Goal: distill from the voiceover script a click hook that can hold up across platforms, use only the Qiuzhi direction, select the 1 best-matching template according to the script conflict, directly generate 1 9:16 visual master of “person + scene + accurate Chinese title,” then recompose and generate 3 true-size covers (3:4/4:3/16:9) with the same visual proposition on real canvases, shared across four platforms and 7 logical entry points. The three ratios must each be recomposed and generated separately; do not treat cropping or temporary framing on the publishing page as adaptation. Do not split into textless base images, typography scripts, scorers, or multi-stage compositing.
Inputs and Boundaries
episode-config.schema_versionis 4; if the identity reference or script binding is not satisfied, stop and do not modify the contract on your own.- Read the current
01-口播稿.md、12-发布/多平台发布内容.mdand the locked Douyin publishing information. - The person reference is fixed to this issue's
05-封面/reference/jeffrey-reference.jpg, and is passed to the image generation model as an image reference. - ⑥ First generate 3 complete 9:16 candidates, then generate 3 shared true-size covers for default 01; when Jeffrey explicitly switches to 02 or 03, these 3 must be regenerated for the new rank before ⑫, and finished assets from the old rank must not be mixed in. Do not simply crop out the person, main object, or title.
- Do not write “imitate a certain blogger's style” in the prompt. Use only
references/douyin-cover-study.md's Qiuzhi observation andreferences/qiuzhi-template-library.md。 - By default, generate only 3 Qiuzhi direction candidates; the Raphael and Zhuzige directions do not enter default output for now.
- Arrange the three in descending order of recommendation as
01 → 02 → 03. When Jeffrey has not explicitly switched, subsequent publishing automation preferentially uses01as the source candidate for the published cover.
Execution
1. Distill cover text
Distill from the voiceover script:
主标题: 4–10 Chinese characters, at most two lines; retain product names that truly affect understanding.辅助短句: 0–8 Chinese characters; do not add it when the main title already makes things clear.画面事件: what Jeffrey is doing, and what specific problem or result he faces.
Prohibit empty phrases such as “AI magic tool” or “double efficiency” that can be applied to any topic. Cover text must not exceed the facts of the voiceover script.
2. Select 1 best-matching Qiuzhi template
Fully read references/douyin-cover-study.md 's Qiuzhi section and references/qiuzhi-template-library.md, first filtering templates by the core conflict of the voiceover.
- First internally filter out the 5–8 semantically best-matching templates.
- Final selection:1template that semantically hits this issue's hook best; output as a single cover (starting from v1.19, no more 3 ranked candidates are produced).
- Comply with
1个基础模板 + 最多1个主物件机制。 - In principle, avoid repeating the template family already used in the previous issue (anti-repetition rule).
- Do not default to astronauts, cockpits, futuristic offices, or laboratories based on “AI feel”;
CONDITIONAL_ONLYstill use them only when the semantics clearly match.
Each prompt must describe in one pass:
- 9:16 Douyin complete cover;
- Jeffrey's identity, position, expression, action, and clothing;
- one main scene or main object directly related to this issue;
- The exact original text, line breaks, position, color, and thick black outline of the main title and optional auxiliary short sentence;
- the layout pattern corresponding to that direction;
- No typos, extra text, platform UI, watermarks, QR codes, additional people, or irrelevant decorations.
2.5 Pre-generation identity analysis (v1.17, derived from Lao Han AI product DH Cover v4)
The fixed identity constraint sentence has been deprecated. Before generation, a vision model must first be used to analyze the current issue's jeffrey-reference.jpg，dynamically extractedtruly visible features and write them into each prompt: face shape, hairstyle, glasses style, beard form, mole position, skin tone, current clothing, and expression. It is prohibited to write "the same real person in the reference photo" as the sole identity description. Identity analysis features update when the reference image changes, and old descriptions must not be reused across issues.
Facial fidelity red line (Jeffrey 2026-09-20): The head must be kept photoreal exactly as in the reference photo, preserving the real skin texture, pores, and facial structure; Q-versionization, 3D stylization, skin smoothing, or any face change are prohibited; the prompt must explicitly state "face rendered photorealistic exactly as the reference photo, real skin texture, no beautification, no stylization". The only variable items are: clothing, expression, action, scene, props, and head-to-body ratio. The only legitimate source of the big-head feel = shrinking the body: The head keeps natural, realistic proportions and shape (no face widening, enlarging, feature enlargement, or cartoonization); all Q-version contrast is achieved by shrinking the 3D body.
