image-prompt-direction

name	image-prompt-direction
description	Crafts structured prompts for visual AI image generators (Midjourney, DALL-E, Stable Diffusion) using systematic prompt architecture: subject, style, composition, lighting, mood, and technical parameters. Includes negative prompting and model-specific parameter tuning. Trigger: "write a Midjourney prompt", "image prompt for", "DALL-E prompt", "Stable Diffusion prompt", "art direction for AI", "negative prompt", "style reference", "prompt engineering for images", "consistent character". Do NOT use for text-based prompt engineering (use prompt-crafting), for editing existing images (use image-editor), or for generating images directly (use imagegen).
license	Apache-2.0
compatibility	{"clients":["openai-codex","gemini-cli","opencode","github-copilot"]}
metadata	{"owner":"codex","domain":"image-prompt-direction","maturity":"draft","risk":"low","tags":["image-generation","prompt-engineering","art-direction","midjourney","dall-e","stable-diffusion"]}

Purpose

Provides a systematic methodology for crafting effective prompts for visual AI image generators. Covers prompt architecture, style vocabulary, composition controls, lighting direction, negative prompting, model-specific parameter tuning, and consistency techniques so that generated images reliably match the creator's intent.

When to use this skill

• User asks to write, improve, or debug a prompt for Midjourney, DALL-E, Stable Diffusion, or similar
• Task requires art direction: defining visual style, mood, composition, or lighting for AI generation
• User needs negative prompts to eliminate artifacts or unwanted elements
• Consistency work: maintaining a character, style, or scene across multiple generations
• User wants to understand model-specific parameters (--ar, --v, --s, CFG scale, sampling steps)
• Batch prompt creation for a series of related images (e.g., product shots, character sheets, storyboards)
• User says "this image isn't coming out right" and needs prompt diagnosis

Do not use this skill when

• The task is text-based prompt engineering for LLMs — use prompt-crafting
• The user wants to edit or manipulate an existing image — use image-editor
• The user wants to directly generate an image (not craft the prompt) — use imagegen
• The task is brand identity or logo design process — use brand-guidelines for strategy, then this skill for generation prompts
• The user needs to build a UI that displays images — use frontend-design or react-typescript

Operating procedure

Step 1 — Creative brief intake

Gather these inputs before writing any prompt:

1. Subject — what is the image of? (person, object, scene, concept)
2. Purpose — where will this image be used? (social media, website hero, print, concept art, personal)
3. Target model — which generator? (Midjourney v6, DALL-E 3, SDXL, Flux, other)
4. Aspect ratio — landscape (16:9), portrait (9:16), square (1:1), or custom
5. Style direction — any reference images, artists, art movements, or aesthetic keywords
6. Constraints — brand guidelines, content policies, elements that must NOT appear
7. Quantity — single image or series; if series, what varies between images

Step 2 — Prompt architecture

Build the prompt using this layered structure, from most to least important:

[Subject & Action] + [Environment/Setting] + [Style & Medium] + [Composition & Framing] + [Lighting] + [Mood & Atmosphere] + [Color Palette] + [Technical Parameters]

Subject & action layer

• Lead with the primary subject: "a weathered lighthouse keeper", not "an image of a man"
• Include action or pose if relevant: "standing at the railing, looking out to sea"
• Specify key details: age, expression, clothing, distinguishing features
• Use concrete nouns over abstract descriptions

Environment / setting layer

• Ground the subject in a specific place: "on a rocky coastal cliff during a storm"
• Include time of day and season if relevant to lighting/mood
• Add depth cues: foreground, midground, background elements

Style & medium vocabulary

Category	Examples
Photorealistic	editorial photography, 35mm film, DSLR, Hasselblad, Canon EOS R5
Illustration	digital illustration, concept art, matte painting, book illustration
Traditional media	watercolor, oil painting, gouache, charcoal sketch, ink wash
Graphic styles	vector art, flat design, isometric, pixel art, low-poly, voxel art
Art movements	Art Nouveau, Art Deco, Impressionism, Surrealism, Ukiyo-e, Bauhaus
Contemporary	synthwave, vaporwave, cyberpunk, solarpunk, cottagecore, dark academia

Composition & framing controls

Technique	Prompt language
Rule of thirds	"subject positioned at the left third intersection"
Golden ratio	"golden spiral composition"
Leading lines	"converging railway tracks drawing the eye to the vanishing point"
Depth of field	"shallow depth of field, bokeh background", "f/1.4 aperture"
Camera angle	"bird's-eye view", "worm's-eye view", "eye-level", "Dutch angle"
Shot type	"extreme close-up", "medium shot", "wide establishing shot", "over-the-shoulder"
Symmetry	"perfectly symmetrical composition", "bilateral symmetry"

