too, consider a scripting approach. Core idea: hand the repetitive content generation to the machine, keep the aesthetic judgment and business decisions for yourself.

One CLI for Every Stage of Writing a Paper — Proposal to Defense

张洲诚（Zack.ZHANG） — Fri, 17 Jul 2026 11:13:51 +0000

The stress that hits during thesis season

The paper isn't even done, and you're already drowning in tools: plagiarism check on one site, AI-content detection on another, a paraphraser here, a translator there, references formatted by hand, and a LaTeX error that eats half an hour of googling. A huge chunk of your energy goes into copy-pasting between a dozen tools.

Then there's the "AI smell" problem. In 2025 China's Ministry of Education folded AIGC detection into the quality framework for degree theses, with a clear line: AI assistance is allowed, AI ghostwriting is not. Many schools now treat the AIGC score as a hard gate alongside plagiarism. The catch: detectors disagree wildly — the same paper can score 70% on one service and 5% on another, and even acknowledged classics get flagged.

I'm no academic heavyweight — I just got tired of tool-hopping and wondered: could one tool absorb all the mechanical, repetitive parts of paper writing? Running it through Bailian's CLI (bl), it turned out nine stages, proposal to defense, really can be covered by a single command.

⚠️ Up front: everything below is writing assistance — gathering material, structuring ideas, polishing phrasing, fixing formatting. It does not write your paper, and it is not a way to beat the detector. The ideas, data, and conclusions are yours to research, verify, and stand behind.

Nine stages, one capability

① Proposal  → ② Lit review  → ③ Draft skeleton
④ Polish    → ⑤ Paraphrase  → ⑥ De-AI phrasing
⑦ Translate → ⑧ LaTeX check → ⑨ Defense slides

Under the hood they're all bl text chat — one long-form chat command. Swap the prompt, do a different job. No separate site or subscription per stage.

Install (two lines)

npm install -g bailian-cli
bl auth login

You'll need Node.js. New users get free credits.

Default model is qwen3.7-max.

A few representative stages

Proposal — the problem often isn't that you can't write, it's that you haven't figured out what to write. Let it give a solid frame, then fill it with your research:

bl text chat \
  --system "You are an academic writing assistant. Produce standard, verifiable drafts for the author to research and revise." \
  --message "For the thesis 'Multimodal Machine Translation for Low-Resource Languages', draft a proposal: background, state of the art, goals, roadmap, novelty, timeline" \
  --max-tokens 6000

The frame is its work; the research is yours — verify every claim and citation against real literature. For references, use it only to format (e.g. APA 7th), never to invent citations — fabricated references are textbook misconduct.

De-AI phrasing — the headache of 2026. Students who wrote every word themselves still get flagged because the prose is too tidy:

bl text chat --message "The text below reads mechanical and templated. Rewrite it to sound more natural and personal without changing the academic meaning: <paragraph>"

To be crystal clear: this exists to improve mechanical phrasing in genuine writing so wrongly-flagged original work reads more human. It is not for laundering ghost-written text. If the content was AI-written to begin with, no rewriting makes it honest.

The remaining stages — lit review, draft skeleton, polish, paraphrase, translation, LaTeX check, defense slides — follow the same pattern; long documents go through --messages-file.

Cost

The whole flow is basically token cost — a few dollars for an entire paper, versus per-paper paraphrasing sites or the several-hundred-yuan "human de-AI" grey market. New users get free credits.

The honest limits

It won't do your research — ideas, novelty, and data are yours.
Verify every citation — models fabricate references; use it only for formatting.
Paraphrase/de-AI aren't magic — if content is ghost-written, no tool saves it.
Rules vary by school — some require declaring AI assistance; check first.

Start with whichever stage hurts most.

Start now (free credits)

This is a walkthrough of a writing-assistance tool. Every stage is framed as assistance, not ghostwriting. Follow your institution's and journal's academic-integrity and AI-use policies; originality, data accuracy, and integrity are the author's own responsibility.

Zero-to-one brand kit with one CLI tool: from logo to social posts

张洲诚（Zack.ZHANG） — Wed, 15 Jul 2026 08:33:42 +0000

AI logo and mascot generators are everywhere right now, but one question kept nagging me: these tools nail a single striking image, yet brand identity was never about one image. A logo, cups, cards, stickers, posters, social shots — a dozen pieces that only read as a "brand" when they share one consistent look. One pretty image is easy; a consistent set is the hard part.

