Dylan Cover Generator

Create a 3:4 vertical social cover through a confirmation-first flow. Infer the best theme, title, value label, and topic elements from the user’s actual content, show the proposed cover plan, and generate only after explicit confirmation. Use exactly three non-blendable themes: 科技资讯风格, 纸张拼贴风, and 粗粝编辑风. Treat assets/style-references/ as style references only; the platform-like counts have been removed.

Persistent character library

Keep reusable character inputs in assets/characters/; the registry is assets/characters/registry.json.

• On the first upload, register the local image with scripts/character_registry.py add --id <id> --image <path> --name <display-name> --set-default.
• If the user says “use my character” and exactly one profile exists, use it automatically. If multiple profiles exist, use the registry default; if no default exists, ask which one to use.
• For “切换形象 / switch character”, run scripts/character_registry.py use <id>. Never overwrite another profile silently.
• Preserve the original upload as the identity anchor. Do not replace it with a generated output.
• Resolve the character in this order: explicit name in the current request → registry default → only registered profile → ask.

Composition selection

Choose one primary archetype and keep its skeleton intact. Do not blend all six layouts together.

Archetype	Best for	Composition
impact-explosion (style-01)	challenge, transformation, “打爆/制霸” claims	oversized 2–3 line title; character bursts from center; flying topic objects; strong bottom anchor
skill-battle (style-02)	skill gap and technical mastery	stacked title; intense character in lower half; cracked/radial dark background; one tool cue
choice-ui (style-03)	comparisons and “which one” questions	question headline; thoughtful central character; floating app windows/cards around the body
versus-comparison (style-04)	A-vs-B and competing products	headline top; character between two opposing subjects; balanced left/right weight
tutorial-hand-off (style-05)	tutorials, workflows, scripts, data	headline top; device/panel pushed toward camera; 3–4 labeled foreground cards
reveal-board (style-06)	showcases, before/after, AI results	character left or lower-left and pointing; screen/grid/evidence board on the right

Theme selection

Choose exactly one theme before selecting the rest of the visual language:

Theme	Input value	Automatic recommendation
科技资讯风格	tech-news	technical news, product operation, tutorials, comparisons, results, or strong high-impact visuals
纸张拼贴风	paper-collage	product dry goods, concept explainers, practical lessons, or topics needing tactile visual evidence
粗粝编辑风	editorial-grit	product judgment, B2B analysis, decision frameworks, or serious editorial content

When a user says “style” or “风格”, map that request to theme and use one of the three values above.

• If the user does not name a theme, infer one from the content and present it as a recommendation; never silently choose a theme and generate.
• If the user asks for a tactile, paper, collage, sticky-note, workshop, or approachable product explainer, use paper-collage.
• If the user asks for an editorial, rough paper, magazine, B2B, product judgment, or serious analysis look, use editorial-grit.
• If the user asks for technical news, product operation, a tutorial, a comparison, a result showcase, or a strong high-impact visual, use tech-news.
• If the content is a concept explanation, model introduction, definition, short lesson, or one-scenario application without a subject request, recommend paper-collage.
• Never blend the themes. Each theme keeps its own composition skeleton.
• paper-collage uses a saturated blue field, tactile torn-paper layers, a yellow series strip, a cream headline sheet, a paper duration sticker, one software window, and 2–3 sticky-note role cards. Keep the collage dimensional and energetic.
• editorial-grit uses a black or charcoal paper-grain field, cream oversized typography, orange rules/labels, blue geometric blocks, one large software panel, and editorial callout cards. Keep the layout analytical and asymmetric; do not turn it into a generic neon tech poster.
• For abstract models such as the Fogg Behavior Model, use a compact three-node/arrow motif or a behavior-curve motif as the only topic element. The visual should suggest the model, not turn the cover into a lecture diagram.

