Package a piece of video material (script/outline/subtitles/verbatim transcript/audio/video) into a complete publishing plan directly usable on the selected platform: cover title, platform title, work description, platform tags. After confirmation, generate cover creative directions and images.
Overview
Method reference jennie-dingding-cover-packager: First determine "what in this episode's content is truly worth clicking," then decide "how the visuals can let viewers understand at a glance." Process only one episode of content at a time.
Fixed delivery of four parts:
- Cover packaging: condensed packaging brief, title derivation, 3 cover titles with strategic differences, recommended line breaks, 3-5 creative theme directions, cover pre-review, platform size adaptation.
- Platform title: generated only for the platform selected by the user.
- Work description: by default one set shared by all selected platforms (including YouTube).
- Platform tags: generated only for the platform selected by the user.
Process:包装简报 → 标题策略 → 封面标题 → 平台标题 → 封面主题 → 封面预审 → 作品描述 → 平台标签 → 待用户确认 → 生图 → 标题合成出成品封面。
When to Use
- User provides a video script/transcript/audio and wants cover titles + publish copy
- User mentions cover packaging / video packaging / publishing plan / title / tags for one episode of content
- User wants cover art direction (creative scenes) before generating the actual cover image
- User says "make cover title description tags for this episode's video"
When NOT to Use:
- Batch packaging of multiple episodes (only one episode at a time)
- The user only wants to edit the video or make subtitles (out of scope)
- Title creation for pure articles/image-text (without video material)
Core Pattern
Step 1: Material and platform confirmation
Minimum input: one piece of material for this episode + at least one target platform (Xiaohongshu/Douyin/WeChat Channels/Bilibili/YouTube, multiple selections allowed).
- When material is missing, ask the user to provide a script, outline, subtitles, verbatim transcript, audio, video, or complete description; when audio/video cannot be transcribed, clearly state the limitation and do not pretend to understand.
- When material and platform are complete, execute directly by default (Mode A), and do not let follow-up questions become a blocker; if the user wants to control the direction, they may answer at most 2 questions (who it is for/what to do after watching), or provide account information to customize the tone (Mode B/C).
- When the user has already provided a title/cover, review it first and give a "keep/partially modify/redo" conclusion; do not replace confirmed content without permission.
Step 2: Extract the packaging brief
Extract from material facts and present with at most 6 items: one-sentence theme, target audience and scenario, pain point/reason to watch, result/conflict/reversal/evidence, content type, most suitable user action. Content facts take priority; account information is only for calibration.
Step 3: Title derivation and generation
First complete the derivation chain:内容事实 → 独特结果 → 人群动机 → 单一点击承诺 → 主策略+备选策略 → 压缩 → 兑现检查, then generate 3 cover titles with genuine strategic differences + 1 recommendation + 2-3 lines of line-break layout + titles for each platform.
Select 1 main strategy and 1 backup from the table below, generate at least 8 candidates from within the corresponding formula pool, then eliminate:
| Strategy | Suitable content | Representative formula |
|---|---|---|
| Judgment/choice type | A vs B, how to choose, worth it or not | [好的特质] VS [坏的特质] 如何区分；你到底要不要 [话题] |
| Pain point/pitfall avoidance type | Pitfalls, correction, loss | [领域] 这 [N] 件事千万别做；[不想要的结果] 的最根本原因 |
| Identity immersion type | Clear audience, "this is talking about me" | 给 [一群人] 的一个忠告；没有 [资源]，也能 [结果] |
| Result checklist type | Tutorials, checklists, collectible | [结果] 的 [N] 个技巧；如何不 [讨厌的事] 就能 [结果] |
| Experience reversal type | Personal testing, review, before-and-after changes | 我如何在 [时间] 内 [结果]；从 [不想要] 到 [想要] |
| Curiosity gap type | Counterintuitive, insider, cognitive upgrade | [一群人] 不会告诉你的建议；为什么 [大家都觉得好的事] 其实有害 |
| Interactive test type | Take a quiz: which type are you | 测测你的 [属性]；[N] 秒看出你是哪种 [人群] |
| Digital anchor type | Data, efficiency, comparison | [N] 分钟完成 [原来要 N 小时的事]；靠 [方法] 拿到 [数字结果] |
Show the user only the primary/backup strategy and the reasons for adopting the cited formulas. Prioritize specific results, real reversals, and visually expressible statements; do not default to capturing generic anxiety of the "stop doing.../are you still..." type.
Step 4: Cover concept and pre-review
Build a causal brief of "title promise → content function → visible input → character action → visible output"; after internally diverging at least 8 directions, show only 3-5. Each direction must have: one memorable core visual engine, a clear character role, and a visible input-action-output loop.
The cover must simultaneously pass four checks:
| Check | Question |
|---|---|
| Semantic accuracy | Does the image express the function the content truly fulfills (rather than literalizing the topic noun)? |
| Visible causality | Can you see the input, character action, and output? |
| High concept | Does it use an unexpected but accurate physical mechanism to amplify the value? |
| Low visual noise | Do the title, core object, and character action form a single visual flow? |
Eliminate: related venues with no causality, ordinary spaces plus props, grand realism but weak semantics, empty backgrounds, information stacking. Realism is not a quality tier—a real-person IP may keep only the face realistic, with the rest using light 3D/toy/paper-cut/staged expression according to the account's historical covers.
Step 5: Description and tags
- Description: a concise, natural, general description with a reason to watch, shared across all platforms (including YouTube).