So I ran a full test, upfront about the framing: suppose you're building a brand's visuals from zero to one — can a single CLI carry it? For a concrete target I made up an indie coffee brand, "Cloudy Coffee," and used the bl CLI (Alibaba Cloud Model Studio) to take it from logo to social images.

What came out

The logo — the Chinese text renders cleanly, which most non-Chinese models fail at:

The mascot, Cloudy — front, side, expressions, all the same character:

Setup

npm install -g bailian-cli    # needs Node.js >= 18
bl auth login --api-key sk-xxxxx

Grab a free API key from the Model Studio console.

The core idea: the "anchor" trick

A general image model is stateless — it re-interprets your prompt every time, so a set drifts in style. The fix: write the brand's visual DNA as one fixed block and paste it into every prompt, plus a fixed --seed:

Color  : warm latte beige and caramel brown, a touch of sky blue
Char   : Cloudy, a chubby cute cloud mascot with rosy cheeks
Style  : soft hand-drawn watercolor, cozy healing

Same seed + same anchors = the style stops wandering. It's like locking a picture-book character's face — except you're locking the whole brand's vibe.

Building the set

Logo:

bl image generate --model qwen-image-2.0-pro --size 1:1 --seed 88 --watermark false \
  --prompt "Minimalist coffee brand logo, a cloud merged with a coffee cup, warm latte beige and caramel with sky blue, soft watercolor, Chinese text '云朵咖啡' and 'Cloudy Coffee', white background"

A 3x3 sticker pack — nine emotions, one cloud:

VI / merchandise flat-lay:

Other assets only change --size (3:4 social, 9:16 poster) and the scene — anchors and seed stay fixed.

The honest limits

Consistency isn't 100%. Anchors + seed hold the overall style; fine details still drift.
Text can break. Short CJK titles are usually fine; long strings produce typos — generate a few and pick.
Commercial use needs due diligence. AI image copyright and font licensing vary — verify before trademarking or large-scale use.
No full VI spec. A complete brand book is still a professional job.

Try it

Install bl | Get a free API key.

The whole trick: encode your brand's color, character, and style as fixed anchors, add a fixed seed, and let the model paint everything in one consistent world. Swap in your own brand and run it.

Based on hands-on testing with Bailian CLI (bl 1.4.2). Brand case is fictional for demonstration.

I Built an AI Storybook Generator for My Kid With One CLI

张洲诚（Zack.ZHANG） — Tue, 14 Jul 2026 05:45:37 +0000

It's summer break, and I wanted to make a personalized picture book for my 5-year-old. Every tool I tried fell short the same way: "AI storybook" apps generate images, but the hero's face changes every other page — and there's no narration, no video, no way to make it an interactive, tappable book.

I ended up doing the whole thing with one command-line tool: Bailian CLI (bl). It writes the story, generates consistent illustrations, synthesizes narration, and I wrapped it into an interactive HTML book. Here's how.

Install

npm install -g bailian-cli   # needs Node.js >= 18
bl auth login --api-key sk-xxxxx

Grab a free key from the Model Studio console (free credits are enough to run all of this).

Step 1 — Write the story

bl text chat --model qwen3.7-max \
  --system "You are a children's picture-book author. Warm, simple language for a 5-year-old." \
  --message "Write a 3-act short story 'Xiaoyu and the Firefly', 2-3 sentences per act."

Step 2 — Consistent illustrations (the whole game)

The trick to keeping the same character across pages: bake the look into a fixed anchor you reuse in every prompt, plus the same --seed.

ANCHOR="A cute 5-year-old girl named Xiaoyu, round face, big eyes, short black bob hair with a red hairpin, yellow dress, white sneakers, soft watercolor picture book"

bl image generate --model qwen-image-2.0-pro --size 4:3 --seed 71 \
  --prompt "$ANCHOR, standing in a meadow at dusk holding a glowing lantern, one firefly nearby"

Only swap the scene; the hero stays put:

Same Xiaoyu across all three. qwen-image-2.0-pro even renders in-image text well — here's a 4-panel comic with speech bubbles:

Step 3 — Narration

bl speech synthesize --text "At dusk, Xiaoyu strolled across the meadow with a little lantern..." --voice <voice_id> --out voice1.mp3
# bl speech synthesize --list-voices  # to pick a voice

Step 4 — Assemble the interactive book

Drop each act's {image, text, audio} into a small HTML template with page-flip + play buttons:

~200 lines of HTML; the images and audio are all bl output. From "write me a firefly story" to a tappable e-book — without opening a single app.

