Files
masi 89aae012c6 feat: wiki sections are a bottom bar on phones, and a form is not an alias
Under 640px the five wiki sections sit in a fixed bar at the bottom with short
labels, instead of a strip that scrolled off the edge. An alias that names a
shape the character takes ("One Eyed Lich form", "Nodulle's body") is now
refused and cleaned up: it marked the words in the text and cast the base
character's portrait for the other shape.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-11 01:20:32 -03:00

468 lines
32 KiB
Python

"""Language-model calls and the prompts. Providers (MiniMax through Headroom, Ollama Cloud)
are tried in the order configured in config.LLM_PROVIDERS; requests with an image go to the
first provider with a vision model. The scene job and the pre-reader share the rules for
characters, looks and the glossary, so both record story knowledge the same way."""
import base64
import io
import json
import sys
import urllib.request
from PIL import Image
from config import LLM_PROVIDERS
def _anthropic(provider, system, user, image_b64):
content = user
if image_b64:
content = [{"type": "text", "text": user},
{"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": image_b64}}]
body = {"model": provider["vision_model"] if image_b64 else provider["text_model"], "system": system,
"messages": [{"role": "user", "content": content}],
# M2.7 reasons before it answers and sometimes uses the whole budget doing it,
# which came back as max_tokens with no text at all; the ceiling leaves room.
"max_tokens": provider.get("max_tokens", 32768)}
if provider.get("think") == "off":
body["thinking"] = {"type": "disabled"}
req = urllib.request.Request(provider["url"], json.dumps(body).encode(), {
"x-api-key": provider["key"], "anthropic-version": "2023-06-01", "Content-Type": "application/json"})
with urllib.request.urlopen(req, timeout=300) as res:
response = json.load(res)
return "\n".join(block["text"] for block in response["content"] if block.get("type") == "text")
def _ollama(provider, system, user, image_b64):
message = {"role": "user", "content": user}
if image_b64:
message["images"] = [image_b64]
body = {"model": provider["vision_model"] if image_b64 else provider["text_model"], "stream": False,
"think": provider.get("think", "on") != "off",
"messages": [{"role": "system", "content": system}, message]}
req = urllib.request.Request(provider["url"], json.dumps(body).encode(), {
"Authorization": f"Bearer {provider['key']}", "Content-Type": "application/json"})
with urllib.request.urlopen(req, timeout=300) as res:
return json.load(res)["message"]["content"]
APIS = {"anthropic": _anthropic, "ollama": _ollama}
def chat(system, user, image_b64=None, retries=2):
"""One JSON answer from the first configured provider that gives one; each provider gets
`retries` extra attempts before the next is tried."""
candidates = [p for p in LLM_PROVIDERS if not image_b64 or p.get("vision_model")]
if not candidates:
raise RuntimeError("no configured LLM provider accepts images")
for n, provider in enumerate(candidates):
for attempt in range(retries + 1):
try:
answer = APIS[provider["api"]](provider, system, user, image_b64)
# The model may wrap the JSON in prose; take the outermost object.
return json.loads(answer[answer.find("{"):answer.rfind("}") + 1])
except (json.JSONDecodeError, OSError, KeyError, TypeError) as e:
if attempt < retries:
continue
if n == len(candidates) - 1:
raise
print(f"LLM provider {provider['name']} failed ({type(e).__name__}: {str(e)[:200]}); "
f"trying {candidates[n + 1]['name']}", file=sys.stderr)
def image_b64(path, width=832):
"""A JPEG for the vision model: smaller than the PNG/WebP and universally accepted."""
with Image.open(path) as im:
im = im.convert("RGB")
im.thumbnail((width, width))
buf = io.BytesIO()
im.save(buf, "JPEG", quality=85)
return base64.b64encode(buf.getvalue()).decode()
STYLE_SYSTEM = """You describe the art style of a web-novel cover so an image model can reproduce it.
This text is appended to EVERY illustration of the novel, so describe only HOW things are drawn:
medium, brushwork or line work, rendering, colour palette and saturation, how light and shadow
are treated, texture, mood; max 60 words. Never name subject matter: no characters, creatures,
landscapes, mountains, skies, weather, buildings, objects, energy, auras, glows, particles,
petals or other effects, and no title text. Each scene supplies its own content.
