Generate a 3:4 vertical cover image for Lin Yuebanzi's (LQ) WeChat Channels/Douyin/Xiaohongshu short video with one click, using MuleRun's Nano Banana 2 Edit API, with LQ's own avatar as the reference image, producing a cover that follows the viral patterns of short video platforms.
Scope of this skill
This skill only generates 3:4 vertical short video covers, used for short video thumbnails on platforms such as WeChat Channels/Douyin/Xiaohongshu.
| Dimension | Fixed scope of this skill |
|---|---|
| Ratio | 3:4 vertical (API uses 3:4, 1080×1440) |
| Use case | WeChat Channels/Douyin/Xiaohongshu short video thumbnails |
| Visual style | 11 options (default: black background with dark grid + cyan-blue-yellow high contrast + giant title dominance) |
| Text position | Middle 2-3 lines of giant title + top hook bar + bottom tag |
| Facial expression | Stable and confident smile, looking at the camera, like a mentor sharing tips |
| UI elements | Dark workflow screenshots + Agent dialogue fragments + result thumbnails |
Trigger determination:
- User mentions "video" "Douyin" "WeChat Channels" "Xiaohongshu video" "vertical" "3:4" "thumbnail" → use this skill
- User mentions "WeChat Official Account" "article title image" "widescreen" "2.35:1" → do not trigger this skill
When to trigger this skill
Explicit trigger scenarios:
- User says "make a cover for this video" "video cover image" "WeChat Channels cover"
- User says "Douyin cover" "Xiaohongshu video cover" "short video title image"
- User says "vertical cover" "3:4 cover" "9:16 cover"
- After finishing the voiceover script, the user asks "what about the video cover" "the cover hasn't been made yet"
- Any scenario related to "the last step of preparing an image before publishing a short video"
Scenarios not to trigger:
- WeChat Official Account widescreen article cover (2.35:1 banner title image)
- WeChat Moments nine-grid
- Avatar, logo, poster
- Non-thumbnail assets such as in-video overlays and subtitle bars
Core process
This skill is executed in five steps (Step Zero to Step Four),each step has a quality check point, and you do not proceed until it passes。
Step Zero: Choose a style
Before generating the prompt, first ask the user which cover style to choose. List the 11 styles in references/style_catalog.md and let the user choose one:
| Number | Style name | Corresponding template |
|---|---|---|
| 1 | Dark tech title style | references/styles/dark-tech-title.md |
| 2 | Bright YouTuber style | references/styles/bright-youtuber.md |
| 3 | Dual-screen contrast narrative style | references/styles/split-screen-narrative.md |
| 4 | Product key visual style | references/styles/product-hero.md |
| 5 | Collage information hurricane | references/styles/collage-info-storm.md |
| 6 | Workbench scene style | references/styles/workspace-scene.md |
| 7 | Glitch warning style | references/styles/glitch-warning.md |
| 8 | Agent call-to-action style | references/styles/agent-hype.md |
| 9 | Card display persona style | references/styles/card-showcase.md |
| 10 | Cute pet code style | references/styles/cute-mascot.md |
| 11 | Handheld device display style | references/styles/device-showcase.md |
When the user does not choose, default to Style 1 (dark tech title style).
After selecting a style, read references/styles/{风格}.md 's complete template, and always follow the brand baseline in references/brand_baseline.md .
After selecting a style, all subsequent fill-ins must follow that style's template; do not mix elements across styles.
Step 1: Understand the voiceover and extract cover materials
Fully digest the video voiceover script and extract six key materials from it:
- Main title (large text at the top) — 4-8 characters,must be "plain language" and cannot use jargon. Rewrite the core pain point of the script using the most common words. For example:
- ❌ "Connected the API but can't get online?" (jargon)
- ✅ "Claude can't get online?" (plain language)
- Subtitle (large text at the bottom) — 4-6 characters, give apromise/solution. For example "Solved in 10 seconds" "Done in one click" "Done in 3 steps"
- Auxiliary subtitle (below the main title) — 6-12 characters, supplement the context. For example "Claude Desktop · Pitfall Guide" "Nanny-level tutorial"
- UI element narrative (the visual story in the middle of the image) — must be a **"problem→solution" two-screenshot comparison**:
- Element 1:error/failure statescreenshot (dominant visual)
- Element 2:success statescreenshot (auxiliary, smaller size)
- Connect the two with an arrow to tell a "❌ → ✅" transformation story
- Expression:consistently use "confident smile, looking at the camera, like a mentor"— never use surprised/wide-eyed/exaggerated expressions, because LQ's persona is "steadily sharing useful content," not a "shock-style blogger"
- Gesture —— does not need to stay consistent with the reference image, choose naturally according to style/composition. For example: dark tech title style uses "index finger slightly raised pointing at the title," product main visual style uses "finger pointing at the UI," dual-screen comparison style uses "both hands pointing left and right."
