ai-draw-cue-word-project

Build high-consistency AI character prompts across five model syntax families (natural-language / conversational / MJ-Niji / SD / domestic API) and multi-format outputs (image / storyboard / comic panel / video / 3D) using a weight & ratio precision-control workbench. Ships character-anchor sheets, composition & shots references, posture-emotion mapping, reference-image capability matrix, multi-character spatial relations, storyboard/panel layout rules, speech-bubble positioning, LoRA management, temporal-consistency checklist, 3D generation matrix, video prompt template, conversational edit chain, negative-word bank, and a P0/P1/P2 validation checklist. All example characters in this skill are adult (18+) women; no minors allowed.

AI Drawing Composition Template (ai-draw-cue-word-project)

A quality-control workbench for generating character prompts that stay consistent across models and across a series of images. The setting layer (anchors, composition, validation) is model-agnostic; the syntax layer is adapted per model family.

Safety & Content Policy (read before use)

• No sexualization of minors. Never generate, refine, or optimize imagery that sexualizes, eroticizes, or exploits anyone depicted as a minor or under 18. This applies to every model family and every syntax flow in this skill.
• Adult characters by default. All example characters in this skill (including Nanwang and Song Jiuning) are depicted as adult women (18+). Do not reintroduce minor ages ("16-year-old", "teen girl", "loli", "少女", "未成年") into anchors, prompts, or logs.
• No mixing of age stages. A character version is either adult-only or absent; never use cross-stage head-to-body ratios, childlike features, school uniforms, or any other visual cues historically associated with minors in adult-character anchors.
• Per-version adult declaration. Every new row added to CharacterAnchors.csv / FeatureDetails.csv / NaturalLanguagePromptTemplate.csv must include an "Adult (18+)" cell. Prompts missing this cell are invalid and must be regenerated.
• User responsibility. You are solely responsible for ensuring generated content complies with applicable laws and platform policies in your jurisdiction. This skill is a prompt-engineering workbench, not legal or safety clearance.
• Fictional / non-sexual character design only. The workbench is intended for fictional, non-sexual character-concept and illustration work. If you are unsure whether a request is appropriate, do not generate it.

v2.9.0 safety-patch changelog (2026-09-01)

• Hardened Safety & Content Policy: added No mixing of age stages and Per-version adult declaration rules.
• All example characters explicitly declared as adult (18+) women.
• Old SKILL.md.bak-20260829 soft-deleted to .workbuddy/.trash/2026-09-01/.

When to use

Use this skill when the user needs to generate or refine an AI image prompt for a character — especially when consistency across multiple images or across different models matters (illustrations, storyboards, character cards, series).

Workflow (run through once per generation)

1. Character anchors — read references/CharacterAnchors.csv. Pick or create an image version row: identity anchor (≥1.6), temperament tag (≥1.3), palette, signature props, negative anchoring.
2. Composition & shots — read references/CompositionAndShots.csv. Fill the Section A config (aspect ratio, composition type, camera angle, shot scale, focal length, lighting, layers, occlusion). Choose composition type from the Composition Type Reference table in references/NOTES.md (Section B). Look up posture by temperament in the Posture-Emotion Mapping table (references/NOTES.md Section D) — never free-hand it.
3. Model & reference strategy — read references/ReferenceImageCapabilityMatrix.csv to pick the model's consistency method (--cref, LoRA, native reference, or verbatim anchor reuse for DALL-E 3). For series work, follow the reference-first workflow in references/ReferenceFirstWorkflow.md (P0).
4. Assemble the prompt — read references/NaturalLanguagePromptTemplate.csv for the syntax of the target family (see the five rows below).
5. Validate — run the P0/P1/P2 checks in references/Checklist.csv. Any P0 failure (identity anchor, core features, anchor-description reuse, negative coverage) means regenerate.
6. Batch production — for series / production runs, follow the pipeline in references/ProductionPipeline.md (P0): 12-image pre-flight, drift labels, retry rules, versioned reference pack.

Five syntax families

Family	Key rules
Natural language (Flux / DALL-E 3 / SD3)	Full English paragraph: scene+shot → anchors → posture → lighting → background. No weight brackets, no plus signs. Anchor sentences first, verbatim across images.
Conversational (GPT-4o / Gemini 2.5 Flash Image)	Complete description in round 1, then single-point fixes in conversation. Never change anchor sentences. First image acts as the native reference.
MJ / Niji	Core anchors in the first 20 tokens; CLI params appended at the end (--ar, --cref, --cw, --s, --niji). --no filters only 4-6 highest-risk words.
Stable Diffusion 1.5 / XL / SD3	Comma-separated tags: identity anchors (1.5+) → core features (1.3-1.5) → posture/composition → lighting (0.7-0.9). Use (tag:weight). Separate Negative Prompt box.
Domestic API (Jimeng / Kling / Doubao / Tongyi Wanxiang)	Chinese natural-language paragraph + API params (aspect_ratio / size). Negative-word and reference-image support vary per vendor — follow official docs.