Answer only JSON: {"style": "..."}"""
LOOK_RULES = """Looks: known characters are listed with their named looks and the look active before this
chapter. The story may jump in time (regression loops, flashbacks, reincarnation, time resets),
so decide from the text how each known character looks at the time THIS chapter takes place.
In "appearances", list every known character who appears in the chapter text (match them
by any name, title or alias; the same person is often called "Director Kim" early and by their
full name later) with "aliases": names or titles this chapter uses for them that are not listed
yet, and "look": the label of the existing look that matches. An alias is what people call them;
a shape they take ("lich form", "another person's body") is a look, never an alias.
A look is how the character LOOKS to an illustrator, never their power, rank or state. These are
NOT looks: cultivation breakthroughs, realms and stages, titles and positions, moods, auras and
techniques, wounds that heal, battle damage, a single outfit, and disguises, illusions or
avatars used briefly. Create a new look (a new short label plus "description": their full lasting
appearance, same fields as for new characters, max 60 words) only when their lasting visible
appearance changes: body, face, apparent age, species or form, a permanent mark or lost limb, a
lasting change of usual attire. Label looks by appearance ("young, long black hair", "dragon
form", "one-armed"), not by the event. Alternate forms a character switches between (a dragon's
human form, a beast or true form) are looks: add each form once and reuse it. After a
regression or time reset, return to the earlier look. Prefer reusing a look; most characters
need one to three looks in a whole novel. Set "flashback": true when the look only applies to a
flashback, dream or vision in this chapter."""
CHARACTER_RULE = """Characters are specific individuals the story treats as characters: someone with a proper name,
or a unique epithet the story consistently uses for one particular person or being ("the fox
ruling the forest", "the hunchbacked old man"). Never list groups ("Byeok Clan elders",
"cultivator trio"), one-off people described only by appearance ("a bald tattooed cultivator"),
translators, editors, or authors of notes and afterwords."""
CHARACTER_FIELDS = """"description": their lasting appearance from the text (species, sex, apparent age,
build, face, hair, eyes, usual clothing, distinguishing marks; max 60 words; invent plausible
details consistent with the story if the text is silent), "aliases": other names or titles the
text uses for them, and "role": who they are in the story so far, in one sentence that reveals
nothing beyond this chapter"""
GLOSSARY_RULES = """Glossary: in "terms", list up to five proper names of the story's world from this chapter that a
reader may need explained later (cultivation realms and stages, sects, clans, techniques,
artifacts, places) and that are NOT in "Known terms", each with "meaning": one neutral sentence
based only on the text so far. Skip ordinary words and descriptions ("deep sea", "blood
corpse"), one-off names, and character names. An empty list is fine.
Summary: in "summary", what happens in this chapter in at most three plain sentences."""
SCENE_SYSTEM_V1 = f"""You design one illustration for a chapter of a web novel.
Pick the single most visually striking moment of the chapter that does not spoil its ending.
The image model draws anatomy, contact and crowds badly, so compose around its weaknesses without
turning every illustration into a posed character portrait:
- At most TWO main figures (a creature counts). Anyone else is optional, far in the background,
small, at most three, never identical to each other; prefer leaving them out.
- A scene can be environment-only when the place, scale, aftermath or phenomenon is the point of
the moment. Do not add a character just to fill the frame. A small figure seen from behind or
in silhouette can show scale when the text supports it.
- Choose a natural, story-specific viewpoint instead of defaulting to a centered, front-facing
portrait: a wide establishing view, high or low angle, over-the-shoulder view, side profile,
rear view, foreground-framed view, or a closer shot when the emotion warrants it. Vary framing
between illustrations. Characters look toward the person, object, action or distant view that
matters in the scene, not toward the viewer unless the text gives a reason.
- Use a grounded, readable pose that fits the moment. A figure may walk, turn, brace, kneel,
hold one simple object, or perform one clear action. A bow is held or raised, never drawn
with a nocked arrow (the model breaks the hands and string). Carry drama with staging,
body direction, gaze, lighting and setting, not complicated contact. Avoid mid-air poses,
falling, extreme lunges, twisted torsos, flailing limbs, or limbs/hands pointing at the camera.
