character-sheet-builder

name	character-sheet-builder
author	Noustiny
license	MIT
version	1.0.0
description	Produce a cast sheet of 1–4 named characters from a story seed. Each entry has a canonical-but-IP-free visual description and a hero-portrait prompt ready to drop into a text-to-image model. Downstream: portraits become the reference frames every subsequent beat image will condition on, locking character consistency across the whole storyboard. Fires once per story at canvas-enter, after story-copyright-detector.
metadata	{"hermes":{"tags":["character-sheet","reference-image","portrait-prompt","story-bible","visual-consistency"],"homepage":"https://github.com/noustiny/noustiny-skills","related_skills":["story-copyright-detector","visual-prompt-builder"]}}

Character Sheet Builder

Every text-to-image render reinvents the characters unless the model has a reference to condition on. This skill solves that by producing the character bible up front. One hero portrait per principal character → cached → passed as a reference to every later beat. Result: the same Aang face, the same Katara braid, the same Tony Stark armor, across every frame of the storyboard.

When to use

Fire exactly once when the canvas opens with a fresh seed, immediately after story-copyright-detector tells you the franchise. Never re-fire for the same story.

Input shape

{
  "seed": "Aang opens his eyes for the first time in a hundred years…",
  "franchise": "avatar-airbender",
  "storyRegister": "animated-feature, cel-shaded 2D animation, crisp ink outlines, vibrant saturated palette, Studio Mir-style atmospheric shading",
  "allow_ip_names": true
}

• seed — the full logline
• franchise — slug from the detector (marvel, avatar-airbender, lotr, null, …). null means original seed.
• storyRegister — the visual register token set to append to every portrait_prompt so the character sheet's aesthetic matches the beat-level frames.
• allow_ip_names — when true, the caller's image model accepts franchise IP tokens (typically Gemini / SeeDream on animated franchises). Portrait prompts MAY then include named characters ("Katara from Avatar: The Last Airbender…"). When false (FLUX path / strict-filter models), portrait prompts MUST stay IP-free (the existing "bald monk boy with blue arrow tattoo" sanitised form).

Output contract

Return exactly one JSON array, nothing else. First character [, last ].

[
  {
    "name": "Aang",
    "description": "twelve-year-old bald monk boy, blue arrow tattoo down the brow, wide silver-grey eyes, saffron-and-orange layered robes, slim athletic child frame, bare feet, ever-curious expression",
    "portrait_prompt": "hero portrait, full-body, twelve-year-old bald monk boy with a blue arrow tattoo on his brow, wide silver-grey eyes, saffron-and-orange layered robes, bare feet, light neutral studio backdrop, three-quarter pose, soft key light, animated-feature, cel-shaded 2D animation, crisp ink outlines, vibrant saturated palette"
  },
  {
    "name": "Katara",
    "description": "fourteen-year-old brown-skinned Water Tribe girl, cobalt fur-trimmed parka with ivory trim, dark hair in twin looped front braids, bright blue eyes, hopeful open face",
    "portrait_prompt": "hero portrait, full-body, fourteen-year-old brown-skinned teenage girl, cobalt blue fur-trimmed parka, ivory bone trim, dark hair styled in twin looped front braids, bright blue eyes, light neutral studio backdrop, three-quarter pose, soft key light, animated-feature, cel-shaded 2D animation, crisp ink outlines, vibrant saturated palette"
  },
  {
    "name": "Sokka",
    "description": "fifteen-year-old brown-skinned Water Tribe boy, cobalt fur-trimmed parka, high wolf-tail topknot, sharp jaw, wry skeptical expression, boomerang sheath at his back",
    "portrait_prompt": "hero portrait, full-body, fifteen-year-old brown-skinned teenage boy, cobalt blue fur-trimmed parka, dark hair pulled into a high wolf-tail topknot, wry skeptical expression, boomerang sheath at back, light neutral studio backdrop, three-quarter pose, soft key light, animated-feature, cel-shaded 2D animation, crisp ink outlines, vibrant saturated palette"
  }
]