Visual contract

1. Use a 3:4 vertical canvas for every output, regardless of theme. Verify the rendered width-to-height ratio is exactly 3:4 before delivery; reject square, landscape, or near-4:5 outputs. For tech-news, reserve the top 28–38% for the headline and keep the face out of that text block. For paper-collage and editorial-grit, reserve enough top space for the series strip/label and headline without shrinking the title below phone readability.
2. For tech-news, use a near-black/navy base with blue, red, magenta, or warm rim lighting and controlled depth elements. For paper-collage, use a saturated blue or cobalt field with paper texture, torn edges, and tactile shadows. For editorial-grit, use black/charcoal grain, cream paper typography, orange accents, and restrained blue blocks.
3. Make every text element use one fixed typeface: Source Han Sans SC Heavy (思源黑体 Heavy). Keep the typeface and weight unchanged between the optional series name, time badge, and main headline; never switch fonts for different durations or colors. Make the headline huge, condensed, extra-bold, and phone-readable. Use white plus one or two accent colors, thick outline/shadow, and no more than 2–3 lines.
4. For tech-news, preserve the selected character’s identity, hair, face, proportions, and signature clothing; change pose and expression only for the topic. For paper-collage and editorial-grit, omit the character entirely unless the user explicitly requests one.
5. Turn the topic into one or two concrete visual metaphors such as an app window, device, document, robot, screen, chart, or card.
6. Keep a clear foreground/midground/background hierarchy for tech-news. For paper-collage, build foreground paper, midground software window, and background field layers. For editorial-grit, use an asymmetric editorial grid with the software panel and callout cards at different depths.

Value point plus headline

Prefer a two-text hierarchy whenever the topic has a clear time, efficiency, result, or benefit promise:

• value_point: a short independent value label such as 1分钟, 省下3小时, or 直接看结果.
• title: the single main question or claim, such as 如何判断需求价值？.

Treat value_point as a primary visual promise, not as a weak subtitle. Give it strong contrast, a warm accent color, a badge/ribbon/outlined panel, and enough scale and spacing to read immediately at phone size. Keep the main headline large and dominant below or beside it; do not merge the two texts into one line.

Extract the exact user-provided value point and never invent a number, duration, result, or benefit claim. If a phrase such as “1分钟告诉你……” appears, extract 1分钟 as the value point and remove that lead-in from the proposed title. If no value point is provided and the topic clearly needs one, derive only a non-quantified phrase from the user’s wording; otherwise mark it as needing confirmation. By default, the cover contains exactly these two required text elements and no subtitle or extra labels. A user-provided series_name is the only optional third text element.

Fixed typeface and duration badge

When value_point is a duration, apply this fixed visual system:

Duration	Badge color	Meaning
≤1分钟	green #43B96B	relaxed and easy
>1且<5分钟	yellow #F2C94C	serious and focused
≥5分钟	red #D94A4A	urgent and higher effort

• Use Source Han Sans SC Heavy (思源黑体 Heavy) for the optional series name, time badge, and main headline. Do not substitute another typeface, italic style, or condensed variant between covers. If the image renderer cannot reliably typeset the font, leave clean text space and typeset all approved text afterward with this same font.
• Keep the time badge in one fixed slot for tech-news. On a 1080×1440 canvas, use x=86px, y=562px, width=346px, height=122px; scale these coordinates proportionally for other 3:4 sizes. For paper-collage, the mapped duration color may use a tactile paper circle/sticker placed beside the software window; for editorial-grit, it may use an orange clipped label below the headline. Preserve the mapped duration color and phone readability.
• If the headline collides with the fixed badge slot, shorten the headline, reduce its line spacing, or reduce its size within the existing hierarchy. Never move the time badge to another corner or reposition it based on the duration.

Optional series name

• series_name is optional. If the user does not provide one, omit it entirely and preserve the two-text layout.
• If provided, use the exact user text; never invent, shorten, number, or turn it into a slogan.
• Apply a theme-specific fixed placement while keeping the label subordinate. For tech-news, use a single-line label in the upper-left corner at approximately x=6%, y=3.5%. For paper-collage, place it inside the yellow torn-paper strip above the headline. For editorial-grit, place it as a small cream label above the headline beside the orange rule. Scale the slot proportionally for other 3:4 sizes.
• Keep it visibly subordinate: use the same Source Han Sans SC Heavy typeface at roughly 30–40% of the headline size, with restrained contrast and no badge, glow, outline, perspective, or decorative treatment. It must not compete with the title or time badge.
• series_name does not count toward the 12-visible-character title limit. Check it for clipping and collision, but never enlarge the series name to solve a layout problem.