Bonus: it's also a tutor

Same bl text chat, different persona:

bl text chat --system "You are an IELTS speaking examiner." --message "Score this answer on fluency, vocabulary, grammar, pronunciation and suggest fixes: 'I very like play basketball on weekend.'"

Honest limitations

Consistency isn't 100% — generate a few, pick the best.
Video (video generate) is async; poll with bl video task get.
Assembling PDF/video/HTML is a separate step (template included).
Adults should vet the content — AI is the pen, not the author.

Try it

All examples run on Bailian CLI (bl 1.4.2); illustrations, comic, and the book UI are real bl output. Model Studio pricing.

The AI Image Toolkit: One CLI for Background Removal, Portraits, Manga, and 3D

张洲诚（Zack.ZHANG） — Fri, 10 Jul 2026 10:56:32 +0000

AI image editing tools are everywhere, but each does only one thing. Remove.bg does cutouts. Miaoya does portraits. Midjourney does generation. To chain 5 needs, you switch between 5 platforms.

Bailian CLI packs 28+ image capabilities into one CLI. Here are 10 tested scenarios.

Installation

npm install -g bailian-cli
bl config set api_key YOUR_KEY

Get your key from the Model Studio console — free credits for new accounts.

6 Core Scenarios

1. Background Removal

bl image edit --input photo.jpg --prompt "Automatically detect the person and remove the background, preserving edge details including hair strands"

2. AI Object Removal

bl image edit --input cat.jpg --prompt "Remove the sofa and clutter in the background, replace with clean Japanese-style minimalist living room, keep only the orange cat"

3. Style Transfer

bl image edit --input street.jpg --prompt "Convert this real city street photo into Japanese anime style, Studio Ghibli style, bright blue sky and fluffy white clouds"

4. AI Portrait

bl image generate --prompt "Generate 5 vintage film photography shots, visible film grain, authentic lighting, retro tones, Fujifilm filter"

5. Manga Generation

bl image generate --prompt "Generate a 4-panel manga about a cat programmer's day, cute healing style"

6. 3D Figure

bl image generate --prompt "Generate Pikachu 3D figure, ABS or PVC or resin material, high detail, smooth surface"

Advanced

Case	Command	Description
Single-Image LoRA	`bl image edit`	Character consistency generation
AI Headshot	`bl image edit`	T-shirt to business suit
9-Grid Comic	`bl image generate`	Photo to comic strip
Time Machine	`bl image edit`	B&W to color

Batch Processing

for file in photos/*.jpg; do
  bl image edit --input "$file" --prompt "Remove background" --output "output/cutout_$(basename $file)"
done

Cost

Operation	CLI	Traditional
Background removal	$0.07/image	Photoshop 30 min
AI portrait	$0.07/image	$1.40/shot
Manga	$0.07/image	$70-300/strip
3D figure	$0.07/image	$70-300/figure

Limitations

Requires CLI basics 2. Text rendering errors 3. Details may mutate 4. 3D is render, not model file

Install | Free API Key

I Built a 6-Step E-Commerce Visual Pipeline Using Only CLI Commands

张洲诚（Zack.ZHANG） — Thu, 09 Jul 2026 06:09:52 +0000

The Problem

E-commerce product listing requires visual assets at scale. Each SKU needs: 5 white-background shots, 4 lifestyle scenes, 3 model photos, 2 marketing banners, and 1 showcase video.

Industry pricing (2026 China market reference): basic product photography runs $110-700/set, full shoots with models and video reach $1,400+. For 20 products, that's tens of thousands just on visuals.

As a developer, when I see repetitive work, I think about scripting. So I went looking for an automation-first solution.

Why GUI Tools Failed Me

Every AI image tool I tried only solved one step:

Tool A does background removal, not model try-on
Tool B generates great images but loses product details
Tool C is full-featured but costs $30-70/month with image caps

The real problem wasn't quality — it was the manual pipeline between tools. Download from A, upload to B, download from B, upload to C. Repeat 20 times.

The Fix: CLI = Scriptable = Batchable

I discovered Bailian CLI (bl) — it wraps image generation, image editing, and video generation into a single command-line tool.