references/brand_baseline.mdhas complete gesture options.
Key judgment: Is the main title plain language?
Show the main title to aperson who completely does not understand technologyand see if they can understand it instantly?
- "API", "Token", "Webhook", "protocol" → not plain language, must be rewritten
- "can't get online", "doesn't work", "can't find it", "can't search it" → plain language
If unsure, ask yourself: "Can my mom/my wife understand it?"
Step 2: Generate the complete prompt according to the template
According to the style selected in Step 0, go to references/styles/{风格}.md to read the prompt template for the corresponding style, and fill in the materials extracted in Step 1. All styles must follow the brand baseline in references/brand_baseline.md .
Key insistence — brand baseline (cannot be changed no matter the style):
- Aspect ratio:3:4 vertical (API uses 3:4)
- Character:Chinese male programmer around 35 years old, short hair and glasses,wearing a dark casual T-shirt, half-body
- Default expression:confident smile, looking at the camera, resolutely eliminate "surprise/wide-eyed/exaggerated" (unless the failure-warning style requires a slightly serious tone, but shock is still not allowed)
- Title principles: must be plain human language, 4-8 character main title, technical jargon must be translated into everyday vocabulary
- Default visual hierarchy: the title is the lead protagonist, the person takes up 20-30%, does not steal the scene
- Watermark: bottom right "@林月半子 🦞", white text with outline
- Fixed reference image: use the LQ avatar as the edit reference image, maintain facial feature consistency
Key insistence — the default visual of Style One (dark tech title style) (other styles are replaced according to the catalog template):
- Background tone: dark tech feel, butdo not hard-code the background into the same black-background template; it can be changed by theme into a dark workbench, dark UI canvas, dark gradient, dark collage
- Main visual: 2-3 lines of giant Chinese title, must be the first protagonist of the image
- Gesture: one hand slightly raised, index finger pointing up to highlight the key point, do not steal the title's spotlight
- Title color: a three-part combination of cyan-blue #00D4FF + white text with cyan outline + bright yellow #FFD93D
- Information layer: top hook bar + middle UI fragments + bottom tags, but all can only serve the main title
Parts that can differ per piece:
- Selected style
- Top/bottom title text
- Auxiliary subtitle
- Specific content of the UI screenshot (change according to the video theme)
- Error message text content
- Specific background tone and scene (within the range allowed by the style)
Step 3: Call the MuleRun API to generate images
Use the linyuebanzi-image-gen skill's script, passing in the complete prompt:
python linyuebanzi-image-gen/scripts/generate.py \
--mode edit \
--prompt-file ./prompt.txt \
--images "https://r2.cloudnative101.net/assets/lq.png" \
--blocklist references/blocklist.txt \
--aspect-ratio 3:4 \
--name-tag video-cover-{视频关键词} \
--output-dir ./video-covers
The script will automatically complete: create edit task → poll for results → download image → save prompt and metadata.
Fixed parameter values:
- Reference image URL:
https://r2.cloudnative101.net/assets/lq.png(LQ avatar; if the URL is invalid, you can use this skill'sreferences/lq_ref_face.pnglocal fallback) --mode edit(required)--aspect-ratio 3:4- resolution:
2K(script default value) --output-dir ./video-covers(Current working directoryundervideo-covers/, do not write the generated results into the skill directory)
Authentication:
Read the Bearer token from the environment variable MULERUN_API_KEY . If it is not set, remind the user export MULERUN_API_KEY=sk-xxx。
Batch generation note:
MuleRun limitsto at most 10 active tasks at the same time, using --manifest --parallel to run more than 10 in parallel at once will receive HTTP 429 (Quota exceeded). If there are more than 10, run them in batches, or rerun failed items individually. In addition, edit mode occasionally returns 2 variants at once (file names with the -0/-1 suffix), which is normal; during self-check, look at both and choose the better one.
