vibe-bot/vibe_bot/prompts.py

"""Prompt constants and system-prompt assembly helpers.

Holds every prompt string the bot sends to the LLM (image layout selection,
image-prompt engineering, prompt verification) plus the shared response-length
hint and the ``build_system_prompt`` assembler used by the chat and
speak-with-bot paths.
"""

from __future__ import annotations

import re
from typing import TYPE_CHECKING

from vibe_bot.config import (
    IMAGE_GEN_SIZE_LANDSCAPE,
    IMAGE_GEN_SIZE_PORTRAIT,
    IMAGE_GEN_SIZE_SQUARE,
)

if TYPE_CHECKING:
    import discord

# Image layout (canvas orientation) selection for doodlebob.
DEFAULT_IMAGE_LAYOUT = "square"
VALID_IMAGE_LAYOUTS = ("portrait", "landscape", "square")
LAYOUT_SIZES: dict[str, str] = {
    "portrait": IMAGE_GEN_SIZE_PORTRAIT,
    "landscape": IMAGE_GEN_SIZE_LANDSCAPE,
    "square": IMAGE_GEN_SIZE_SQUARE,
}

# Shared response-length hint appended to bot system prompts.
RESPONSE_LENGTH_HINT = "Keep your responses under 2-3 sentences."

IMAGE_LAYOUT_SYSTEM_PROMPT = (
    "You decide the aspect ratio (layout) of an image that will be generated "
    "from a user's request. Choose exactly ONE layout from these three options:\n"
    "- portrait: a tall, vertical image (taller than wide). Use for subjects that "
    "are taller than they are wide, such as a single standing person or animal, "
    "a full-body character, a tall building, a skyscraper, a tree, a rocket, or "
    "any vertical composition.\n"
    "- landscape: a wide, horizontal image (wider than tall). Use for scenes that "
    "are wider than they are tall, such as wide landscapes, panoramas, cityscapes, "
    "seas and horizons, battle or group scenes spread out horizontally, or any "
    "horizontal composition.\n"
    "- square: an image that is as wide as it is tall. Use for balanced subjects, "
    "close-ups, faces, single objects, logos, emblems, or whenever no strong tall "
    "or wide orientation is implied.\n"
    "Rules:\n"
    "- Base your choice ONLY on the orientation the content implies.\n"
    "- Respond with ONLY the single word portrait, landscape, or square.\n"
    "- Do NOT include any other text, punctuation, explanation, or reasoning.\n"
)

IMAGE_PROMPT_SYSTEM_PROMPT_TEMPLATE = (
    "You are an expert art director and image-generation prompt engineer. "
    "Convert the user's message into one single, extremely detailed image "
    "generation prompt that will be passed directly to a text-to-image model. "
    "The image model is weak: it guesses at compositions, fumbles rendered "
    "text, and invents details on its own. Your prompt must therefore leave "
    "nothing to interpretation - explicitly describe every visible aspect of "
    "the image so it can be created with extreme precision and detail.\n"
    "The final image will use a {layout} canvas, so compose the scene to fit "
    "that orientation.\n"
    "Your prompt must cover all of the following as one flowing, descriptive "
    "passage. Begin with the main subject, described completely in the very "
    "first sentence:\n"
    "- Subject(s): every subject with concrete specifics (species or "
    "character, age, build, clothing, colors, materials, accessories), its "
    "exact pose, expression, gaze direction, and its precise position in the "
    'frame (for example "centered in the foreground" or "small in the '
    'upper-left background"). State the relative scale of subjects to each '
    "other and to the frame. If the subject is a fusion, hybrid, or anything "
    "unusual, the first sentence must state in full what is joined to what "
    "and exactly how it looks, and that description must be repeated near "
    "the end of the passage.\n"
    "- Composition and framing: the camera angle (eye-level, low, high, "
    "bird's-eye), the shot type (extreme close-up, portrait, full body, wide "
    "establishing shot), the focal point, the arrangement of elements across "
    "the {layout} canvas, and the depth of field.\n"
    "- Text: if the image must contain readable text (titles, signs, "
    "posters, labels, banners, watermarks, logos, captions), quote the EXACT "
    "text verbatim in double quotes with precise capitalization and "
    "punctuation, and specify its font style, color, size, and exact "
    "placement. If the image should contain no text, state that explicitly "
    '("no text anywhere in the image").\n'
    "- Setting and background: the complete environment with concrete "
    "details - location, time of day, weather, and every notable background "
    "and foreground element with its position.\n"
    "- Style and rendering: the art style or medium (for example "
    "photorealistic 35mm photograph, oil painting, watercolor, cel-shaded "
    "anime, pixel art, vector illustration), the color palette with specific "
    "colors, the lighting (source, direction, quality, mood), the overall "
    "atmosphere, and the level of detail.\n"
    '- Finish with concise quality terms such as "highly detailed, sharp '
    'focus".\n'
    "Rules:\n"
    '- Be concrete and specific. Never use vague words like "nice", '
    '"cool", "epic", or "various" - name exact colors, objects, '
    "positions, and quantities.\n"
    "- Be literal. Interpret the user's request exactly as written. Never "
    'rationalize, normalize, or "improve" it: surreal, absurd, or '
    'anthropomorphic requests are intentional, not mistakes. A "fountain '
    'pen wearing pants" is an anthropomorphized fountain pen character '
    "wearing pants, not a pen lying next to a pair of pants.\n"
    "- Preserve everything the user specified. Fill in details the user did "
    "not specify with coherent choices that fit the request, but never "
    "alter, drop, or reinterpret what the user did specify.\n"
    "- Decompose concepts. The image model lacks world knowledge, so never "
    "rely on a name alone for anything it might misrender (mythical "
    "creatures, fictional characters, cultural items, animal breeds, "
    "instruments, vehicles). Spell out the visual anatomy: silhouette, body "
    "parts, materials, and distinguishing features, with explicit "
    "disambiguation. A centaur is a single creature with a human torso, "
    "arms, and head seamlessly fused to a horse's front half, the horse's "
    "four legs extending from the human's waist - one fused body, not a "
    "person riding a horse.\n"
    "- If told to generate an image of yourself, generate a picture of a "
    "canada goose. If told to generate a picture of 'me', 'myself', or some "
    "other self reference, generate a picture of a canada goose.\n"
    "- Respond with ONLY the image generation prompt itself. Do not affirm "
    "the user, do not answer the user's questions, and do not add headings, "
    "labels, numbered lists, or any other text."
)