Setup

npm install -g bailian-cli
bl config set api_key YOUR_KEY

Get a free API key — starter credits cover this entire tutorial.

My 6-Step Pipeline

Product: Vintage brown leather crossbody bag.

Step 1: White-background product shot

bl image generate --prompt "vintage brown leather crossbody bag, white background, product photography, 45-degree angle"

For example: "Generate 4 e-commerce hero images in different scene styles, covering dining, kitchen, home, and café scenarios" — just describe what you need in plain language.

Step 2: Lifestyle scene

bl image edit --image ./product-white.png --prompt "Place bag on rustic café table, warm lighting, lifestyle feel"

For example: "Generate a suitable background for the product in the image, conveying a refreshing, cool sensation"

Step 3: Virtual model try-on

bl image edit --image ./product-white.png --reference ./model-photo.png --prompt "Model wearing this crossbody bag naturally, maintain pose"

For example: "Virtual try-on of clothing onto my own body, keeping my face, body shape, and pose unchanged"

Step 4: Style variants

bl image edit --image ./scene.png --prompt "6 style variants: warm soft, cool minimal, vintage film, Japanese fresh, dark luxury, vibrant Instagram"

For example: "Generate 6 different style variants: warm soft-light, cool minimal, vintage film, fresh Japanese, dark luxury, vibrant Instagram" — exact same wording works.

Step 5: Marketing banner

bl image generate --prompt "E-commerce banner, bag centered, bold '50% OFF', gold accents, dark background" --size 1280x720

For example: "Generate a high-quality e-commerce marketing poster suitable for Taobao and Xiaohongshu platforms"

Step 6: Product video

bl video generate --image ./scene.png --prompt "Product rotating on turntable, warm lighting, bokeh" --duration 5

For example: "Generate a product showcase video suitable for Douyin/TikTok and Kuaishou platforms"

Batch Mode

for product in crossbody-bag tote backpack wallet belt; do
  bl image generate \
    --prompt "${product}, white background, product photography" \
    --output "./output/${product}-main.png"
done

20 SKUs done in one coffee break.

Cost Comparison

Based on Bailian official pricing (July 2026): qwen-image-2.0-pro = $0.071676/image, wan2.6-i2v 720P = $0.086012/sec.

	Traditional (China market)	CLI (official pricing)
Total per SKU	~$780-1,850	~$1.44
Time per SKU	5-7 days	15-20 min
Batch capability	Linear	Scriptable

Traditional pricing: 2026 industry survey data. CLI cost: 15 images × $0.072 + 5s video × $0.086.

Honest Take

Key limitations to note:

Works great for: Small-to-medium sellers, daily listings, flash sale prep
Limitations: Text rendering occasionally glitches, fine metal details (buckles/zippers) can "mutate", luxury goods still need real photography

Get Started

If you're the "script everything" type, this might be the highest-ROI tool discovery of the year for e-commerce visuals.

I've Been Trying to Write AI Video Prompts for Months. They All Sucked Until I Found a Formula.

张洲诚（Zack.ZHANG） — Mon, 06 Jul 2026 03:50:20 +0000

The Problem Nobody Talks About

Everyone's posting AI-generated videos — characters speaking with lip-sync, manga panels coming alive, virtual idols dancing. The pitch: "just describe what you want."

I tried. For months. Here's what I got:

Character's face morphed by frame 2
"Slowly looks up" became "violent head shake"
Voice-over sounded like Google Translate
Same prompt, 3 runs, 3 completely different results
No idea what to include or how long the prompt should be

Tutorials were either too vague ("be detailed") or too technical (parameter tuning from line 1).

The real issue: video prompts are structurally different from text/image prompts. You need to simultaneously control visuals, motion, audio, camera, and consistency constraints — in the right order, at the right length.

What I Found

A Skill in the Model Studio official repo called happyhorse-prompt-studio. It doesn't teach you theory — it asks you questions and assembles the prompt for you.

4-phase flow:

1. Inspiration Menu

Shows you 4 "flavors" of what HappyHorse can do:

Flavor	What it does
A · Voiced Manga Drama	Characters talk to each other, with voice + lip-sync
B · Character Voice PV	Single character self-introduction, 8-10 sec
C · Manga Panel Motion	Static manga panel starts breathing
D · Virtual Idol MV	Idol performance with choreography

2. Discovery

Asks you conversationally: character appearance, scene, emotion, dialogue, voice type, art style, camera.