Title length, text, and exclusions

Enforce a hard title limit of 12 visible characters. Count Chinese characters, Latin letters, digits, and punctuation; ignore spaces and line breaks. The separate value_point and optional series_name do not count toward this title limit, but keep both concise and phone-readable. Check the required text elements and the optional series name, when present, in the final rendered cover before every generation.

• If the requested title is longer than 12 visible characters, compress it to the shortest faithful question or claim before generating. Preserve the core subject and tension; remove subtitles first.
• Prefer one value point plus one headline, with no subtitle. Do not put explanatory copy, method lists, warnings, UI labels, or decorative pseudo-text into the image unless the user explicitly asks for them.
• Examples that satisfy the limit: AI怎么操作网页？, 什么网页任务适合AI？, 不是所有网页任务.

Treat user-supplied copy as immutable once it has been compressed or approved. Never invent a claim, product name, number, or comparison winner.

Include the exact title in the generation prompt with line breaks and color assignment. If the renderer mangles Chinese or English lettering, regenerate with clean text space and typeset the copy in a separate post-process.

Never include heart icons, like counts, comment counts, share buttons, usernames, view counts, progress bars, watermarks, feed borders, or other platform chrome. Do not add logos unless explicitly requested or necessary to the topic.

Pre-generation confirmation gate

This gate is mandatory for every cover request, including requests with an explicit theme, title, or value_point. Do not call the image-generation tool, create a final cover, or claim that a cover is ready before the user confirms the proposed plan.

1. Read the user’s actual content: raw topic, script, title idea, duration phrase, requested audience, and visual constraints.
2. Infer the theme using the rules above. Explicit user constraints win; otherwise use the topic intent. If the evidence is mixed, state the uncertainty and recommend one theme rather than hiding the ambiguity.
3. Extract the value label exactly when present. Preserve the user’s duration or benefit wording; do not invent a number. Map a duration to its badge color.
4. Propose one main title. Compress only enough to satisfy the 12-visible-character limit, preserve the core subject and tension, and show the exact proposed copy for approval.
5. Propose the approximate topic elements. Keep them concrete and sparse: for example, a Fogg Behavior Model topic may use three connected nodes, a behavior curve, or a prompt dot. Do not add explanatory labels unless requested.
6. Show this confirmation card in the response and stop:

我准备使用以下封面方案：

主题：[科技资讯风格 / 纸张拼贴风 / 粗粝编辑风]
系列名：[用户提供的系列名 / 不显示]
主标题：[不超过12个可见字符]
时间/价值标签：[用户原文中的标签]（[颜色，如适用]）
主体形象：[使用 / 不使用]
主题元素：[1–3个具体元素]
背景与构图：[一句话说明]

请确认，或告诉我需要修改哪一项。

7. Treat “确认”“可以”“按这个生成” and equivalent approval as permission to continue. If the user asks to change the theme, title, label, elements, or layout, update the confirmation card and stop again. A vague acknowledgment that does not approve the plan is not enough.

Example for the user’s Fogg topic:

主题：纸张拼贴风
主标题：福格模型影响AI设计
时间/价值标签：1分钟（绿色）
主体形象：不使用
主题元素：三个相连节点 + 一条行为曲线
背景与构图：钴蓝纸张底，黄色撕纸条与米白标题纸分层，三个节点和行为曲线作为核心视觉