Rules

• 1 to 4 characters max. Focus on PRINCIPALS actually named or strongly implied in the seed. Skip incidental / unnamed bystanders.
• name — the canonical short name used in the source (first-name only unless ambiguous). Used as the key in the downstream character registry, so it must match whatever the beat prose will reference.
• description — rich physical detail, no IP names, no franchise terms. Thor → thunder-god warrior with blond beard and Norse-style plate armor with red cape, not "Thor". Every description must be reusable verbatim in every subsequent beat.
• portrait_prompt — a full-body hero portrait prompt for a text-to-image model. Must end with the provided storyRegister string verbatim so every portrait and every beat share the same aesthetic grammar.
• IP naming rule — conditional on allow_ip_names:
  • allow_ip_names: true → portrait_prompt MAY name the canonical character ("hero portrait of Katara from Avatar: The Last Airbender, cobalt fur-trimmed parka, twin looped front braids…"). Always follow the name with the rich physical description anyway — the image model gets the best of both if it recognises the name, and the description if it doesn't.
  • allow_ip_names: false → portrait_prompt MUST be fully sanitised (IP-free). "Katara" becomes "fourteen-year-old brown-skinned Water Tribe girl…" every time.
• The description field is ALWAYS the IP-free version (downstream beats reuse it regardless of model). Only the portrait_prompt is conditional.
• For original seeds (franchise: null): invent characters from what the seed implies. If the seed doesn't name anyone, produce one default protagonist + any strongly implied secondary. allow_ip_names is moot for originals — no IP to name anyway.
• Output is strict JSON array. No markdown code fences. No preamble prose ("Here are the characters…"). No trailing notes. First char [, last ].

Anti-patterns

• Do not name an IP character in the portrait_prompt when allow_ip_names: false — that tells the caller the image model will reject named prompts.
• When allow_ip_names: true, do not rely on the name alone — always include the rich physical description as well.
• Do not invent characters not implied by seed or canon.
• Do not exceed 4 entries. A bloated cast sheet wastes image-gen calls.
• Do not output reasoning, step-by-step notes, or "here's my thinking" prose. JSON array only.
• Do not wrap the portrait prompt in quotes or add "--style" flags. Plain prose only.

Worked examples

Marvel Endgame seed

{
  "seed": "Tony Stark holds the Infinity Gauntlet on the Avengers compound battlefield — every possible ending hangs on what he does next.",
  "franchise": "marvel",
  "storyRegister": "superhero-finale, operatic cosmic, IMAX 70mm color grade, high dynamic range, heroic silhouettes"
}

Correct output:

[
  {
    "name": "Tony Stark",
    "description": "lean mid-forties man, dark close-cropped hair, sharp goatee, sharp brown eyes, cracked and battle-worn red-and-gold articulated plate armor with a glowing chest core",
    "portrait_prompt": "hero portrait, full-body, lean mid-forties man with dark close-cropped hair and a sharp goatee, cracked red-and-gold articulated plate armor with a faintly glowing chest core, sharp brown eyes, three-quarter pose, soft key light on a neutral grey backdrop, superhero-finale, operatic cosmic, IMAX 70mm color grade, high dynamic range, heroic silhouettes"
  }
]

Grounded original seed

{
  "seed": "Walt is on the payphone outside the diner in New Hampshire. Everything he is considering would end something.",
  "franchise": null,
  "storyRegister": "quiet literary drama, natural window light, muted earth tones, medium-format stills"
}

Correct output:

[
  {
    "name": "Walt",
    "description": "bald late-fifties man, neat greying goatee, wire-frame glasses, heavy navy winter parka, tired grey-blue eyes, slight hunch at the shoulders",
    "portrait_prompt": "hero portrait, full-body, bald late-fifties man with a neat greying goatee, wire-frame glasses, heavy navy winter parka, tired grey-blue eyes, three-quarter pose, soft key light on a neutral grey backdrop, quiet literary drama, natural window light, muted earth tones, medium-format stills"
  }
]