- Tags: generate only for the selected platform; do not pad with irrelevant hot words; event tags must be explicitly provided by the user.
Step 6: Output the publishing package
Fixed section order:包装简报 → 标题策略与公式 → 封面标题 → 分平台标题 → 封面主题方向 → 封面预审与核对 → 精简作品描述 → 分平台标签 → 待确认. Hide the formula pool, the complete derivation chain, internal candidates, and the scoring process; do not output empty sections for unselected platforms.
Step 7: Generate images after user confirmation
Enter image generation only after the user explicitly confirms the title and topic direction. For a fixed real-person IP, first execute the clear face reference image gate (the reference image must actually be passed into the image tool; if this cannot be done, explain the limitation); if there is no fixed IP, generate directly. After generation, perform quality control according to the pre-review checklist.
Step 8: Composite the title into the finished cover
AI image generation models are prone to garbled text when rendering Chinese; cover textmust always be composited in post, and the model must not draw the text directly:
- The image generation prompt must enforce "no text, no watermark," and reserve title space at the top of the base image (subject centered and lower).
- After the base image passes quality control, use the skill's built-in script to composite the confirmed title onto it:
Differentiation within the same direction: the image generation prompt must not reuse the previous text as-is. When generating images again for the same creative direction, at least two visual variables must be rotated—composition (centered/rule of thirds/diagonal/close-up/long shot), viewpoint (eye level/high angle/slight low angle), lighting (soft morning light/afternoon highlights/warm dusk light/indoor lighting), background era or material (stage/light 3D/paper-cut collage/toy advertisement)—so that multiple generations in the same direction show clear differences while retaining the core visual engine.
python3 scripts/add-title.py --image 底图.png \
--title "第一行|第二行" --out 成品封面.png \
[--stroke "#E65A14"] [--fill "#FFFFFF"] [--top 0.07] [--width 0.92]
The script automatically lays out according to the Quick Reference typography specs: the title block width is about 92%, the top safe margin is 7%, the stroke shadow counts toward the safe margin, and multi-line text distributes width by character count. After generation, you must inspect the final image before delivery: the title has no garbled characters, is not blocked by people/objects, and the stroke color harmonizes with the image; the stroke color should preferably echo the main color of the background image (such as the character's clothing color).
Quick Reference
Usage example
# 用户请求示例（对话中直接说，不需要真的执行命令）
使用 video-cover-packager。
平台：抖音、小红书。
方式：A，直接包装。
素材：本期字幕文件 subtitles.txt
Platform rules quick reference
| Platform | Title | Tags | Cover |
|---|---|---|---|
| Xiaohongshu | 3, about 20 characters, identity immersion/low barrier/collection reason | 5-8 | 3:4 vertical |
| Douyin | 2, about 30 characters, tool/conflict/efficiency/judgment | 4-6 | 3:4 cover layout preferred |
| WeChat Channels | 2, clear and steady, suitable for community forwarding | 3-5 | Vertical (follow 3:4) |
| Bilibili | 2, fully explain the object/method/conclusion | 3-6 | 16:9 horizontal |
| YouTube | 2, keep the core search terms, use the main language of the material | 3-5 strongly related | 16:9 horizontal |
Default proportion for a two-line large title in 3:4 vertical
When there is no account template: the title outline width is about 90%-93% of the canvas, and the height is about 31%-34%; the top safe margin is 6%-8%, and the left and right are 4%-5% each (calculated from the outermost edge of the stroke shadow). Title color should be chosen dynamically according to the brightness of the current background, the character's clothing, and the core object; do not use a fixed palette.
Hard rules
- Do not fabricate numbers, results, features, prices, endorsements, or partnerships that do not exist in the material.
- For multiple platforms in the same video issue, by default use the same cover title, creative theme, IP, and core color scheme, and only recompose according to the ratio (not simply cropping).
- After the user confirms the title, lock the text, line breaks, font size, position, color, and stroke; when modifying, unlock only the specified items.
- For images of real people, use only the user themselves or materials confirmed to be authorized for use.
Common Mistakes
- Batch processing multiple issues of content — This skill packages only one issue at a time; split multiple issues into multiple tasks.
- Start thinking about the cover from "what to draw" — First solve "why it is worth clicking" (the title promise), then decide the visual mechanism; otherwise the output will be a meaningless staged image.
- Treat the literal masking of thematic nouns as creativity(e.g., draw a robot for "AI tools") — the image must express the function completed by the content, forming an input-action-output loop.
- Mechanically cobble together the three titles "steady/conflict/suspense" — The 3 candidates must have real strategic differences, and the core psychological mechanism of the formula must be retained, not forced on afterward.
- Output the entire internal process — Formula pools, reasoning chains, candidate elimination, and scoring are internal work; the user sees only the conclusion and rationale.
- Generate images directly without confirmation — "Give the result directly" does not mean generate images directly; only generate images after the text publishing package and cover pre-review are confirmed.
- Generate empty title/tag sections for unselected platforms — Output only the selected platform.
- Directly crop a vertical cover into horizontal — Each platform ratio requires reallocating the space for the title, people, and scene.
- Let the AI image generation model draw the Chinese title directly — garbled text is guaranteed; generate only a text-free base image, and use the title with
scripts/add-title.pypost-processing compositing. - Delivering only the base image and considering it done — the base image is not the finished product; only after title compositing and viewing verification is it considered a delivered cover.