Weight semantics (SD/MJ numeric baseline)

Weight	Meaning
1.6-2.0	Hard anchor — absolutely immutable, force-kept in every scene
1.3-1.5	Core feature — kept in all scenes, validated as P0
1.0-1.2	Baseline — default strength, fluctuate ≤0.2
0.7-0.9	Atmosphere — adjustable per scene
<0.7	Weak prompt — atmosphere only, never carries key info

v2.6–v2.8 multi-format outputs (image / storyboard / comic panel / video / 3D)

Beyond single-image generation, v2.6–v2.8 adds sheets for serialized and cross-modal work:

Output	Key sheet	Rules
Multi-character scenes	references/MultiCharacterSpatialRelations.csv	Subject size hierarchy, occlusion, gaze direction, relative positioning anchors
Comic panels / pages	references/PanelLayoutRules.csv, references/SpeechBubblePositioning.csv	Panel grid baselines (4-panel / 6-panel / cinematic strip / splash / vertical scroll), bubble placement by speech type (7 types)
Storyboard sequences	references/StoryboardSkeletonTemplate.csv, references/TemporalConsistencyChecklist.csv	Shot-to-shot continuity: 10 panel roles, 3 camera rules, 4 page layouts; 14 temporal checks across 6 categories
Conversational iterative editing (GPT-4o / Gemini)	references/ConversationalEditChain.csv	Single-point edits only, anchor sentences frozen, edit log structure
LoRA management	references/LoRAManagement.csv	Trigger tokens, weight baselines, version pins, conflict detection
Video generation	references/VideoGenerationPromptTemplate.csv	Motion verbs, camera trajectory, duration baselines per model family
3D generation	references/ThreeDGenerationMatrix.csv	Model-specific syntax for 3D asset outputs (Turntable / orthographic / mesh)

v2.7 comic-mode edition

Added PanelLayoutRules (5 layout types), SpeechBubblePositioning (7 bubble types), and TemporalConsistencyChecklist (14 check items across 6 categories). Total sheet count 18→21.

v2.8 comic-workflow closure

Interactive workbench deployed with bilingual (zh/en) comic mode workflow: storyboard skeleton → panel layout → speech bubble positioning → temporal consistency check. All 21 CSV sheets verified against deployed index.html data.

Cross-model red lines

• (word:1.3) works only on SD; MJ uses token position; natural-language models forbid bracket weights.
• MJ --no ≤ 4-6 short words; SD has a Negative Prompt box; natural-language and domestic API models rely on affirmative writing first, negation as fallback.
• Anchor sentences must be reused verbatim — synonym rewriting breaks consistency.
• DALL-E 3 is deprecated (v2.5): no reference-image support; for series or consistency-critical scenes, route to a conversational model (GPT-4o / Gemini) instead of compensating with description only.
• Reference-first (v2.5): for any multi-image series, prepare a reference pack and follow references/ReferenceFirstWorkflow.md; text anchors are the second layer of defense, never the first.

Files

• references/*.csv — the 21 workbench sheets (flat tables):
  • Core (v2.5 baseline): CharacterAnchors, CompositionAndShots, AspectRatioBaselines, FeatureDetails, ModelsAndReferences, ReferenceImageCapabilityMatrix, NaturalLanguagePromptTemplate, NegativeWordBank, Checklist (P0/P1/P2), Usage, ChangeLog, GenerationIterationLog
  • Multi-format (v2.6): ConversationalEditChain, LoRAManagement, MultiCharacterSpatialRelations, PanelLayoutRules, SpeechBubblePositioning, StoryboardSkeletonTemplate, TemporalConsistencyChecklist, ThreeDGenerationMatrix, VideoGenerationPromptTemplate
• references/NOTES.md — supplementary tables & rules (composition types B, two-character interaction C, posture-emotion mapping D, weight semantics, narrative-intensity mapping, parameter interaction rules, strength ranking)
• references/ReferenceFirstWorkflow.md — P0 reference-first workflow (v2.5): decision rule, reference-pack quality gates, per-model methods, two-layer defense
• references/ProductionPipeline.md — P0 production pipeline (v2.5): 12-image pre-flight, drift labels, retry rules, versioned reference pack
• examples/nanwang_spring_garden_prompt.md — finished prompts in four syntax flows for the example character Nanwang

Source

Version-controlled at github.com/nohn3043-arch/ai-draw-cue-word-project (also SourceForge p/ai-draw-cue-word-project and Gitee nohn-ecosystem/aidraw). Apache-2.0. Copyright (c) 2026 NOHN AI TECHNOLOGY PTE LTD.