name: creative-research-first description: 创意研究优先工具：在重要的图像、视频、角色、场景、品牌、时尚、建筑、音乐或声音创作前，先完成有目标的网络研究与参考验证。用于避免凭空创作和提高参考质量。 veRSIon: 1.0.0

Creative Research First

Bottom-Layer Reasoning

APPly sophia-mode retrieval routing and think-one-step-further:

• Inspect supplied artifACTs and local truth before browsing.

• Search only the unresolved knowledge gap.

• Prefer current primary and official sources for platforms, products, technical behavior, lAWS, and specifications.

• Translate sources into creative controls rather than copying reference language.

• Keep research proportional to the task.

Core Intent

For substantive creative work, research before ideation and Prompt writing:

brief -> known assets -> knowledge gaps -> web research -> mechanism extraction -> creative synthesis -> prompt -> source note

Research is not a decorative bibliography. It must change the design, story, sound, camera, material, action, or generation strategy.

Broad Reference Atlas Mode

When the user asks for "大批量通用参考", "不要只适用于这个", "动作库/审美库/转场库", or a reusable creative knowledge base, widen the research scope before nARRowing to the current project.

• Build a reusable atlas across multiple lanes: action/martial/dance/stunt, camera/editing, Aesthetics/lookdev, transitions/VFX, sound, material/light, production design/worldbuilding, character/weapon/vehicle/creature design.

• Include adjacent genres and media even when they are not immediately useful for the current prompt; label them as future-use routes.

• Extract every source as a mechanism and prompt control, not as a title list.

• After the atlas exists, select a small subset for the current project. Do not let the current project erase useful General references.

aesthetic Research Gate

For any non-trivial base aesthetic / visual master / premium style / 画面质感 / 审美方向 decision, do not rely on mEMOry, generic taste words, or a few familiar names. Run a focused but broad enough reference sweep before writing the visual master.

Minimum sweep:

• search 3-5 distinct reference lanes, such as film cinematography, animation surface treatment, production design, photography/editorial finish, VFX/compositing, game/CG lookdev, graphic design, or relevant niche artists/studios;

• inspect more than the first obvious result when taste is open or the user complAIns the result is cheap;

• compare at least 4-8 candidate anchors when the aesthetic is still unresolved;

• keep only the smallest compatible stack after research.

Extract aesthetic references as mechanisms, not as name lists:

reference anchor -> what it controls -> visible mechanism
-> prompt language -> risk / what not to import

If a reference is useful only for one layer, scope it tightly:

Fortiche / Arcane -> hand-painted 3D surface and character material hierarchy
John Wick 4 -> large-space hard light and practical color pools
The Raid -> readable action geography and contact proof
Spider-Verse -> sub-0.15s graphic impact frames only, not comic palette drift

Do not let a famous name override the project's existing style, story, action physics, continuity, or user's deliberate anchors.

Cultural / Design Sedimentation Gate

For every substantial original character, world, faction, costume, weapon, creature, vehicle, ritual, profession, action system, or style-led image/video prompt, run a small web research pass before final prompt writing unless the user explicitly says not to browse or the task is a trivial wording edit.

Do this as a default universal step, not only when the user complains that one specific image "缺少文化". The goal is to add sedimentation: cultural memory, craft logic, material history, body technique, social use, symbolic consequence, and design lineage.

Minimum useful pass:

• search the relevant culture/craft/object/movement/profession/design tradition, preferably using museum, academic, practitioner, official, or high-quality reference sources;

• extract 3-6 usable mechanisms, not decorations: silhouette logic, material process, construction, ritual/social use, movement principle, color ownership, motif placement, wear/weathering, and forbidden CLIché;

• translate findings into original design controls: costume structure, weapon function, scene rules, action rhythm, light/material response, and negative constraints;

• save or cite compact sources when browsing materially changes the result.

Avoid shallow cultural sticker work:

• do not paste random symbols, patterns, sacred objects, scripts, or national clichés onto a generic pretty character;

• do not make every character ancient, mythic, religious, or ornate just because culture was researched;

• do not let researched references overwrite the user's chosen style or current character direction;

• when the user wants modern, cute, sexy, sci-fi, street, PV, or action-showcase work, still research the relevant design lineage and material/lifestyle logic, then express it lightly.