Lighting vocabulary

Lighting type	Visual effect	Prompt language
Rembrandt	Triangle of light on shadow-side cheek	"Rembrandt lighting, dramatic triangle shadow"
Rim / back lighting	Glowing edge outline	"rim lighting, backlit silhouette glow"
Golden hour	Warm, soft, directional	"golden hour sunlight, warm tones, long shadows"
Studio	Clean, controlled	"studio lighting, softbox, white background"
Chiaroscuro	High contrast, dramatic	"chiaroscuro lighting, deep shadows, Caravaggio-inspired"
Overcast / diffused	Soft, even, no harsh shadows	"overcast sky, diffused natural light, soft shadows"
Neon / colored	Stylized, cyberpunk	"neon-lit, pink and cyan light spill, wet reflections"
Volumetric	God rays, fog beams	"volumetric lighting, god rays through fog"

Mood & atmosphere

• Use emotional adjectives: "melancholic", "triumphant", "eerie", "serene", "chaotic"
• Pair mood with environmental cues: "desolate wasteland" vs. "lush paradise"
• Color temperature reinforces mood: warm = comfort/energy, cool = calm/isolation

Step 3 — Negative prompting

Define exclusions to steer the model away from common failure modes:

Universal negative prompt baseline:

blurry, low quality, low resolution, watermark, text, signature, cropped, out of frame, worst quality, jpeg artifacts

Anatomy-specific (for human subjects):

deformed hands, extra fingers, extra limbs, fused fingers, mutated hands, bad anatomy, malformed limbs, extra arms, extra legs, cloned face

Style-specific negatives:

• Photorealistic: illustration, cartoon, anime, painting, drawing, CGI
• Illustration: photo, photograph, realistic, photographic
• Clean design: cluttered, busy background, noisy, grain

Guidance: Only include negatives relevant to your target style. Over-stuffing the negative prompt can constrain the model too aggressively and produce flat results.

Step 4 — Model-specific parameters

Midjourney (v5/v6)

Parameter	Purpose	Range / values
--ar	Aspect ratio	1:1, 16:9, 9:16, 3:2, 2:3, 21:9
--v	Model version	5.2, 6, 6.1
--s (stylize)	How strongly Midjourney applies its aesthetic	0–1000; low = literal, high = artistic
--c (chaos)	Variation between results	0–100; low = similar, high = diverse
--q (quality)	Rendering time/detail	0.25, 0.5, 1 (default), 2
--no	Negative prompt	--no text, watermark, blurry
--seed	Reproducibility	Any integer; same seed + same prompt ≈ same image
--sref	Style reference image URL	--sref <url> to match a visual style
--cref	Character reference image URL	--cref <url> to maintain character consistency

DALL-E 3 (via ChatGPT / API)

• No direct parameter flags; control via detailed natural language in the system prompt
• Specify aspect ratio in the request: "wide landscape format", "tall portrait format", "square"
• Use "I NEED the image to exactly match this description" to reduce DALL-E's prompt rewriting
• Negative concepts go in the prompt body: "without any text or watermarks"
• Seed control: not user-facing; use very specific descriptions for consistency

Stable Diffusion (SDXL / SD 1.5)

Parameter	Purpose	Recommended range
CFG Scale	Prompt adherence vs. creativity	5–12; 7 is a good default
Sampling steps	Detail vs. speed	20–50; 30 is a good default
Sampler	Denoising algorithm	Euler a (fast), DPM++ 2M Karras (quality), DDIM (consistent)
Seed	Reproducibility	Any integer
Resolution	Output dimensions	SDXL: 1024×1024, 1216×832, 832×1216; SD 1.5: 512×512, 768×512
LoRA / embeddings	Style or character fine-tuning	Specify model name and weight (e.g., <lora:stylename:0.7>)
Negative prompt	Separate field for exclusions	Use the negative prompts from Step 3

Step 5 — Consistency techniques

For maintaining visual coherence across a series:

1. Seed pinning — use the same seed number across variations; change only the element you want to vary.
2. Reference images — provide reference URLs (Midjourney --sref, --cref) or img2img in SD.
3. Style transfer — lock the style by referencing a specific artist, movement, or prior generation.
4. Character sheets — generate a multi-angle character reference sheet first, then use it as reference for scene images.
5. Prompt templating — create a base prompt template with {variable} slots; only swap the variable per image.
6. Consistent terminology — use the exact same descriptive phrases for recurring elements across all prompts in a series.