2.6 Single-image contract (v1.17, derived from DH Cover v4)
- Aspect ratio hard contract: The top of the head to the chin occupies 16%–26% of the image height, and the character overall occupies 48%–72%; a close-up hero shot above the waist, with full-body shots or the character shrunk at the bottom prohibited.
- Expression tension clause: The expression must have clear tension such as surprise, excitement, certainty, or doubt;an ordinary smile or relaxed standing pose is unacceptable。
- Action phrase contract: The character's action is written as a complete phrase of 4–40 characters and includes one of lifting, opening, operating, inspecting, pointing, or confronting; it cannot merely say "operating".
- Layout is not fixed: It is specified by the selected template as close-up wide angle, low camera position, strong perspective, or main object extending forward; it is not fixed as "lower half of the character + top horizontal row".
- Warm-cold contrast: At least two strong contrast theme colors in warm-cold or light-dark must be formed, and the character and main object must have clear silhouette separation; a single blue tone across the whole image, collage partitioning, isolated ID photos, and multiple Jeffreys are prohibited.
- Advertising KV three-piece set (derived from a 2026-09-20 real-shot comparison; without it, only reaches 75% of Qiuzhi): ① Add a warm golden glow edge around the real person's head (a seam-hiding technique that "welds" the real face into the 3D scene); ② surround the character with 3-4 small glowing 3D pendants (theme elements, filling the space around the character); ③ an overall high-saturation glossy finish (explicitly write HIGH-SATURATION punchy colors, bright glossy finish), with cinematic dark tones prohibited.
- General contract for lighting and camera (added in the 2026-09-20 S5 acceptance omission audit, required on every image and unrelated to template family—dimensions not explicitly specified = AI random)：
- Character three-light system: a) Theme object glow tinting the face (explicitly state from which direction the light tints the face, such as "lure glow lighting his face from below-left"); b) cool rim light outlining the character ("strong cool rim light outlining his silhouette"); c) volumetric god rays entering the frame. If any of the three lights is missing, the character will look "flat".
- Forward-extending action perspective: The character must have a first-hand/main propextended toward the camera("toward the viewer/lifting toward the camera"), creating a selfie-style perspective dynamism; static, symmetrical, front-facing standing-pose composition is prohibited.
- Curved arrangement of hanging elements: For surrounding elements, explicitly state "in a loose arc around him, filling the space around him"; rigid top-bottom symmetrical distribution is prohibited.
- Atmospheric particles: At least one of dust particles/bubbles/light beams.
- Universal saturation: Any subject matter (including noir/suspense/crisis dark tones) must maintain HIGH-SATURATION and a brightness baseline; only change hue, do not lower brightness.
2.7 Quality gate (v1.17, field-by-field per-image review)
After each image is generated, review it field by field; if any field is false, only modify the original prompt and regenerate that image (max 2 regenerations per image):
exactTitle(compare at character-by-character zoom, including English spelling)/samePerson／singlePerson／faceProminent(face ≥ 1/6 of the frame)/expressiveAction／expressionContrast／heroObjectClear／compositionIntegrated／thumbnailLegible／noExtraText／noWatermark。
Text typo lesson (C07 record): The published cover's English title was generated as "vibe coking" (should be coding), while the manual review record mistakenly wrote "zero errors"—the AI's text error rate cannot be trusted to manual casual inspection. All text must be checked at zoom character by character, and English words must be spelled out letter by letter for comparison.
2.8 Observation of real Qiuzhi 2046 samples (provided by Jeffrey on 2026-09-20, 9 desktop screenshots, top likes 500k)
- Big-head Q-version compositing: photoreal real-person head (original face) + 3D cartoon body, head-to-body ratio clearly big-headed; not a fully realistic person.
- Extra-large 3D staggered title: yellow/white letter faces + thick black outline + hard shadow, some characters 3D-extruded; arranged in a staggered manner by semantic hierarchy (characters vary in size), occupying the upper 1/3–1/2; mixed Chinese and English ("AI 导演" "Hermes 教程" "抖音 Agent" "超级豆包" "Seed 2.1 实测" "AI 短剧").
