SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsaikey-cover

aikey-cover

Design & creation

View repository

Updated 10-18 22:09

About this skill

aikey-cover: one set of cover images

First read aikey/references/家族公约.md(if unreadable → say plainly "the convention can't be read, so the red lines can't be guaranteed" and stop,do not substitute memory for the convention), then read aikey/references/交互规范.md(🔴 wrap up per §4's three-piece set — Recap · Before/After · next step given as numbered options; if information is missing, per §4 usemultiple-choice questionsto ask,only one at a time; for cases where this doesn't apply see §5), then read memory {vault}/08-技能记忆/aikey-cover/ + _通用/. The criteria are in references/规则卡.md, and the starting table of platforms and ratios is in references/平台尺寸表.md; read both before starting work.

This skill exists for only three reasons — other models can draw images too:

  1. One layout per ratio. Stretch one image into four ratios and either the face or the big text will be cut off in at least one version
  2. Cover text crosses the red lines. The cover is the front door viewers see first; prohibited words, truthful claims, and not naming names cost the most here
  3. Delivery is verified. A filename saying 16:9 doesn't count; only the pixels read out count

Step one: gather the three inputs

Input If missing, what to do
Subject material image(your own photo / video frame grab) Ask. 🔴 Do not substitute someone else's face, stock portraits, or a screenshot from the previous item
Background image(scene, product interface, mood image) Ask; if the user says "no background", use a solid color or gradient background, as stated
Theme(what this piece is about, or give the cover text directly) Ask. If there's a script / a video, read the script / watch the video, don't rely on just one sentence

Look over the materials first and state clearly each image's role (who is the subject, which background elements to keep). If several are missing, say them all at once, but per the interaction rulesask only one multiple-choice question at a time。

How far along is the video? Judge by aikey-title/references/拍摄进度分档.md . If the video is already shot and can't be reshot, the cover should only promise what's actually in the video.

Step two: cover text (handled centrally via /aikey-title)

Call it centrally per family convention 〈8〉②: read /aikey-title 's criteria and decide the cover headline on the spot,and state in the report which structure number of it was used。

  • If the user gave exact wording → use it as is, but still run the red-line check; if it fails, say where and give a fix, without changing it on your own
  • If not given → distill from the theme / script, keeping each line's character count within platforms[].cover_text_limit; if the platform doesn't configure this item, follow /aikey-title 's standard for that platform
  • If two languages are wanted → one image per language,with the two languages matching in meaning and each reading naturally, without piling both languages into one image. If only one language is given, translate the other faithfully, adding no numbers or conclusions absent from the original
  • Red lines follow convention 〈3〉,with platform differences still accounted for: The full set for domestic public platforms takes effect,domestic_public: false platforms according to their rules

Step 3: Define the delivery checklist

Read config.yaml of platforms, cross-reference with references/平台尺寸表.md to determine which "language × ratio" combinations are needed this time.

  • If it can be calculated → list it in a table for the user to see,start directly, without asking item by item about size, directory, or style
  • platforms is not in the size table → ask only this one platform what ratio it needs
  • If the user named a platform, language, size, or number of images for this request → follow this request
  • Only fill in one missing image (e.g., only missing Bilibili 16:9, only missing Instagram 9:16) → make only the missing one, do not redo the full set

File names must includelanguage + ratio + orientation, e.g.:中文-3比4-竖版.png、英文-9比16-竖版.png、中文-16比9-横版.png. For revised versions add a suffix before the extension:中文-4比3-横版-v2.png。

Step 4: Generate images

Use the image generation / editing tool in the current environmentthat can accept reference images. If an image model is specified in the environment, use it; read credentials from environment variables,never write them into any file。

  • Generate each "language × ratio" separately once, and write the prompt clearly: target ratio and pixels, original cover text and language, the respective roles of the character image and background image, the character features and color scheme to preserve
  • Feed both original assets in every time. If you only modify the next image from the previous result, the person will look less and less alike. If the tool allows, you can additionally use the first result as a style reference
  • Within one set, the character, color scheme, and font style are consistent; each ratio rearranges the character, text, and whitespace according to its own canvas
  • No watermarks, no platform logos, no swapping the person
  • Chinese characters are the least stable part of AI image generation. If the text is wrong twice on the same image → switch approach: generatea textless base image, lay out the text in HTML and screenshot it onto the image (text is controllable), and note the switch in the production notes
  • If no usable image tool → say so clearly, deliver "the prompt + layout plan for each image",do not automatically switch to an API that requires separate payment

Step 5: Verify

Look at each image, then use a read-only tool to read theactual pixels(e.g., sips -g pixelWidth -g pixelHeight、python3 -c read with PIL). Check item by item:

  • Every "language × ratio" in the checklist is present, and the ratio is correct (file names are not evidence)
  • Cover text matches character by character, with no typos, missing characters, or garbled text
  • Faces and large text do not touch the edges or get cut off; ratios with safe-area requirements are checked against the rule card
  • The person is the same person, with no obvious deformities

Redo any non-compliant image separately; if the target size cannot be achieved, report theactual sizeand the unfinished items.Only say it is complete after everything has been checked, and if partially failed, keep the compliant ones and only redo the failed ones.

Step 6: Save to disk

Directory:{vault}/{config.paths.drafts}/封面/YYYY-MM-DD-主题/

  • Use the user's local current date for the date; if the user specified one, use the specified date
  • Keep the topic in Chinese, remove characters that cannot go into file names such as / and other characters that cannot go into file names,do not interpret the topic as a path
  • If a directory with the same name already exists on the same day: add -02、-03for a new task; if revising the same set → reuse the directory, add -v2，for new files, do not overwrite, unless the user explicitly says to overwrite
  • The generated image comes from the tool's output locationcopyit in; find the file by the real path returned by the tool, do not guess the path
  • The original material stays untouched

Write in the same directory 制作说明.md: theme, cover text for each language (and the /aikey-title structure number), each image's language / file name / actual pixels, final prompt, tools used, and whether it was switched to HTML text overlay. Next time you revise an image, continue from here.

Delivery

## 封面：[主题]　｜　片子进度：[档]　｜　目录：[绝对路径]
| 文件 | 语言 | 比例 | 实际像素 | 给哪个平台 |
|---|---|---|---|---|
封面字：[中文] / [英文]　｜　结构：S[编号]（/aikey-title）　｜　红线：[✓ / 改了哪处]
未完成：[无 / 哪张、为什么]

Wrap up with the three-part interaction spec. The next-step options are usually: revise a certain image / add a ratio for a platform / go post (after posting, come back to /aikey-retro）。

Absolutely never do

  • Do not publish, do not upload. Stop at producing images on disk
  • Do not stretch or crop an existing image to pass it off as another ratio
  • Do not use substitutes for missing material (other people's faces, old images, stock portraits)
  • Do not claim that a certain margin "won't be blocked on any platform". The safe zone is a design baseline; actual platform cropping is subject to the upload interface
  • Do not add numbers, results, or promises to the cover text that are not in the original draft

Memory

Self-check before finishing: which layout / color scheme / font did the user reject (record as "correction")? Which set of covers was validated as effective by /aikey-retro click data (record as "effective method", with values + platform)? Which platform's actual cropping differs from the size table (record as "correction", and come back to update the size table)? Write to 08-技能记忆/aikey-cover/, and check for duplicates first.

Don't know the next step → go back to /aikey。