Generation workflow

1. Complete the pre-generation confirmation gate and receive explicit approval. Do not skip this step because the user supplied structured fields.
2. Parse the approved theme, optional series_name, value_point, title, topic, archetype, character, aspect-ratio, and output-format; use 3:4 and the fixed Source Han Sans SC Heavy typeface unless the user approved another supported output constraint.
3. If value_point is a duration, classify it with the fixed badge map and reserve the fixed badge slot before composing the headline.
4. If theme=tech-news, choose the closest single archetype and inspect its bundled reference image. If theme=paper-collage or theme=editorial-grit, skip character resolution and use the selected theme’s paper/editorial skeleton. Read references/style-anatomy.md for the selected theme’s decisions.
5. Resolve or register the character only for approved tech-news; never let a character in a style reference silently override the selected profile. Never introduce a character into paper-collage or editorial-grit without explicit approval.
6. Write one concise prompt containing only the approved theme, optional series name, copy, fixed font, fixed badge coordinates, title layout, and allowed topic elements for that theme. If no series name was approved, explicitly omit it.
7. Generate with the image-generation tool. For tech-news, include the selected character anchor and chosen style reference only. For paper-collage and editorial-grit, include no character; use no style reference unless the approved plan explicitly names one.
8. Inspect at phone-size scale and verify the output dimensions are exactly 3:4. Regenerate if the aspect ratio, theme boundaries, approved copy, title hierarchy, font consistency, time-badge position/color, topic motif, or clean bottom edge is unclear.
9. Return the image and state the approved theme. State the archetype and character profile only for tech-news.

Prompt template

Create a polished 3:4 vertical social-media cover using the user-approved plan: theme [THEME],
value point [VALUE_POINT], optional series name [SERIES_NAME], title [TITLE], and topic motif
[MOTIF]. Do not add or alter any approved text or visual element. If `SERIES_NAME` is absent,
do not add a series label.

For theme `tech-news`: use composition archetype [one archetype and reference file] and
keep its layout [title zone, subject placement, object placement, depth direction]. Use
exactly two required text elements and no subtitle: value point “[VALUE_POINT]” as a large,
high-priority badge in the mapped color, locked to the fixed slot x=8%, y=39%, width=32%,
height=8.5% of the 3:4 canvas; exact main headline “[TITLE]” on [line breaks], with [color
hierarchy], using the same fixed typeface Source Han Sans SC Heavy (思源黑体 Heavy), huge
extra-bold lettering, thick dark shadow and slight perspective. Use character profile
“[PROFILE]”; preserve identity, hair, face, proportions, and clothing, with [pose/expression].
Visualize [TOPIC] using [one or two concrete objects]. Use a near-black/navy background,
directional rim light, controlled glow, cinematic contrast, and crisp phone-size readability.
If `SERIES_NAME` is present, render the exact series name as one small, single-line,
low-contrast label in the fixed upper-left corner around x=6%, y=3.5%, using the same font at
30–40% of the headline size, with no badge, glow, outline, or perspective. Keep it subordinate.

For theme `paper-collage`: use a tactile 3:4 torn-paper collage. Use a saturated blue/cobalt background, a yellow torn-paper strip containing the exact
series name “[SERIES_NAME]” when present, a large cream torn-paper sheet for the exact main
headline “[TITLE]”, and a mapped-color paper duration sticker “[VALUE_POINT]”. Place one
abstract software window on a layered paper sheet and connect it to 2–3 sticky-note cards
using only approved topic labels. Use tactile paper grain, torn edges, tape, realistic paper
shadows, and a strong foreground/midground/background hierarchy. Keep the title dominant and
phone-readable. No character, photo, cinematic neon, platform chrome, or extra words.

For theme `editorial-grit`: use a rough black/charcoal paper-grain editorial cover. Render the
exact series name “[SERIES_NAME]” as a small cream label above the headline beside a narrow
orange rule when present. Use a huge cream main headline “[TITLE]”, an orange clipped duration
label “[VALUE_POINT]”, one large software panel offset to the right, and 2–3 editorial callout
cards on the left using only approved topic labels. Add restrained blue blocks, orange dots,
thin connector lines, asymmetric magazine spacing, and tactile grain. Keep the design serious,
analytical, and product-focused. No character, photo, generic neon glow, platform chrome, or
extra words.

## Bundled resources

- `assets/style-references/style-01.jpg` through `style-06.jpg`: cropped composition references with platform count strips removed.
- `assets/characters/registry.json`: persistent selected-character registry.
- `scripts/character_registry.py`: add, list, and select character profiles.
- `references/style-anatomy.md`: visual anatomy, theme selection, and selection notes.