When Research Is Required

Browse before final prompt creation when:

• the user asks for a new creative concept, scene, character, cover, world, fight, vehicle, script, score, or sound plan;

• the task references a profession, historical period, culture, real object, physical mechanism, musical tradition, production process, or unfamiliar aesthetic;

• the task is an original character/weapon/costume/worldbuilding prompt and would otherwise risk being merely "pretty" without cultural, craft, material, or social depth;

• current AI model/platform controls matter;

• the user asks for high quality, authenticity, advanced aesthetics, Professional terminology, or broad generality;

• the user asks for or critiques base aesthetics, 画面质感, 审美, 高级感, style bible, visual master, non-generic style, or says the result feels cheap;

• previous attempts failed and external references can clarify the mechanism.

Skip or minimize browsing when:

• the user asks for a tiny wording change;

• the task is a direct transformation of complete supplied material;

• current local files are the authoritative source;

• the user explicitly says not to browse;

• research would expose private material or add no decision value.

Research Lanes

Use the smallest necessary combination:

1. Domain truth: how the real object, craft, profession, culture, music, movement, or system works.

2. Reference mechanics: examples that solve composition, motion, sound, structure, lighting, material, editing, or emotional problems.

3. Platform truth: current official model capabilities, modes, limits, reference syntax, duration, aspect ratio, audio, and controls.

4. Audience/context: platform conventions, genre expectations, accessibility, or use case when relevant.

Workflow

1. Inventory:

• user references;

• current project files;

• active capsules;

• known constraints.

2. Write 2-5 research questions that would materially change the ouTPUt.

• When the user knows only part of the relevant vocabulary, include an anchor-gap question: which director, studio, film, animator, cinematographer, martial art, dance, stunt, weapon system, or sound reference best names the missing function?

• For aesthetic work, include at least one baseline taste question: what reference systems solve the desired finish, lighting, texture, color hierarchy, and anti-cheapness better than the current prompt?

3. Search:

• use official/primary sources for current technical facts;

• use credible Educational, institutional, museum, studio, or practitioner sources for craft;

• compare multiple examples for taste and aesthetics.

4. Extract each useful finding as:

source fact/reference -> mechanism -> visible/audible control -> prompt language -> risk/boundary

5. Synthesize:

• combine mechanics into an original direction;

• do not imitate one living artist or copy one copyrighted scene.

6. Write the prompt only after the research brief is coherent.

7. Cite a compact source list when browsing was used.

Read references/research-playbook.md for query and extraction templates.

Output Discipline

For normal final delivery, keep the research compact:

研究提炼：
- ...

参考机制：
- ...

完整提示词：
...

来源：
- link

Do not bury the prompt under a long research report unless the user asks for one.

Coordination

• Use $creative-workbench-toolkit when research is about open-source tools, GitHub projects, assistant infrastructure, long-term workflow stacks, reference libraries, prompt evaluation, automation, or whether something should be installed/adopted.

• Use $reference-hunting-board for visual, motion, film, photography, action, and material reference boards.

• Use $cinematic-music-sound-design for music and sound research.

• Use official model documentation before $jimeng-sd2-prompting or other platform-specific packAGIng.

• Use $visual-style-aesthetic-direction, $production-design-worldbuilding, $screenwriting-story-craft, or $cinematic-audiovisual-language after research establishes the relevant mechanics.

• Use $capsule-engine only after a researched mechanism is validated in real work or explicitly adopted as architecture policy.

Guardrails

• Do not browse broadly without research questions.

• Do not treat search ranking as authority.

• Do not use outdated model tutorials when current official documentation exists.

• Do not copy prompts, lyrics, scripts, visual layouts, or proprietary workflows verbatim.

• Use named references as high-density creative anchors when they genuinely fit. Verify their relevance, preserve the selected name, and add mechanics only as clarification or guardrails.

• Do not force research findings into the prompt when they conflict with the user's brief.

• Do not claim a source proves an aesthetic judgment; distinguish evidence, precedent, and taste.