wtf is ai-sticker-pack?

Ai-sticker-pack is not a ready to run application. It is a written explanation of an architecture pattern for giving any chatbot the ability to pick and send its own stickers or meme images at the right moment in a conversation, along with code snippets you are meant to adapt rather than copy directly. The author is upfront that the backend language, chat platform, and vision model are all up to you, what this repository hands you is the idea and a list of mistakes already made so you do not repeat them. The core problem it solves is that a chatbot cannot look at an image and know when it is funny or appropriate to send, and asking a vision model to look at every candidate sticker on every single message would be slow and expensive. The solution splits understanding from using. When you first upload a sticker, a vision model looks at it exactly once and writes a short name and a one sentence description covering any text on the image, its emotional tone, and when it would fit. That name and description get stored in a small database table alongside the image's URL. From then on, the chatbot never looks at the image again. It only ever sees the short text description injected into its prompt, and when it wants to send a sticker it writes a marker like [sticker:name] in its reply, which the backend catches and swaps for the actual image before sending. The write up walks through four pieces: a simple stickers table with id, name, URL, and description columns, a one time vision call at upload time using any OpenAI compatible vision model, a function that formats stored descriptions into a prompt block for the bot, and backend code that intercepts the sticker marker to render an inline image or call Telegram's sendPhoto API. A large section covers practical pitfalls: Telegram's sendPhoto needs a publicly reachable image URL with no authentication, text and sticker markers can arrive as separate messages if not batched together, sticker images look better floating outside a chat bubble than crammed inside one, images should be size capped so they do not dominate the screen, and streaming multi-byte UTF-8 text in chunks can corrupt characters if concatenated the wrong way. The project is released under the MIT license, though it notes that copyright of the sticker images themselves belongs to their original creators.