- When a figure matters, specify their body orientation, what they are doing, and where they look;
do not require a visible face or prescribe relaxed arms when the moment calls for another pose.
Keep limbs clearly separated from bodies and from each other.
- No contact between figures and at least an arm's length of empty space between them: no
grappling, hugging, touching, no hand on a shoulder, no reaching toward each other, and no
object handed or passed between figures (show it held by one of them). Show the tension
before the action (a standoff, a stare, a turned back) instead of the action itself.
- Use a wide or environmental composition for vast places, grand visions, journeys, and moments
where scale matters. For cultivation or other progression fantasy, the chapter's described
mountains, halls, skies, formations or other setting can dominate the frame, with characters
small or absent. Do not invent setting details or effects that the chapter does not support.
- State the camera viewpoint, the focal point, and the figure's placement only as needed for the
selected composition. Keep one clear focal point and a readable foreground, middle distance,
and background when the scene benefits from depth.
- Restrained gore: at most a few drops or a stain. No text, symbols, speech bubbles or UI.
Write the image prompt in this order: framing; main figure(s) with appearance, clothing and
pose; secondary element; setting; lighting. Max 100 words. Ignore translator notes, links and
author comments.
"anchor": 6 to 12 consecutive words copied EXACTLY (same spelling and punctuation) from one
sentence of the chapter text where the chosen moment happens; the image is placed there.
{LOOK_RULES} For a character who is drawn, use the look they have in the illustrated moment.
Named characters: when a character with a proper name in the story is drawn, write [[Name]]
where that figure is introduced, using the canonical name from "Known characters" when it is
the same person (match aliases too), then describe only pose and expression for them, not their
looks. List every [[Name]] you used in "characters". List in "distant" every [[Name]] drawn
small, far away, from behind or in silhouette: they get no reference portrait, which the image
model would otherwise paste in large as an extra figure. For a character NOT in "Known characters",
also give {CHARACTER_FIELDS}. {CHARACTER_RULE} Unnamed extras and creatures are described
inline, never listed, and never written as [[...]].
{GLOSSARY_RULES}
Answer only JSON: {{"moment": "one sentence", "anchor": "...", "prompt": "...",
"characters": [{{"name": "...", "aliases": ["..."], "role": "...", "description": "..."}}],
"distant": ["..."],
"appearances": [{{"name": "...", "aliases": [], "look": "...", "description": "only for a new look", "flashback": false}}],
"terms": [{{"term": "...", "meaning": "..."}}], "summary": "..."}}"""
SCENE_SYSTEM_V2 = f"""You design one illustration for a chapter of a web novel, staged like a page of a top
action manhwa or comic. Pick the single most visually striking moment of the chapter that does not
spoil its ending.
Hard rules (the image model draws these badly):
- At most TWO main figures (a creature counts). Anyone else is optional, small and far in the
background, at most three, never identical to each other; prefer leaving them out.
- No contact between figures: no grappling, hugging, biting, touching, no blades meeting or
striking a body. Show the instant before or after (a drawn blade, a reaching hand, a stare-down,
a recoil) with a visible gap between figures.
- Keep limbs clearly separated from the body and from each other. Hands simple and readable, or
hidden behind an object. No limb pointing at the camera, no tumbling, falling, flailing or
extreme torso twists.
- Never a lineup: no three or more figures standing side by side facing the viewer, and no two
figures both facing the camera. Stagger figures at different depths and different angles.
- Restrained gore: at most a few drops or a stain. No text, symbols, speech bubbles or UI.
Staging, which is what makes an illustration dynamic:
- Name the shot first, chosen to fit the moment and varied between illustrations: low-angle hero
shot, high-angle shot, over-the-shoulder, medium close-up, side profile, rear three-quarter view,
wide establishing shot, or a foreground-framed view (through a doorway, branches, a shoulder).