Step 6 — Iteration and refinement

1. Generate initial results with the crafted prompt.
2. Diagnose issues using this checklist:
   • Wrong style → adjust style/medium keywords; add style-negatives
   • Wrong composition → make framing/angle instructions more explicit
   • Wrong lighting → specify lighting type by name; add "no flash photography" if too flat
   • Artifacts → strengthen negative prompt; adjust CFG or steps
   • Too literal / too abstract → tune stylize (MJ) or CFG (SD)
3. Document what changed between iterations and why (Iteration Notes).
4. After 3+ failed iterations, reconsider the fundamental approach rather than tweaking parameters.

Decision rules

• Prompt length > 75 tokens (SD) or overly complex → simplify; focus on the most important 5–7 descriptors. Models lose coherence with prompt-stuffing.
• Conflicting styles in one prompt (e.g., "photorealistic watercolor") → pick one style; use blending only with explicit intent and model support.
• No aspect ratio specified → always specify; default square rarely matches the use case.
• No negative prompt → always include at minimum the universal baseline negatives.
• Anatomy in the image → always include anatomy-specific negatives for human/animal subjects.
• Requesting text in images → warn that most models render text poorly; suggest adding text in post-production. DALL-E 3 handles short text better than others.
• Brand or commercial use → flag content policy and licensing considerations for the target model.
• Reproducibility required → mandate seed pinning and document all parameters.
• Vague style words ("cool", "nice", "beautiful") → replace with specific aesthetic vocabulary from the style table.

Output structure

Deliver these sections for each prompt request:

1. Prompt Brief

Summary of creative intent: subject, purpose, target model, aspect ratio, style direction.

2. Full Prompt

The complete, ready-to-paste prompt text with all layers applied.

[Full prompt text here]

3. Negative Prompt

Separate negative prompt text (for SD; for Midjourney, integrated via --no).

[Negative prompt text here]

4. Parameter Settings

Model-specific parameters in a copy-pasteable format:

--ar 16:9 --v 6 --s 250 --seed 42

or

CFG: 7 | Steps: 30 | Sampler: DPM++ 2M Karras | Seed: 42 | Resolution: 1024x1024

5. Iteration Notes

Table tracking changes across prompt versions:

Version	Change made	Reason	Result
v1	Initial prompt	—	—
v2	Added "Rembrandt lighting"	First version too flat	Improved shadow depth

6. Consistency Guide (for series)

Template, fixed elements, variable elements, seed, and reference image strategy.

Anti-patterns

• Prompt-dumping — cramming 20+ disconnected style keywords into one prompt. Models lose coherence; pick 3–5 complementary descriptors maximum.
• Ignoring aspect ratio — generating square images for a 16:9 banner creates cropping headaches or distortion. Always match output to intended use.
• No negative prompts — omitting negatives in Stable Diffusion especially leads to watermarks, text artifacts, and anatomical errors that are trivially preventable.
• Vague style words — "beautiful", "stunning", "epic" carry almost zero signal for the model. Replace with concrete aesthetics: "golden hour rim lighting" beats "beautiful lighting".
• Conflicting style instructions — "photorealistic anime oil painting" fights itself. Be intentional about style blending or pick one.
• Ignoring model strengths — using Stable Diffusion for text-heavy graphics (DALL-E 3 is better) or DALL-E for precise parameter control (SD is better). Match task to model.
• No seed documentation — generating a great image and not recording the seed or parameters means it cannot be reproduced or iterated upon.
• Over-relying on artist names — ethical and legal grey area; also, model knowledge of specific artists varies. Describe the visual qualities you want instead.
• Skipping iteration notes — repeating the same failed adjustments because previous attempts weren't documented.

Related skills

• imagegen — directly generating images (this skill crafts the prompt; imagegen executes it)
• prompt-crafting — text-based prompt engineering for LLMs (sister skill for non-visual prompting)
• brand-guidelines — ensuring generated images align with brand identity standards
• image-editor — post-production editing of generated images
• worldbuilding-research — generating consistent visual world-building assets
• canvas-design — composing generated images into layouts and designs

Failure handling

• If the user provides no subject or purpose, ask: "What is the image of, and where will it be used?"
• If the target model is not specified, default to Midjourney v6 parameters but note the assumption and offer alternatives.
• If the user's description contains conflicting style directions, surface the conflict and ask them to prioritize.
• If the prompt exceeds the model's effective token window, summarize and offer both a full and condensed version.
• If the user expects photorealistic text rendering, warn about model limitations and suggest compositing text in post-production.
• If consistency across a series is failing, recommend switching to a seed-pinning + reference-image strategy before further iteration.