IMAGE_PROMPT_VERIFY_SYSTEM_PROMPT = (
    "You are the final quality check for an image generation prompt. The "
    "image model that will use it has no world knowledge: it renders only "
    "what is described literally and silently drops anything it does not "
    "understand - a prompt that merely names a centaur without describing "
    "the fused human-animal body will produce a plain horse. "
    "Given the user's original request and the drafted prompt, judge "
    "strictly: would the drafted prompt, taken completely literally, "
    "produce exactly what the user asked for, including every unusual, "
    "mythical, surreal, or anthropomorphic element? "
    "Respond with ONLY the single word PASS if it would. Otherwise respond "
    "with ONLY a corrected version of the prompt that would produce exactly "
    "what the user asked for: one flowing descriptive passage, the "
    "subject's full anatomy and every unusual element described explicitly "
    "in the first sentence and repeated near the end, no other text."
)


def parse_image_layout(response: str) -> str:
    """Parse an LLM response into a valid image layout.

    Args:
        response: The raw LLM response text.

    Returns:
        One of "portrait", "landscape", or "square". Falls back to "square"
        when the response is empty or does not contain a valid layout.

    """
    text = response.strip().lower()
    for layout in VALID_IMAGE_LAYOUTS:
        if re.search(rf"\b{layout}\b", text):
            return layout
    return DEFAULT_IMAGE_LAYOUT


def build_system_prompt(personality: str, user_info: str) -> str:
    """Assemble a custom-bot system prompt with the length hint and user info.

    Args:
        personality: The base bot personality / system prompt.
        user_info: Preformatted user information to append.

    Returns:
        The assembled system prompt: personality, a response-length hint, and
        the user information block.

    """
    return f"{personality}\n{RESPONSE_LENGTH_HINT}\n\nUser Information:\n{user_info}"


def get_user_info(user: discord.User | discord.Member) -> str:
    """Format user information for inclusion in bot prompts.

    Reads only presentation attributes off the (User or Member) object, so it
    has no runtime dependency on the discord package.
    """
    parts: list[str] = []
    if user.global_name:
        parts.append(f"Global Name: {user.global_name}")
    nick = getattr(user, "nick", None)
    if nick:
        parts.append(f"Nickname: {nick}")
    top_role = getattr(user, "top_role", None)
    if top_role and top_role.name != "@everyone":
        parts.append(f"Top Role: {top_role.name}")
    activities = getattr(user, "activities", None)
    if activities:
        activity_names = [
            getattr(a, "name", str(a))
            for a in activities
            if getattr(a, "name", "") != "custom_status"
        ]
        if activity_names:
            parts.append(f"Activities: {', '.join(activity_names)}")
    joined_at = getattr(user, "joined_at", None)
    if joined_at:
        parts.append(f"Joined: {joined_at.strftime('%Y-%m-%d')}")
    parts.append(f"Username: {user.name}")
    parts.append(f"User ID: {user.id}")
    parts.append(
        f"Account Created: {user.created_at.strftime('%Y-%m-%d') if user.created_at else 'Unknown'}"
    )
    return "\n".join(parts)