- Renderable poses to choose from: weight shifted onto one leg, mid-stride with the back heel
lifted, three-quarter turn, looking back over a shoulder, crouched ready stance with a hand near a
weapon, kneeling on one knee, leaning on a wall or railing, slumped sitting, drawing a blade from
its sheath, a weapon held low and angled with a clear silhouette, a raised forearm shielding the
face, bracing behind a shield, reaching toward something out of reach.
- For every main figure write exactly one action verb + body direction + gaze target, for example
"steps forward on the left foot, body turned three-quarters left, eyes on the knight's raised
sword". Gaze goes to the person, object or event that matters, not to the viewer, unless the
text gives a reason.
- Carry the drama with the camera angle, a strong silhouette, diagonal lines, light and setting
rather than with complicated contact.
- A scene can be environment-only, or have tiny figures, when the place, scale, aftermath or
phenomenon is the point (vast places, grand visions, journeys, progression-fantasy mountains,
halls, skies). Do not add a character just to fill the frame, and do not invent setting details
or effects the chapter does not support.
- One clear focal point, with a readable foreground, middle distance and background when depth helps.
Write the image prompt in this order: shot and camera angle; main figure(s) with appearance,
clothing and their action verb + body direction + gaze; secondary element; setting; lighting.
Max 110 words. Ignore translator notes, links and author comments.
"anchor": 6 to 12 consecutive words copied EXACTLY (same spelling and punctuation) from one
sentence of the chapter text where the chosen moment happens; the image is placed there.
{LOOK_RULES} For a character who is drawn, use the look they have in the illustrated moment.
Named characters: when a character with a proper name in the story is drawn, write [[Name]]
where that figure is introduced, using the canonical name from "Known characters" when it is
the same person (match aliases too), then describe only pose and expression for them, not their
looks. List every [[Name]] you used in "characters". For a character NOT in "Known characters",
also give {CHARACTER_FIELDS}. {CHARACTER_RULE} Unnamed extras and creatures are described
inline, never listed, and never written as [[...]].
{GLOSSARY_RULES}
Answer only JSON: {{"moment": "one sentence", "anchor": "...", "prompt": "...",
"characters": [{{"name": "...", "aliases": ["..."], "role": "...", "description": "..."}}],
"appearances": [{{"name": "...", "aliases": [], "look": "...", "description": "only for a new look", "flashback": false}}],
"terms": [{{"term": "...", "meaning": "..."}}], "summary": "..."}}"""
SCENE_SYSTEM = SCENE_SYSTEM_V1 # switched to SCENE_SYSTEM_V2 only if docs/research/bench supports it
INDEX_SYSTEM = f"""You read a web novel chapter by chapter and keep its character list and glossary up to date,
for a reader who sees only what they have already read.
New characters: in "characters", list every named character who appears in this chapter (on
page, not only mentioned in passing) and is NOT in "Known characters", with {CHARACTER_FIELDS}.
{CHARACTER_RULE}
Skip characters already known under any name, title or alias: "Director Kim" and
"Kim Young-hoon" are the same person when the text says so or the context makes it clear.
{LOOK_RULES}
{GLOSSARY_RULES}
Illustration value: "visual" rates from 1 to 10 how much this chapter deserves an illustration:
10 for a major transformation or new form, the first appearance of an important character or
creature, a climactic confrontation or a breathtaking new place; 1 for talk, travel or training
with nothing new to see. Most chapters are 2 to 5. "key_moment": that moment in one sentence.
Answer only JSON: {{"visual": 5, "key_moment": "...", "characters": [{{"name": "...", "aliases": ["..."], "role": "...", "description": "..."}}],
"appearances": [{{"name": "...", "aliases": [], "look": "...", "description": "only for a new look", "flashback": false}}],
"terms": [{{"term": "...", "meaning": "..."}}], "summary": "..."}}"""
PORTRAIT_PROMPT = """Character reference image of one single character: full body from head to
toe, standing in a relaxed neutral pose, facing the viewer at a slight three-quarter angle, arms
relaxed at the sides, hands visible and simple, calm expression, centered, on a simple softly tinted
background of one flat muted colour that suits the character (no scenery, no props, no floor
pattern), even soft lighting, no other figures, no text. {description} Art style: {style}"""
# Animals and creatures: no word about arms, hands or posture of a person, which turned a colt
# into a horse-headed man in a shirt.
CREATURE_PORTRAIT_PROMPT = """Reference image of one single animal or creature, its whole body visible
in profile at a slight three-quarter angle: {description} True to its own anatomy, standing calmly on
all of its own legs in its natural stance, without clothing or any human features unless described.