3. Prompt Assembly

Assembles using the HappyHorse Formula:

Scene + Subject + Motion + Audio + Quality

Key techniques:

@「Image n」 syntax locks character identity across shots
Dialogue ≤15 characters (split shots if longer)
Japanese prompts work best (HappyHorse is JP-optimized)
Always end with キャラの顔・髪・衣装が変わらない (face/hair/outfit stays unchanged)

4. Quality Check

Auto-reviews: completeness, compliance, cost estimate, optimization tips.

Before vs. After

Dimension	Writing myself	With Prompt Studio
Attempts needed	10-20 before one usable	2-3 to satisfaction
Time per prompt	1-2 hours (research + trial)	5-10 minutes
Character consistency	Collapsed by frame 2	R2V syntax locks it
Common pitfalls	Learned by failing	Built-in troubleshooting

Honest Limitations

No guarantee of perfection on first try (AI video has inherent randomness)
Can't replace your creativity (character design is still yours)
Max ~30 seconds per generation
Requires API key (free tier available)
Japanese prompts get best results; CN/EN work but JP is optimized

Setup

npx skills add modelstudioai/skills --skill happyhorse-prompt-studio -g

Then tell your AI assistant: "I want to make a video with HappyHorse"

Execute the assembled prompt:

bl video generate --image ./char.png --prompt "[assembled prompt]" --download output.mp4

Links

Open-source Skill. Free tier available. Results subject to model randomness.

I've Been Watching Vibe Coding Posts for Months. I Finally Shipped Something Playable

张洲诚（Zack.ZHANG） — Fri, 03 Jul 2026 10:51:31 +0000

The Vibe Coding Gap

My feed has been full of "I built X with AI in one afternoon" posts for months. Chrome extensions, SaaS tools, mini-games — everyone seemed to be shipping.

I wanted to make an interactive fiction game. Branching story, multiple endings, character art, music. A visual novel you play in the browser.

So I tried the standard vibe coding approach:

"Build me a branching story engine in React" → AI wrote something. It ran. Then I spent 3 days debugging state management bugs in the branching logic.
"Generate character portraits" → Got images, but each one looked like a different art style. No consistency.
"Add background music" → AI suggested downloading royalty-free tracks. Now I'm managing audio files and dealing with licensing.
"Make choices meaningful" → The branching logic AI wrote had fake branches everywhere — choose A or B, see the same next scene.

After two weeks: a broken prototype with inconsistent art and no audio. The "vibe" was gone.

Realization: vibe coding is great for single-domain tasks. Multi-domain projects (story + logic + art + audio + frontend) fall apart because each prompt is isolated.

What novel-game Does Differently

Found novel-game in Alibaba Cloud's Model Studio Skills repo. It's not a "prompt AI to help you code" workflow. It's a pre-packaged pipeline that handles the entire interactive fiction production at once.

One input:

Use novel-game to create an interactive fiction game.
Genre: Sci-fi thriller. Programmer discovers company AI has become sentient.
Player decides what to do. 15 minutes. Image mode. Include BGM.

One output (25 minutes later): a complete, playable game in the browser.

What I Got vs What Vibe Coding Gave Me

Aspect	My 2-week vibe coding attempt	novel-game (25 min)
Branching	Buggy, fake branches	Real forks (choice = different scenes)
Art	Inconsistent styles	Batch-generated, consistent
Audio	Gave up	Procedural Web Audio (zero files)
Save system	Didn't get to it	Auto-save + 3 manual slots
Mobile	Didn't get to it	Portrait-first, touch-ready
Status	Broken prototype	Playable game

Why It Works Where Generic Vibe Coding Didn't

The difference: novel-game embeds game design expertise, not just code generation.

Things I didn't know I needed until I hit them:

"Choice = fork" principle — every major option must lead to genuinely different scenes (no fake branches)
When to merge branches back vs let them diverge to separate endings
Genre → visual theme inference (sci-fi → minimalist modern, wuxia → ink wash)
Procedural audio that matches scene mood without external files
Asset pre-generation (zero API calls at runtime)

These are lessons indie game devs learn over months. novel-game packages them into the pipeline. You skip the learning tax.

