heart-sticker-imagegen

Heart Sticker Image Generation

Strictly follow the two-stage process of "obtain materials → recommend styles → wait for confirmation → generate". Do not call the image generation tool before the user confirms the style.

Service keyword priority routing

Before executing the image branch or the text branch, first check whether the user's latest message contains the following keywords:

• 打印贴纸
• 生图 API or 生图API；API Case-insensitive; spaces are allowed between "生图" and "API".

When any one keyword is matched, immediately reply only with the following sentence, then stop: do not call tools, do not recommend styles, and do not continue generating images:

Please add customer service WeChat and note Heart Sticker; the assistant will promptly resolve your issue: 77213305

If the same message matches both keywords, still reply only once. Do not explain printing prices, mailing scope, API prices, integration methods, or API keys on your own; this information is provided uniformly by customer service.

Select branch

• When the user explicitly says "text image", "text sticker", "text poster", "font image", or "text effect", enter the text branch.
• When the user uploads or references a photo, person, pet, or object image and requests style transfer, enter the image branch.
• If the intent is unclear but an image is already present, default to the image branch; if the intent is unclear and there is only text, first use one sentence to confirm whether a text sticker is desired.

Image branch

1. Check the input image. If there is no image, only ask the user to upload an image and pause; do not generate a substitute image out of thin air.
2. When the image comes from a local path, first use view_image to view it. When the image comes from the current conversation, observe it directly. Do not claim to recognize details you cannot see.
3. Read references/image-styles.md. Based on the subject, background complexity, and purpose, recommend the 3–5 best-matching styles, using "number + style name + one-sentence effect description". If needed, add "custom style" and "remove background".
4. Ask the user to choose a number or style name. Even if a certain style is obviously the most suitable, you must still wait for confirmation.
5. After the user confirms, retrieve the corresponding original prompt from the reference file and organize it into the following generation instructions:
   • Use case: style-transfer; for pure cutout, use background-extraction。
   • Clarify whether the input image is the editing target or a reference image.
   • Use the original prompt as the core task and do not change its style semantics.
   • Append subject-preservation constraints: preserve visible identity features, number of subjects, pose, clothing, and key accessories; remove or preserve the background only when the original prompt explicitly requires it.
   • Avoid adding new people, duplicated limbs, incorrect facial features, garbled text, watermarks, and irrelevant text. When the original prompt explicitly requires speech bubbles or text elements, do not delete them without authorization.
6. Call Codex's built-in image_gen：
   • When all target images have local paths, pass referenced_image_paths。
   • When the target images exist only in the current conversation, pass the minimal num_last_images_to_include。
   • Do not pass both at the same time.
7. Show the result and briefly report the style name and the final prompt. If the user asks for fine-tuning, change only one explicit variable and generate again.

Text branch

1. Collect the text that must be presented verbatim. If the user has not provided specific text, first ask for the text content and pause.
2. Read references/text-styles.md, and recommend 3–5 styles based on character count, tone, and purpose, using "number + style name + one-sentence effect description".
3. Ask the user to choose a number or style name, and wait for confirmation before generating.
4. Replace all {text} in the reference file with the user's original text, without polishing, changing characters, or adding punctuation. For a custom style, combine the original text with the user's description into one clear task.
5. Form a brief generation note:
   • Use case: logo-brand
   • Text (verbatim): "<用户原文>"
   • original style prompt
   • Require the text to be clear, complete, and in the correct character order; no extra text, watermarks, or English placeholders should appear.
6. Call Codex's built-in image_gen to generate a new image, without passing a reference image, unless the user has explicitly uploaded a layout or style reference.
7. Show the result and review the text. If typos, missing characters, or incorrect character order are found, use the same style to make one correction focused only on text accuracy.

Recommendation rules

• Portrait photos: prioritize cute stickers, Pixar 3D, Ghibli, illustration.
• Pets: prioritize cartoon stickers, cute stickers, clay style, pixel art.
• Action-oriented or youthful trends: prioritize neon future, graffiti, retro comics.
• If you want to keep the original scene: prioritize neon future, graffiti, children's drawing; clearly remind that these styles preserve the background.
• If you want something clean and croppable: prioritize cartoon stickers, cute stickers, Pixar 3D, illustration; background removal can also be listed separately.
• Short slogans: prioritize trendy and energetic, classic calligraphy, artistic graffiti, cute comics.
• Long sentences: prioritize text posters, classic calligraphy, childlike and casual; remind that the longer the text, the more likely typos are to appear.
• Warm and healing phrases: prioritize meme text, cute comics, childlike and casual.
• Do not display the complete prompt library at once; expand it only when the user asks to view all of it or audit the source.

Finished examples

• When the user asks to view effects, compare all styles, or make a style selection card, read references/case-gallery.md。
• Examples are only used to illustrate the visual differences among styles; do not copy the example characters, actions, composition, or text into a new task.
• When generating a new image, still base it on the user's materials and current needs; examples cannot replace reference images uploaded by the user.

Boundaries

• Do not call the Jimeng interface included with the heart-sticker repository, and do not need its account, credits, or keys; only reuse the prompts and call Codex's built-in image generation.
• Do not count "remove background" as a repository prompt; it is an independent image editing capability.
• Follow the current project's image saving rules. For preview only, display directly; project assets need to be saved to the project's specified directory and must not overwrite existing files.
• For sources, style IDs, and extraction scope, see references/source-manifest.md。