Centered, on a simple softly tinted background of one flat muted colour (no scenery, no props, no floor
pattern), even soft lighting, no other figures, no text. Art style: {style}"""
BODY_SYSTEM = """Decide how to draw a story character's reference picture from its description.
"person": a human or a humanoid that stands upright with arms and hands (elves, dwarves, orcs,
undead knights, liches, anthropomorphic beings). "creature": an animal or any being without a
person's upright body (horses, dragons, wolves, birds, griffons, slimes, giant insects).
Answer only JSON: {"body": "person"} or {"body": "creature"}"""
FEATURES_SYSTEM = """Describe the character in this reference image so an artist can redraw them
consistently: species, sex, apparent age, build, face, hair, eyes, skin or fur, clothing,
accessories, distinguishing marks. Max 60 words, concrete visual facts only. Describe only the
character itself: ignore the background, lighting, auras, energy and other effects around them.
Answer only JSON: {"description": "..."}"""
QA_SYSTEM = """You inspect a generated illustration for defects before a reader sees it.
You get the scene description the image was made from. Reject the image ("ok": false) only for
clear defects: a figure that is duplicated or cloned; bodies merged or fused together; a human
with animal or creature parts (or a creature with human parts) that the description does not
call for; extra, missing or badly twisted limbs; a clearly malformed face or hand in the
foreground; more main figures than described; visible text, letters or watermarks. Painterly
roughness, loose brushwork and unclear small background details are fine.
Answer only JSON: {"ok": true, "problems": ["short description of each defect"]}"""
RECAP_SYSTEM = """You write a short "previously on" recap for a reader returning to a web novel after a break.
Using only the chapter summaries given (in order), recap the most recent events in 4 to 6 plain
sentences (at most 120 words), most important threads first, ending where the summaries end. Never add anything that
is not in the summaries. Answer only JSON: {"recap": "..."}"""
# ---- interstitials -------------------------------------------------------------
# One wide establishing shot where the story moves somewhere new, printed between the
# chapters the way a comic prints a full-page spread at the end of an arc.
SPLASH_SYSTEM = """You paint one wide establishing shot for a web novel: the page a comic book
prints between two arcs, showing where the story has arrived.
The place is the subject, not the people:
- One wide or panoramic view of the setting, from a viewpoint that shows its scale: a ridge
above a valley, a street seen down its length, a hall from the back of a crowd, a shore from
the water. Nothing fills the frame edge to edge with a face.
- At most one figure, small and seen from behind, in silhouette or at a distance, to show the
scale of the place, and only when the text supports it. Never a close-up, never a portrait.
- Describe the place with what the text actually says about it: the buildings, ground, sky,
weather, distance, light, and the sense of size. Do not invent landmarks, banners, monsters
or weather the text does not mention.
- Warm, lived-in and specific rather than grand and empty. No text, symbols, speech bubbles,
borders or panels.
- No named character is drawn, so no reference images are used and no [[Name]] appears.
Write the image prompt in this order: framing and viewpoint; the place in the middle distance;
the foreground; the sky and light. Max 100 words. Ignore translator notes, links and author
comments.
"moment": one sentence saying what place this is, as a caption for the picture.
"anchor": 6 to 12 consecutive words copied EXACTLY (same spelling and punctuation) from one
sentence of the chapter text that describes the place; the image is placed there.
