tekkiiiii/the-agency/image-prompt-engineer
Claude Code and agent skill by tekkiiiii/the-agency
- community
Checking that the SKILL.md file is still in the repository.
Use when writing, improving or debugging an image-generation prompt: \"write an image prompt\", \"Midjourney prompt\", \"DALL-E prompt\", \"Stable Diffusion prompt\", \"negative prompt\". Any generator (also Flux, Higgsfield, Leonardo, Ideogram): structure, parameters, style vocabulary, composition, iteration.
Install
npx vskill@latest install tekkiiiii/the-agency/image-prompt-engineerSKILL.md
The same SKILL.md was published earlier as tekkiiiii/the-agency/image-prompt-engineer.
Image Prompt Methodology Skill
Teaches HOW to write a great image-generation prompt, for any generator and any subject. This is a methodology skill, not a lookup library — it gives the structure, vocabulary, and per-generator parameter syntax needed to construct a prompt from scratch or fix one that isn't producing the right result.
Not in scope here: a curated database of ready-made prompts (see gpt-image-prompts skill for 476+ examples you can browse/search/adapt), and not UI-screenshot-specific composition rules (see imagegen-frontend-web / imagegen-frontend-mobile for those). This skill is the general-purpose methodology underneath all three.
When to use this skill
- Asked to write an image prompt and no existing curated prompt fits the need
- Asked to improve, debug, or extend an existing prompt that isn't producing the right output
- Building a prompt for a generator not covered by another skill (Stable Diffusion, Flux,
Higgsfield, Leonardo, Ideogram, etc.)
- Any agent (image-prompt role, AI-video role, content writers, design agents)
needs to produce or QA an image-generation prompt
Core Methodology — The Five-Layer Prompt Structure
Every effective image prompt is built in layers, ordered most-important-first. Most generators weight earlier tokens more heavily, so put what matters most up front.
- Subject — who/what is in frame, with concrete (not vague) descriptors
- Bad: "a woman" → Good: "a woman in her 30s with curly red hair, wearing a tailored navy blazer"
- Action / Composition — pose, framing, what's happening, where elements sit
- Subject position (centered, rule-of-thirds, foreground/background), interaction with environment
- Environment / Setting — location, time of day, weather, background treatment
- Lighting — source, direction, quality, color temperature (this is the single highest-leverage
lever for photorealism and mood — never leave it implicit)
- Style / Technical — art style or photography genre, camera/lens specs if photoreal,
color grade, quality boosters, generator-specific parameters (always last)
Template:
Concrete beats vague every time. "Shallow depth of field, f/1.8 bokeh" beats "blurry background" — name the mechanism, not the visual effect, because generators were trained on photography/art captions that use mechanism-language.
Per-Generator Parameter Reference
Midjourney
Parameters go at the end of the prompt, each prefixed --.
| Param | Purpose | Common values | |
|---|---|---|---|
--ar | Aspect ratio | 16:9, 1:1, 9:16, 3:2 | |
--v | Model version | 6.1, 6, niji 6 (anime) | |
--style | Style preset | raw (less opinionated/more literal), cute, scenic (niji) | |
--stylize (--s) | How strongly MJ's house aesthetic is applied | 0–1000, default 100 | |
--chaos (--c) | Variation between the 4 grid outputs | 0–100 | |
--weird (--w) | Unusual/experimental aesthetics | 0–3000 | |
--no | Negative prompt (exclude elements) | --no text, watermark | |
--seed | Reproducibility | any integer | |
:: weighting | Emphasize/de-emphasize a term | red hair::2 blue eyes::1 |
DALL-E / GPT-Image-2
No bracket-parameter syntax — everything is natural language. Precision comes from explicit, plain-English description rather than tokens.
- Specify aspect/composition in words: "wide cinematic shot," "square product photo on white background"
- Specify lighting and perspective explicitly — the model follows literal descriptions well
- Style is steered by naming a medium/genre directly: "isometric vector illustration,"
"35mm analog film photo," "studio Ghibli-style animation still"
- No native negative-prompt field — exclusions must be phrased positively
("clean background" instead of "no clutter") or stated as a constraint in the sentence ("no text or logos anywhere in the image")
- See
gpt-image-promptsskill for 476+ tested examples across 5 categories in this exact style
Stable Diffusion (and forks: SDXL, SD3, Automatic1111/ComfyUI workflows)
Token-weighted, supports a true negative-prompt field.
| Param | Purpose | Notes | |
|---|---|---|---|
| CFG Scale | How strictly the output follows the prompt | 7–12 typical; higher = more literal, less creative | |
| Sampler | Denoising algorithm | DPM++ 2M Karras, Euler a, DDIM — affects detail/speed tradeoff | |
| Steps | Denoising iterations | 20–40 typical; diminishing returns past ~40 | |
| Negative prompt | Separate field listing what to avoid | blurry, deformed hands, extra fingers, watermark, low quality | |
(token:1.3) weighting | Emphasize a term | Higher number = stronger pull | |
| LoRA / embedding tags | Style or subject fine-tune reference | <lora:name:0.8> |
Flux (FLUX.1, Flux Pro/Dev)
Natural-language-first like DALL-E, but rewards longer, highly detailed descriptive paragraphs over keyword-stacking. Strong photorealism — lean on real photography vocabulary (lens, film stock, lighting setup) rather than generic quality tags.
Higgsfield
Built for motion/video-adjacent generation — prompts need a motion/camera-move layer in addition to the static-image layers above.
| Concept | Purpose | Examples | |
|---|---|---|---|
| Motion style | How the subject/scene moves | subtle parallax, slow zoom in, dynamic action | |
| Camera move | Virtual camera behavior | dolly in, orbit left, static lockoff, handheld shake | |
| Reference image/motion | Anchor to an uploaded asset | use motion_control for recast/puppeteer/motion-transfer |
- Use
models_explore(action:'recommend')to pick the right Higgsfield model for the goal before generating.