Limitations (being honest)

Art style is "digital painting" — you can't guarantee anime style or photorealism
BGM is procedural — loops become noticeable in longer sessions
Only works for interactive fiction / visual novels (not action games)
Costs money — my 15-min game was ~$2. Free credits cover the first one.

Setup

npm install -g bailian-cli
bl auth login
npx skills add modelstudioai/skills --skill novel-game -g

Needs: Node.js >= 18 + Free API Key

The Takeaway

If you've been watching vibe coding posts and wondering why your own attempts don't produce finished products — it might not be you. Generic "tell AI to write code" breaks down on multi-domain projects.

The answer (at least for interactive fiction) isn't better prompting. It's domain-specific packaging — someone already figured out the hard parts and wrapped them into a pipeline.

25 minutes vs 2 weeks. Same person, same idea, different approach.

GitHub: modelstudioai/skills
Free API Key: Get one
Bailian CLI: Install

I Wanted to Try AI Video But Kept Getting Stuck. This Tool Got Me My First Complete Video in 12 Minutes

张洲诚（Zack.ZHANG） — Wed, 01 Jul 2026 10:29:22 +0000

The Gap Between "Wanting to Try" and "Actually Doing It"

You've probably seen AI-generated videos on social media. Sora, HappyHorse, Kling — the results look amazing. You want to try it too.

But then you hit a wall:

Which model? text-to-video? image-to-video? What resolution? What parameters?
How much will it cost? API pricing is per-second. One video requires multiple model calls. What if the result sucks?
Where's the full path? Tutorials teach you to generate a 5-second clip. Then what? Script? Editing? Voiceover? Music?
Other creators make it look easy — but they clearly spent months learning. You don't have months.

The result: you bookmark 20 tutorials, register 3 accounts, and never produce a single complete video.

What If Someone Packaged All That Experience For You?

That's exactly what spark-video does.

It's a Skill from Alibaba Cloud's Model Studio (Bailian) official repository. It takes the accumulated best practices from experienced AI video creators — model selection, shot design, quality control, editing — and packages them into an automated pipeline.

What you do: Type one sentence → confirm 4 times → get a complete mp4.

What it does: Write script → design shots → select models → render → quality check (auto-reshoot bad frames) → stitch → add voiceover + BGM → output.

My First Video (12 Minutes, Zero Prior Experience)

Use spark-video to create a 30-second video.
Content: A cat watching a sunset on a city rooftop. Warm, cozy vibe. 16:9.

What happened:

AI wrote a 4-shot script → asked "OK?" → I said yes
AI designed each shot + showed cost estimate (~$1.50) → "OK?" → yes
AI rendered all shots (one auto-reshooted due to low quality score) → "OK?" → yes
AI stitched final video with BGM → "Final version OK?" → yes

12 minutes later: I had my first complete AI video.

No model selection. No parameter tuning. No editing skills needed.

Why This Works for Beginners

The design philosophy is: hide complexity, expose decisions.

You decide	spark-video handles
What video you want	Script writing
"OK" or "change this"	Shot design
"OK" or "too expensive"	Model selection + rendering
"OK" or "reshoot that"	Quality control + retries
"OK" or "tweak audio"	Stitching + voiceover + BGM

On Cost (The #1 Fear)

Cost estimate shown before rendering starts
Typical 30-second video: $1-3
New users get free credits — first video essentially costs nothing
Compare: randomly calling APIs yourself without best practices = 10x wasted spend on failed attempts

Installation (3 Minutes)

npm install -g bailian-cli
bl auth login
npx skills add modelstudioai/skills --skill spark-video -g

Requirements: Node.js + Bailian API Key (free) + ffmpeg

Or just tell your AI assistant:

Install spark-video for me. Handle Node.js, bailian-cli, and ffmpeg setup.

Under the Hood (Optional Reading)

For the technically curious, spark-video uses a multi-agent architecture:

Producer: Orchestrator (no production work, only routing)
Screenwriter: Script generation
Director: Shot design + render prompt engineering
Cast: Character consistency management (prevents "face changing" between shots)
Clip-Review / VFX-Review: Automated QA via vision models (score ≥ 7.0 = pass)
Stitch: ffmpeg composition + TTS + BGM mixing

Key patterns:

DAG scheduling (parallel across scene groups, serial within groups for continuity)
Retry-with-escalation (3 auto-retries, then escalate to user)
Cost gate (GATE 2 shows estimate before spending)

But you don't need to know any of this to use it.