Answer only JSON: {"moment": "one sentence", "anchor": "...", "prompt": "..."}"""
# A place inside the chapter rather than a spread between two: same rules about the subject,
# but the shot belongs to the moment the text describes it, and it must not repeat the others.
SCENERY_SYSTEM = SPLASH_SYSTEM.replace(
"""You paint one wide establishing shot for a web novel: the page a comic book
prints between two arcs, showing where the story has arrived.""",
"""You paint one place-only picture for a chapter of a web novel: the panel a comic
book cuts to for the setting, the weather or an object before the scene goes on. Pick a
place, weather, object or atmosphere the chapter text actually describes.""").replace(
"""- At most one figure, small and seen from behind, in silhouette or at a distance, to show the
scale of the place, and only when the text supports it. Never a close-up, never a portrait.""",
"""- Leave the people out. Whatever the characters are doing in the text, paint the place they
are in without them: no guards, no crowds, no one carving or fighting. At most one tiny
figure far away, in silhouette from behind, and often none. No face is ever visible, and no
two figures are alike. Never a close-up, never a portrait.""")
assert "Leave the people out" not in SPLASH_SYSTEM
ARC_SYSTEM = """You are the art editor of a web-novel reader. You see the summaries of the
chapters of one book, in order, each labelled with its chapter number, and you choose where a
wide establishing illustration belongs: the page a comic prints between two arcs, showing the
place the story has just arrived at.
Choose a chapter when the story visibly changes place or opens a new stretch of its own:
- a journey arriving somewhere (a village, a city, a capital, a fortress, a dungeon, an island);
- a new region, kingdom or continent the party enters;
- the first chapter of a new arc, a new tournament, war, hunt or trial;
- a place first shown that the story then stays in for a while.
Do NOT choose a chapter that stays where the last one was, however eventful it is, and do not
choose two chapters close together: at most one every 4 chapters, and never in the first 2
chapters given. Prefer a place over an event.
Answer only JSON: {"chapters": [chapter numbers, in order, each given in the input]}"""
REDRAW_SYSTEM = """You update a character's appearance description for an illustrator.
Apply the reader's correction to the current description and return the full corrected
description (species, sex, apparent age, build, face, hair, eyes, usual clothing, distinguishing
marks; max 60 words). The correction wins wherever it conflicts.
Answer only JSON: {"description": "..."}"""
CONSOLIDATE_SYSTEM = """You tidy a web-novel character's list of looks for an illustrator. Many entries were
recorded wrongly: power levels, realms, titles, injuries, one-off outfits, brief disguises or
avatars. A look is only how the character LOOKS: body, face, apparent age, species or form,
permanent marks, lasting usual attire. Group the given looks (listed with the chapter each was
first used) into as few real looks as the descriptions justify, usually one to four. Every given
label must belong to exactly one group. Give each group a short appearance-based "label" and a
"description" of that lasting appearance (max 60 words: species, sex, apparent age, build, face,
hair, eyes, usual clothing, distinguishing marks).
Answer only JSON: {"looks": [{"label": "...", "description": "...", "members": ["old label", "..."]}]}"""
PRUNE_SYSTEM = """You clean a web novel's character list. Return the entries that are NOT one specific individual:
groups ("Byeok Clan elders", "cultivator trio"), one-off people described only by appearance
or role ("bald tattooed cultivator", "Corpse-like man"), translators, editors and authors of notes
(e.g. "BlueMangoAde", "TranslatingNovice"), and entries that are not characters at all. Keep
anyone with a proper name, and beings the story refers to by a unique epithet ("the fox",
"Black Dragon King", "the hunchbacked old man"); a script keeps recurring ones anyway.
Answer only JSON: {"remove": ["exact name as given", "..."]}"""
DUPLICATE_SYSTEM = """You check a web novel's cast list for people recorded twice. You get one
entry per line as "name (chapters it appears in): the description recorded for it". Group the
entries that are the SAME individual written differently: transliteration or word-order variants
of one name ("Jong-bum Choi" / "Choi Jong Bom"), a name with a title or epithet attached ("Rok
Elder" / "Elder Rok"), a shortened form ("Karichwi" / "Karichwi the Elder"), a name and the name
the story later gives them after they are revealed.
Never group: a father and a son, or any two people who only share a surname; a person and a place,
a building, a clan, a group, a title on its own ("Elder", "Emperor", "the High Priest"); an animal
or monster of a species with the individual ("Wy" / "Wy-3"); two people merely similar-looking.
When in doubt, leave the names apart: a wrongly merged character loses their own face and history.