- Props = visual translation of the title: holding the main object toward the camera (film projector, glowing phone, notebook, magic hat, tablet).
- High-saturation 3D themed scene: space moon, snow mountain pegasus, rocket launch, beach, magic stage—change the worldview each issue, volumetric light + rim light.
- Character centered and large: bust/half-body occupies 40–60%, wide-eyed open-mouthed surprised expression looking straight at the camera.
- Platform corner badge: place a small badge like "共创/前沿科技激发计划/独家" in the corner (optional, not required).
2.9 Verified baseline sample sheet and standard prompt template (v1.17 final)
-
samples/baseline-cover-03-kv.png= the official baseline approved by Jeffrey ("03KV version is good"),samples/baseline-cover-03-kv-prompt.txtis its complete generation prompt (Seedream 5.0 lite, seed=33, identity reference). -
Standard prompt template structure(for new issues only replace the theme variables, do not change the skeleton):
KV定调(HIGH-SATURATION punchy glossy)→HEADLINE标题合同(两行错落/白+语义色3D挤出/宽88-94%/英文眉题)→CENTER人物合同(photoreal头硬保真+身份分析动态特征+脸16-26%人物48-72%+表情张力+动作短语+主物件=标题视觉翻译)→环绕发光挂件×4(主题元素弧形)→SCENE主题世界观(冷暖对比/体积光/轮廓分离)→文字精确渲染约束+负面清单。 -
After generation, the 2.7 quality gate must be followed (exactTitle zoomed character-by-character comparison is the first checkpoint).
-
Scene narrative paragraph guardrail: target 20–100 characters for scene + action + prop + lighting, hard cap 200 (reject if there is risk of truncation over the limit); three-candidate differentiation = 3 different template families × 3 sceneIds (stage large-screen / handheld handheld prop / dramatic dramatic scene).
-
03 formula conservation (Jeffrey's ruling on 2026-09-21: established after two degradation counterexamples at S5/S20)：
samples/baseline-cover-03-kv-prompt.txtis the master template; for new issuesonly five slots may be replaced——{theme scene worldview}{main object}{title group}{themed costume design}{expression}——the remaining sentences (selfie-style forward-extended arm perspective, HIGH-SATURATION KV tone-setting, three-light system, ratio numbers, negative list)preserve verbatim. Two hard rules: ① the costume must be upgraded and designed according to the theme (e.g., a hoodie with glowing circuit patterns),it is forbidden to reuse the everyday outfit from the reference photo; ② the pose must be dynamically extended forward (selfie-style arm toward the camera / main object raised toward the camera),static compositions such as standing and pointing are forbidden. Negative example archive: S5 detective (symmetrical standing pose) and S20 selection table (plain polo + pointing) were both judged degraded by Jeffrey for breaking the formula. - After generation, it must pass the 2.7 quality gate (exactTitle character-by-character enlarged comparison is the first checkpoint). -
Scene narrative paragraph guardrail: scene + action + prop + lighting target 20—100 characters, hard upper limit 200 (reject if over limit due to truncation risk); three-candidate differentiation = 3 different template families × 3 sceneIds (stage large-screen / handheld prop / dramatic scene).
-
Three-light system sentence pattern (copy verbatim into the prompt during assembly; tested in S20 on 2026-09-21: omitting it loses the light effect)：
LIGHTING ON CHARACTER (mandatory three-light system): 1) strong cool rim light outlining his entire silhouette and catching the edges of his hair, separating him from the background; 2) the {主题光} glow lighting his face from the {方向} (dual-side color wash if split scene); 3) a soft top light keeping his face the brightest area of the whole cover.and retain in the SCENE paragraphvolumetric god raysandfloating dust particles catching the light。 -
Rejected route (Jeffrey 2026-09-20, do not try again): the composite method of "generating a headless body + pasting a real person's photo head" was rejected—strong pasted-on feel, fake head-body connection, and expressions can only reuse the source photo. This skill maintains the single-pass image generation route (Seedream + identity reference), with quality guaranteed by the Section 2.5-2.8 contract and gates (the 03KV sample has been verified as usable).
Shared text constraints:
Render only the following Chinese cover text, exactly as written, with no missing, substituted or extra characters: “[主标题]” and, only when provided, “[辅助短句]”. The Chinese text must be large, crisp, correctly spelled and readable at phone-thumbnail size. Do not render any other words, letters or numbers.
