Make two decisions for a VoiceDrop article:whether to include a cover image, andwhich style of prompt to use for generation. Core principles: real material takes priority over AI art; style follows the article's temperament, never one trick for everything.
Input
All three are accepted: VoiceDrop article (MCP read_article, by title or ID) / local markdown / directly pasted body text.[[photo:KEY]] refers to user-taken real photos already inserted in the article.
Step 1: Whether to include one (before any style discussion)
| Situation | Judgment |
|---|---|
| The article already has real photos and they are relevant to the topic | Do not include. The first image is the cover (real photos > AI art; an AI cover would wash away the sincerity of a real-photo article). Point out which one to use [[photo:KEY]]。 |
| No photos, will be published / shared | Include |
| No photos, very short casual musing (<100 characters) | First ask "is this going to be published on its own"; only include if it will |
| Real photos exist but are unrelated to the topic (screenshots, receipts) | Treat as no photos, include |
Step 2: Style selection (must go through this table, defaulting to one kind is forbidden)
| Article temperament | Style | Tone |
|---|---|---|
| Speculative / opinion / methodology | Minimalist conceptual image | One concrete object as metaphor, large negative space |
| Life essay / nature / walking | Watercolor light wash | Light, transparent, light-colored background |
| Nostalgia / memories / family | Film / oil painting | Warm tones, graininess |
| Tech / tools / AI / tutorials | Editorial cover | Concrete objects (cards, panels, canvas), avoid generic robots |
| Humor / self-deprecation | Cartoon light illustration | Relaxed, not rowdy |
| Grief / mourning / illness | Sketch or very light watercolor | Low saturation, restrained, no faces,absolutely no cartoon / advertising style |
When unsure of the temperament, read the full text before judging; do not guess from the title.
Step 3: Assemble the prompt (hard specs)
- Ratio 2.45:1, size 1568x640(VoiceDrop cover image spec; the local wjs-publishing-wechat line is 900×383, do not mix them up)
- The cover image's main title is distilled from the article title into 6–10 Chinese characters(minimalist conceptual images may use 1–4 large characters), must be clearly readable
- Composition: large title on top, main visual below, text stretching full width left and right
- Color scheme varies with the topic, default light background; articles in the same batch must not clash in color or composition
- Avoid list (write into the prompt): garbled text, too much small text, real brand logos, pure atmospheric wallpaper, heavy blue-purple gradients, cheap marketing poster feel
Step 4: Generate the image (only do it when the user requests)
gpt-image-2-skill go through --provider codex，--size 1568x640; or within VoiceDrop hand the prompt to new_photo。
Output contract (these four lines are fixed for each article)
判断：配 / 不配 —— 一句理由
风格：<Step 2 表里的风格名>
Prompt：<完整可直接用，含 2.45:1 / 1568x640 与避免清单>
（不配时第三行改为：封面 = [[photo:KEY]]，横向裁 2.45:1 的建议裁法）
Red Flags
- Two consecutive outputs in the same style → go back to Step 2 and re-evaluate
- The text contains real photos but you are still writing an AI prompt → go back to Step 1
- 900×383 or 2.35:1 appears in the prompt → the specs were copied from the wrong line
- Cartoon / bright colors appear in a sorrowful text → start over