Who Is This For?

Anyone who's been wanting to try AI video but hasn't started
People who find existing tools too complex or too expensive to experiment with
Content creators who want to skip the trial-and-error phase

Not for: feature films, frame-perfect animation, photorealistic human faces.

The Real Barrier

The barrier to AI video creation was never talent or technical skill. It was not having a simple enough starting point.

spark-video is that starting point. Expert methods, officially verified, packaged for beginners.

Your first AI video is 10 minutes away.

GitHub: modelstudioai/skills
Bailian CLI: Install

- API Key: Free

title: "AI Video Production in One Prompt: From Script to Final MP4 in 10 Minutes"
published: true
description: "How spark-video turns your AI Agent into a full video production pipeline — screenplay, storyboard, render, QA, and stitch, all automated."
tags: ["ai", "video", "productivity", "tutorial"]
cover_image: ""

canonical_url: "https://github.com/modelstudioai/skills/tree/main/skills/spark-video"

AI Video Production in One Prompt

The Problem Nobody's Solving

Here's what AI video tools look like in 2026:

Sora/Kling: Generate stunning 5-10 second clips. Then you write the script yourself, stitch clips yourself, add voiceover yourself, mix audio yourself.
CapCut/templates: Select a template, drag in your assets. Creative freedom? Zero.

The gap: There's no tool that takes "I want a product ad" and delivers a complete MP4. Until now.

What Is spark-video?

spark-video is an AI Agent Skill that turns your coding assistant (Qwen Code, Claude Code, Cursor, etc.) into a full video production pipeline:

Your one-sentence premise
        ↓
Screenwriter (writes multi-scene script)
        ↓
Director (creates shot-by-shot storyboard)
        ↓
HappyHorse model (renders each shot in parallel)
        ↓
Auto QA (vision model scores each clip, retries if < 7/10)
        ↓
ffmpeg stitch + TTS voiceover + BGM mix
        ↓
Complete MP4

You confirm at 4 gates. Creative control stays with you.

Real Examples

Product ad — input:

Use spark-video to create a premium wireless headphone ad.
Product image: ~/headphone.webp
Copy: "AirWave Pro — adaptive noise cancellation, spatial audio, 28h battery."
16:9. Loop BGM.

Result: 30-second product ad. 12 minutes. ~$1 in API costs.

Explainer — input:

Pop-science video, under 3 min: why cats always land on their feet.
Narration mode.

Result: 3-minute explainer with TTS voiceover.

Vertical short drama — input:

Suspense: programmer works late, elevator comes from nonexistent floor B1.
9:16 vertical. Drama mode.

Result: 2-minute vertical short for TikTok/Reels.

Architecture (Why It's Different)

The key insight: spark-video is not a video generator. It's a video production Agent.

6 Sub-Skills

Producer: Orchestrator, manages 4 confirmation gates
Screenwriter: Writes multi-scene screenplay
Director: Creates JSON storyboard per scene
Cast: Manages character consistency (cast.json)
Clip-Review: Auto-QA with vision model scoring
Stitch: ffmpeg concat + audio mixing

DAG-based Parallel Rendering

chain_groups = [
    ["S01-001", "S01-002", "S01-003"],  # sequential (frame continuity)
    ["S02-001", "S02-002"],              # parallel with above
    ["S03-001"]                          # parallel with above
]

Within a chain group: sequential (last frame → first frame chaining).
Between chain groups: parallel (up to 4 concurrent).

Auto QA + Escalation

Render → Vision model scores → >= 7.0 → ACCEPT
                              → < 7.0  → rewrite prompt → retry (max 3)
                                       → exhausted → escalate to Director

Quick Start

# Install
npm install -g bailian-cli && bl auth login
npx skills add modelstudioai/skills --skill spark-video -g

# Use (in your AI Agent)
"Use spark-video to make a product ad. Project: demo, episode 1.
 Product: smart watch. Selling points: 7-day battery, blood oxygen. 30s, 16:9."

Prerequisites: Node.js >= 18, API Key (free), ffmpeg.

When to Use

Use case	Fit
Product ads (30s-2min)	Excellent
Explainers (1-5min)	Great
Short dramas (1-3min)	Good
Social media content	Great
30+ min long-form	Not ideal
Photorealistic live-action	Not ideal