3. Directly generate 1 master version
- Use 1 prompt to call the image provider, and pass in the Jeffrey reference.
- Generate a complete 9:16 image, save as
cover-qiuzhi-9x16.<ext>；extOnly png, jpg, jpeg, or webp are allowed. - The complete image is saved directly to the
05-封面/root directory; do not createbackgrounds/, and do not run a post-layout text script. - If the Chinese text has typos, the person does not look alike, or the theme is drawn incorrectly, only fix the original prompt and regenerate (with the same prompt, reseed at most 2 times).
- Only when the master version is real, decodable, 9:16, and all 3 shared-size final images are real, decodable, and correctly sized, is the ⑥ mechanical threshold satisfied. Character identity, title accuracy, theme consistency, and thumbnail safe zones must be truthfully reported after visual review of each image in the current task, but no new review files or production stages are added.
4. Generate 3 real final images at shared sizes in one pass
At the end of cover-prompts.md , add ## 四平台封面成品, separately write 3 complete generation prompts, call the image provider, and pass in the same Jeffrey reference. The three images must maintain the same visual proposition, main title, person, and main object, but must be recomposed separately according to the real canvas, safe zones, and line breaks:
- 3:4：
cover-qiuzhi-3x4.<ext>, must be 1080×1440; shared to the Douyin vertical cover, WeChat Channels profile card, and Xiaohongshu. - 4:3：
cover-qiuzhi-4x3.<ext>, must be 1440×1080; shared to the Douyin horizontal cover, WeChat Channels share card, and Bilibili homepage recommendation. - 16:9：
cover-qiuzhi-16x9.<ext>, must be 1920×1080; used for the Bilibili personal space.
Only these 3 physical files are generated; platforms with the same aspect ratio reuse them directly; platform attribute differences are handled by their respective titles, descriptions, topics, and publishing fields. The three covers are still bound to the same master version and SHA; no aspect ratio may be simply center-cropped, edge-extended, or passed off as another ratio by dragging the framing on the publishing page.
2.10 Expression and Clothing Variants (a mature solution for changing expressions/outfits, derived from product v2.1)
When the reference photo has a single expression/clothing (e.g., only a smiling ID photo), first use Qwen-Image-Edit-2509 img2img to generate a variant for the current issue, with the standard Chinese sentence pattern:
Keep the face of the same person in the image completely unchanged: {identity_description}. Only change the clothing, expression, and pose: {personAction}. Half-body portrait, facing the camera, natural and realistic portrait photography texture, clean pure dark gray background, even soft lighting. No text, watermark, or second person may appear in the image.
If the variant fails, fall back to the original frame; the variant result also goes through the 2.7 gate. Known trade-off: img2img has a slight AI repaint feel (the priority between this and character fidelity is determined per issue).
2.11 Route Conclusion (full-cycle validation from product v2 to v4; see references/dh-cover-product-lessons.md for details)
Layered compositing (background + real-person cutout + programmatic text) gives exactTitle 100% butthe ceiling of fusion quality cannot reach the Qiuzhi style—the Qiuzhi style's "person integrated into the scene + prop occlusion + unified lighting" can only be achieved by full-image generation (mutually corroborating this skill's "single-pass is the correct solution" conclusion; Brother Zhuzi's realistic-environment style uses layering = laohan-fengmianzhuzige env_photo). Alternative title font: SmileySans-Oblique (OFL, already verified in the APP). Provider has been A/B decided: Seedream 5.0 lite.
-
After generation, the 2.7 quality gate must be passed (character-by-character magnified comparison of exactTitle is the first checkpoint).
-
Scene narrative paragraph guardrail: the scene + action + prop + lighting target is 20–100 characters, with a hard upper limit of 200 (anything over the limit is rejected due to truncation risk); three-candidate differentiation = 3 different template families × 3 sceneIds (stage large-screen / handheld prop / dramatic scene).
-
Rejected route (Jeffrey 2026-09-20, do not try again): the composite approach of "generate a headless body + paste a real-person photo head" was rejected—strong cutout feel, fake head-body connection, and expressions can only reuse the source photo. This skill maintains the single-pass image generation route (Seedream + identity reference), with quality guaranteed by the contract and gate in sections 2.5–2.8 (the 03KV sample has been verified as usable).
