| name | cinephile-wallpaper |
| description | Generate original high-aesthetic posters and wallpapers for films, TV series, animation, anime, tokusatsu, documentaries, shorts, franchises, characters, props, and scenes. Use for Chinese or English requests to make/generate/create a 海报, 电影海报, 影视海报, 角色海报, poster, wallpaper, or device wallpaper for a named audiovisual work or character, e.g. “给我做一张《杰克奥特曼》的海报”. Always run the fixed guided flow: confirm work identity, ask base settings in one turn, ask image references/design requirements in a second turn, then generate via an image model or provide one unified prompt if no image capability exists. |
Cinephile Wallpaper
Core Idea
Turn a real audiovisual work, character, prop, or scene into an original art-poster wallpaper. Use image generation as the poster-making path. Do not scrape or copy existing posters, do not use HTML/CSS as the poster renderer, and do not deliver old cached outputs unless explicitly asked.
Opening And Mandatory Gates
If the user has not provided a work title, character, prop, scene, or URL, start with exactly this one sentence and nothing else:
请问你想生成哪部电影的海报壁纸?
Do not introduce the skill, list features, explain usage, ask about style, or add examples before this sentence.
If the triggering message already names a work/character/prop/scene, do not ask for it again. Resolve ambiguity if needed, then immediately ask the numbered one-turn base settings. Never research, write a prompt, call an image model, or generate from the first user message alone.
Hard gate: generation is allowed only after these are complete:
- work identity is known and ambiguity is resolved;
- base settings are answered or explicitly accepted as saved defaults;
- the separate image-reference/design-request gate is answered or explicitly skipped.
The base-settings prompt and the image-reference/design-request prompt are two separate assistant turns. Do not merge them.
Workflow
- Read
references/workflow.md for the guided flow, setup prompt, image-reference gate, preference memory, output directory, text variants, and low-token final response.
- Before asking setup, load cross-conversation defaults with
scripts/preferences.mjs get when local script execution is available.
- Confirm the work identity. If ambiguous, ask a concise clarification. If already provided, move to base settings.
- Ask the one-turn base settings. It must include only size/device, output folder, text variant, generation mode, and whether to remember defaults.
- After base settings are answered, ask the separate image-reference/design-request gate. Do not merge it with base settings.
- Only after both gates are complete, research the work enough to verify identity, tone, core characters, props, motifs, and recognition anchors.
- If no user-specified style exists, run
scripts/draw-style.mjs before work interpretation affects style. Use the returned style exactly; work content may shape elements and metaphor, not style choice.
- Run
scripts/draw-density.mjs after identifying the work profile, then build a concise visual brief using the relevant reference files listed below.
- Route generation through an image model when available; otherwise return one unified copyable prompt. Never substitute HTML for image generation.
- Save output image(s), compact prompt package, research notes, and manifest. Follow
references/output-schema.md.
- Final response is poster-first: show/link the image path(s) and one short status line. Do not paste research, visual brief, manifest, or prompt unless requested or prompt-only mode is used.
Non-Negotiables
- Auto-detect interaction language; honor explicit output language.
- Ask only the two mandatory setup turns before research/generation.
- Remember defaults for size/device, output folder, text variant, and generation mode when the user asks or the runtime supports it.
- Cross-conversation defaults must be loaded from persistent JSON before setup, not inferred from chat history.
- Normalize saved defaults before using them; never display or save
中英双语/bilingual as a text variant.
- Remember reusable user design ideas, custom styles, typography preferences, and avoid-lists when the user asks to save them.
- In the base-settings prompt, the text-variant question is only about output variants: with text, no text, or both. Never rewrite it as a language choice such as "Chinese-English bilingual" or any fixed language pair.
- Generate text directly inside the image model when text is requested; keep typography designed, short, and integrated.
- Do not darken or globally filter text versions just to improve text readability.
- Default style selection is weighted random and decoupled from work content. Never explain style as "suitable" for the work unless the user specified it.
- Visual density must be selected by
scripts/draw-density.mjs; use its mode and prompt instruction exactly.
- If depicting recognizable real actor/performer faces, attach actual image references; actor names and descriptions are not enough.
- Uploaded references are identity/shape/atmosphere anchors, not compositing layers. Redesign pose, scale, placement, expression, lighting, and composition.
- Without usable face references, characters may still appear through non-face strategies; do not invent readable substitute faces.
- Film-specific props require prop identity locks and visual references when prominent.
- Every request creates a fresh generation unless the user explicitly asks for an old file.
- Keep user-facing output brief and save details to files.
- Do not expose internal command counts, sandbox/cache failures, or repeated image-refusal retries unless the user asks for debugging details.
- If current screen resolution cannot be concretely detected, use a saved confirmed default or ask; do not infer from previous runs.
- If an image tool refuses a title/character/text variant twice, stop retrying that variant and fall back honestly.
- Classify image-generation failures before fallback: server/transient errors are not refusals, not copyright blocks, and not user-consent requests for CLI/API.
- For branded material styles such as LEGO-like bricks, use generic material language in image prompts and avoid brand logos or official marks.
Quality Bar
Reject or revise outputs that look like:
- generic streaming thumbnails;
- official poster copies;
- text-heavy film database cards;
- low-effort AI fantasy art;
- illegible or ordinary model-generated lettering;
- live-action photorealism;
- copied-still photorealism or celebrity glamour portraiture;
- pasted/cutout-looking reference-image figures, traced poses, copied lighting, copied background, or poster designs that visibly reuse the uploaded reference image as a source layer;
- claiming character-face restoration without attached still/crop reference images;
- generating readable actor/character faces after the user requested characters but no uploaded/acquired reference image can be attached;
- invented character designs that do not match the work;
- cluttered desktop backgrounds;
- conventional illustration when the user asked for bolder art direction;
- obvious reserved blank space that weakens the poster composition;
- visually exhausting all-over element coverage;
- generated text with incorrect title, wrong language, misspellings, or pasted-on default typography.
Resources
- Read
references/workflow.md at the start of each run.
- Read
references/design-memory.md when the user provides design ideas, custom styles, or asks to remember preferences.
- Read
references/device-size.md only for device-size lookup.
- Read
references/model-routing.md before generation-mode/provider decisions.
- Read
references/visual-brief.md when writing the brief; run scripts/draw-density.mjs for density selection.
- Read
references/film-tone.md before palette, brightness, contrast, color temperature, or medium decisions.
- Read
references/style-distillation.md before style routing/prompting; run scripts/draw-style.mjs when style is not user-specified.
- Read
references/character-reference.md when characters, props, or scene/object identity need visual anchoring.
- Read
references/artist-grammars.md only when a stronger poster-design grammar is needed.
- Read
references/output-schema.md when saving outputs.
scripts/set-macos-wallpaper.mjs: macOS desktop setter.