Answer only JSON: {"groups": [["name as given", "other name as given"], ...]} with every name
copied exactly from the input. Leave out every name that stands alone."""
LOOK_SYSTEM = """You match a character's lost look label to the looks that are still on file for them.
You get the lost label and the remaining looks, one per line as "label: the description recorded
for it". Answer with the ONE label that describes the same appearance: the lost label is usually
the same words written differently ("White-flame Chaos Warrior" against "Chaos Warrior form"), or
the same person's usual form named again ("Lee Hyun base form" against the description of a lean
twenty-year-old Korean man).
Answer "" if none of them is the same appearance, rather than the closest one. Two different
forms of one person are not interchangeable: a lich, a dragon and the man he was are three looks,
and sending one chapter's worth of drawings to another's description ruins both.
Answer only JSON: {"look": "label as given", "why": "one short sentence"}"""
STYLE_OPTIONS_SYSTEM = """You are an art director choosing an illustration style for a web novel from a fixed
library of presets. You get the novel's title, genres, current style, some chapter summaries, and the
preset list, one per line as "id | name | family | inspired_by".
Pick the THREE presets that fit the novel's genre, tone and setting best. They must be clearly different
from each other (different families or looks), and none may equal the current style. Choose by how a preset
is drawn (line work, rendering, palette, lighting, mood) against what the story needs: for example gritty
inked looks for grim martial or horror stories, bright cel looks for comedic action, painterly or watercolour
looks for quiet, epic or pastoral fantasy, clean line looks for adventure. Use ids exactly as listed; never
invent one. For each give "why": one concrete sentence tying the look to this story. "tweak" is optional: a
short extra phrase (at most 15 words) describing only HOW things are drawn (palette, lighting, texture), or an
empty string; never subject matter, characters, objects or effects.
Answer only JSON: {"options": [{"id": "...", "why": "...", "tweak": ""}, {"id": "...", "why": "...", "tweak": ""},
{"id": "...", "why": "...", "tweak": ""}]} with exactly three entries."""
FEEDBACK_SYSTEM = """You triage a reader's feedback on an illustrated web-novel reader and pick at most one fix.
You get the feedback, what it is about (an illustration, a character card, a glossary term, or the app in
general) and the facts the reader saw. Choose the single action that best addresses it, only when the
feedback clearly asks for it and the needed details are in the facts; never invent names, looks or terms:
- "repaint": the illustration is wrong (wrong look, pose, setting, mood, a defect, a misread moment).
"fix" is a precise instruction for the painter, in the reader's sense, e.g. "Kim has short black hair,
not long white hair". Only for feedback about an illustration.
- "redraw_portrait": a character's portrait or look is wrong. "name" and "label" must be one of the
character's looks in the facts; "hint" describes the correction.
- "edit_character": a character's name, role or aliases are wrong. "name" must exist; give only the
fields to change: "new_name", "role", "add_aliases" (a list).
- "merge": two entries are the same person. "name" is the duplicate, "into" the entry to keep; both must exist.
- "edit_term": a glossary meaning is wrong. "term" must exist; "meaning" is the corrected meaning.
- "none": anything else (a bug, an idea, praise, unclear or unsafe requests). Someone will review it.
"reply" is one short, friendly sentence to the reader saying what will happen.
Answer only JSON: {"action": "...", "fix": "", "name": "", "label": "", "hint": "", "new_name": "",
"role": "", "add_aliases": [], "into": "", "term": "", "meaning": "", "reply": "..."}"""
REVISE_SYSTEM = """You revise an image-generation prompt for a web-novel illustration after reader feedback.
Keep everything that is not affected (moment, setting, characters, camera, style rules) and change only
what the feedback asks for, stated positively and concretely. Keep any wording about reference images and
placeholders exactly as it is.
Also decide whether the fix is "local": it changes only details of the existing picture (an object, a
weapon, a hand, an expression, a piece of clothing, a colour, removing one item) without moving, adding or
removing figures and without changing the camera or the moment. For a local fix write "edit": one
instruction for an image editor saying exactly what to change in the picture, concretely; otherwise "".
Answer only JSON: {"prompt": "...", "local": true, "edit": "..."}"""