Leonardo AI
Preset-driven — prompt is plain language, but model/preset choice carries a lot of the style weight (PhotoReal, Illustration, Anime presets each interpret the same prompt differently). Supports negative prompt field and an "Elements" (LoRA-like) system for style consistency.
Ideogram
Strongest of the mainstream generators at rendering legible text-in-image — if the deliverable needs accurate on-image typography (posters, signage, logos with text), prefer Ideogram and state the exact text in quotes within the prompt.
Style Modifier Vocabulary
Lighting terms: golden hour, blue hour, overcast diffused light, hard direct sunlight, softbox key light, rim/edge light, Rembrandt lighting, butterfly lighting, split lighting, volumetric light/god rays, neon practical lighting, chiaroscuro, backlit silhouette, high-key (bright, low contrast), low-key (dark, high contrast)
Camera/lens terms: 35mm/50mm/85mm focal length, wide-angle distortion, telephoto compression, shallow depth of field (f/1.4–f/2.8), deep focus (f/8–f/16), bokeh, macro, low angle / high angle / bird's-eye / worm's-eye, Dutch tilt, tilt-shift, anamorphic lens flare, long exposure, motion blur
Art style terms: isometric, flat vector illustration, watercolor, oil painting, cel-shaded anime, photorealistic, hyperrealistic, claymation, low-poly 3D, brutalist, art nouveau, cyberpunk, vaporwave, minimalist, maximalist, editorial photography, documentary photography, fine art print
Quality boosters (use sparingly — generator-dependent, can be redundant on newer models): "highly detailed," "8k," "sharp focus," "professional photography," "award-winning" — on modern models (Midjourney v6+, GPT-Image-2, Flux) these add less than they used to; concrete descriptive language outperforms generic quality tags. Prefer specificity over boosters.
Negative Prompt Patterns
Only Midjourney (--no), Stable Diffusion (negative field), and Leonardo have a true negative-prompt mechanism. For DALL-E/GPT-Image-2/Flux, rephrase exclusions as positive constraints inside the main prompt.
Common negative-prompt targets:
- Anatomy failures:
deformed hands, extra fingers, asymmetric eyes, malformed limbs - Unwanted artifacts:
watermark, text, logo, signature, blurry, low quality, jpeg artifacts - Composition failures:
cropped, out of frame, duplicate, cluttered background - Style leakage: name the unwanted style explicitly if the model keeps drifting toward it
(e.g., --no anime if a photoreal prompt keeps returning illustrated results)
Multi-Subject / Scene Composition Rules
- Establish a clear hierarchy — name the primary subject first, secondary subjects after,
with explicit spatial relationships ("in the foreground... behind her, slightly blurred...")
- Cap subject count realistically — most generators degrade past 2-3 named subjects with
distinct described attributes; complex group scenes need either acceptance of some randomness or multiple generation + compositing passes
- Use weighting/emphasis to keep the hierarchy from collapsing (Midjourney
::,
SD (token:1.3)) — without it, generators often give equal visual weight to all mentioned elements regardless of prompt order
- Describe interactions, not just co-presence — "two people shaking hands" generates
more coherently than "two people, a handshake" as a bolted-on fragment
- One scene, one lighting setup — conflicting lighting descriptions for different
subjects in the same frame is a common cause of incoherent results
Iterative Refinement Pattern
- Generate a baseline with the core 5-layer prompt — don't over-specify on attempt 1
- Diagnose the gap — is it composition, style, lighting, or a specific element wrong?
Change ONE layer at a time; changing everything at once makes it impossible to learn what worked
- Reuse the seed (where supported — Midjourney
--seed, SD seed field) when refining
composition/style without wanting the whole image to change
- Escalate specificity, not length — replace vague terms with concrete ones rather
than padding the prompt with more adjectives
- When stuck after 3 iterations, suspect a structural conflict (e.g., asking for both
"soft diffused light" and "dramatic hard shadows") rather than a wording problem — re-read the prompt for contradictions before adding more detail
- Document what worked — once a prompt pattern produces a reliable result for a
recurring need, that's a candidate to add to the gpt-image-prompts library rather than rediscovering it each time
Workflow
- Identify the target generator (if unspecified, ask or default to GPT-Image-2/natural language)
- Identify the subject type — is this photography-style (use the image-prompt role file's genre patterns for portrait/product/landscape/fashion), a UI mockup (route to
imagegen-frontend-web/imagegen-frontend-mobile instead), or general/illustrative (use this skill's five-layer structure directly)
- Build the prompt layer by layer: subject → action/composition → environment → lighting → style/technical
- Apply the generator's parameter syntax from the reference table above
- Add negative-prompt exclusions if the generator supports them; otherwise fold exclusions
in as positive constraints
- If multiple subjects, apply the composition rules above
- Generate, diagnose gaps against the 5 layers, refine one layer at a time
- Run critique-imageprompt on the final prompt before delivery — invoke the critique-imageprompt
agent or paste the prompt into the critique checklist below. Address any FAIL or NEEDS WORK findings before finalizing.
Cross-References
This is the start of the file. Read the full SKILL.md on GitHub.
Security
- Security scanPassed, 100 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingskills/image-prompt-engineer/SKILL.md
Popularity
- Repository stars
- 3
- 7-day trend score
- 7
Stars and forks are for the whole GitHub repository, not this skill alone. Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/tekkiiiii/the-agency/image-prompt-engineer)