Shared text constraints:
Render only the following Chinese cover text, exactly as written, with no missing, substituted or extra characters: “[主标题]” and, only when provided, “[辅助短句]”. The Chinese text must be large, crisp, correctly spelled and readable at phone-thumbnail size. Do not render any other words, letters or numbers.
3. Generate 3 images directly
- Call the image provider separately with the 3 prompts, and pass in the same Jeffrey reference.
- All three generate complete 9:16 images, saved in the recommended order as
cover-01-qiuzhi-9x16.<ext>、cover-02-qiuzhi-9x16.<ext>、cover-03-qiuzhi-9x16.<ext>；extOnly png, jpg, jpeg, or webp are allowed. - Save the complete images directly to the
05-封面/root directory; do not createbackgrounds/, and do not run any post text-layout script. - If there are Chinese typos, the person does not look alike, or the theme is drawn incorrectly, only fix the original prompt and regenerate that image; do not add new processes.
- Only after the three candidates are real, decodable, 9:16, and named in the correct order, and the three shared-size finished images defaulting to 01 are all real, decodable, and correctly sized, is mechanical gate ⑥ satisfied. Character identity, title accuracy, theme consistency, and thumbnail safe area must be visually reviewed one by one in the current task and reported truthfully, but no new review files or production stages are added.
- Jeffrey may explicitly switch the selection to 02 or 03; when there is no explicit switch, release preparation defaults to adapting the platform-required cover ratio from 01, and must not silently switch to 02 or 03.
4. Generate 3 real finished images of shared size at once
At the end of cover-prompts.md , add ## 四平台封面成品, and for the current published rank, write 3 complete generation prompts respectively, call the image provider, and pass in the same Jeffrey reference. The three images keep the same visual proposition, main title, person, and main object, but must be recomposed separately according to the actual canvas, safe zone, and line breaks:
- 3:4：
cover-<rank>-qiuzhi-3x4.<ext>, must be 1080×1440; shared to the Douyin vertical cover, WeChat Channels profile card, and Xiaohongshu. - 4:3：
cover-<rank>-qiuzhi-4x3.<ext>, must be 1440×1080; shared to the Douyin horizontal cover, WeChat Channels share card, and Bilibili homepage recommendation. - 16:9：
cover-<rank>-qiuzhi-16x9.<ext>, must be 1920×1080; used for the Bilibili personal space.
Generate only these 3 physical files, and platforms with the same aspect ratio reuse them directly; differences in platform attributes are carried by their respective titles, descriptions, topics, and publishing fields. The three covers are still bound to the same rank source image and SHA; no aspect ratio may be simply center-cropped, extended, or passed off as another ratio after dragging to frame on the publishing page.
Output
Write to 05-封面/cover-prompts.md：
---
schema_version: 3
script_hash: <01-口播稿.md sha256>
distribution_locked_at: <ISO-8601>
prompt_executor: cover-prompt-strategy
image_provider: <实际provider>
reference_mode: REQUIRED
reference_asset: reference/jeffrey-reference.jpg
reference_sha256: <sha256>
strategy: QIUZHI_SINGLE_MASTER_THREE_SIZES
primary_aspect_ratio: "9:16"
required_generated_candidate_count: 1
required_physical_cover_count: 3
required_logical_cover_usage_count: 7
default_publish_candidate: cover-qiuzhi-9x16
---
## 封面文字
- 主标题：<准确原文与断行>
- 辅助短句：<原文或无>
- 事实绑定：<对应口播稿哪一判断>
## 01｜秋芝方向｜最推荐
- 模板族：<族名>
- 模板：<S编号＋模板名>
- 推荐理由：<为什么最适合本期钩子>
<一条可直接生成完整封面的英文提示词>
## 02｜秋芝方向｜第二推荐
- 模板族：<与01不同的族名>
- 模板：<S编号＋模板名>
- 推荐理由：<为什么排第二>
<一条可直接生成完整封面的英文提示词>
## 03｜秋芝方向｜第三推荐
- 模板族：<与01、02不同的族名>
- 模板：<S编号＋模板名>
- 推荐理由：<为什么排第三>
<一条可直接生成完整封面的英文提示词>
Do not claim to copy any account's CTR or unique style. Here, only the general layout patterns observable in public covers are referenced.