Step 4: Read the image for self-check (must be done after generation)
The generated PNG must be read back with the Read tool and checked item by item against the checklist below. If it does not pass, modify the prompt according to the "Common failure points and countermeasures" table and rerun:
- Title text: check the main title/subtitle/tags word by word for typos, repeated characters, or missing characters (such as repeated text like "无法访问无法访问外部网络")
- Expression: whether it is a confident smile looking at the camera (the failure-warning style allows focused seriousness, but it cannot be displeasure/indifference/shock)
- Watermark:Does "@林月半子 🦞" exist in the bottom-right corner
- Hierarchy:Is the title the primary focus, do people occupy only 20-30%, and does the subtitle press on key UI or faces
- UI elements:Are key phrases (error banners/toggle names) clearly readable; if long garbled sentences are too conspicuous, blur them on rerun per brand_baseline's "text rendering and garbled text control" requirements
Only hand the result to the user after self-check passes; if there are only minor flaws, explain the flaws and rerun suggestions together.
After generation is complete
The script will automatically save .png(image),.txt(prompt),.json(API metadata), three files with the same name, to the --output-dir specified ./video-covers/(under the current working directory), with the naming format {name-tag}-{时间戳}, no manual save-as needed.
Feedback to LQ must include:
- Image file path
- If the image has problems (blurry text/wrong expression/weakened UI elements), suggest which parameter to fine-tune
- If a rerun is needed, tell him which part of the prompt to change
Common failure points and countermeasures
| Failure point | Cause | Countermeasure |
|---|---|---|
| Exaggerated expression/wide-eyed | The prompt did not explicitly exclude a surprised emotion | Explicitly write "NOT shocked, NOT wide-eyed" + describe the target emotion "confident mentor smile" |
| The main title uses jargon (API/Token) | Cutting corners during the first step of distillation | Must go through the "plain-language check": can your mom understand it? If not, rewrite it |
| The background looks like the same image with the text swapped repeatedly | The background template was hard-coded | Explicitly write "background must be newly composed for this topic, not reused 1:1", fix only the color scheme and hierarchy, not the specific composition |
| The UI screenshot is too small to see clearly | The prompt did not emphasize the dominant visual | Write "headline dominates, UI supports the title", the UI can be tilted/cropped/blurred as an atmospheric layer |
| Overall too bright/like a YouTuber thumbnail | Still using the old warm-color background wording | Explicitly write "dark layered background, charcoal/blue-black base, cyan glow accents" |
| The person is too large and steals the scene | The prompt did not limit the person's proportion | Explicitly write "subject relatively small, bottom 20-25% only, text dominates 3-4x more than person" |
| Labels and small images are too cluttered | All information elements were made into the protagonist | Emphasize "all secondary elements are decorative/supporting only" |
| Typos/duplicated characters appear in the title | The Chinese text blocks did not have rendering protection added | Append "Render every Chinese character EXACTLY as written, no duplicated or missing characters" after each Chinese text block (11 styles tested with 0 typos) |
| Long garbled sentences in the UI screenshot | Letting the model render Chinese longer than phrase length | Only put phrases of ≤8 characters as readable text; for long text explicitly require "blur or fade long body text into unreadable soft texture" |
| The subtitle presses on key content | No space was reserved for the subtitle | Specify "reserve a clear zone for the subtitle; must not overlap key UI or the subject's face" |
References
- Style catalog index:
references/style_catalog.md - Brand baseline and general fill-in guidance:
references/brand_baseline.md - Complete prompt templates for 11 styles:
references/styles/{dark-tech-title,bright-youtuber,split-screen-narrative,product-hero,collage-info-storm,workspace-scene,glitch-warning,agent-hype,card-showcase,cute-mascot,device-showcase}.md - Verified case (Claude web-enabled pitfall-avoidance cover):
references/examples.md - General image generation script:
linyuebanzi-image-gen/scripts/generate.py