Image Prompt Director

What this skill does

• **Model + job first:** asks the generator/version + use-case, switches the syntax dialect and aspect ratio
• **Fills every pillar:** subject, action, environment, composition, camera, lighting, color, style, quality — and translates vibes into concrete descriptors
• **Consistency kit for series:** fixed style block + character bible + seed/reference discipline (`--oref`/`--sref`/Kontext/Nano Banana) — the #1 pain
• **Negative prompts done right:** real only in the SD family; elsewhere phrase exclusions positively, never over-stuff
• **Text & logos routed right:** legible text to Ideogram/GPT Image; exact logos/hex composited in a design tool
• **Aspect ratio + safe zones** per platform; **troubleshooting** (hands, text, bleeding, "AI look") with the one fix
• **Rights check:** license per model, copyright, trademark/likeness, labeling → hand off to KI-Transparenz
• **Two Node scripts:** prompt-forge (builds a model-correct prompt) and prompt-lint (QA against vagueness/dialect-mismatch/living artists)

Description

Image Prompt Director is a skill that turns an idea and your brand into a precise, **model-aware** image prompt — one you paste into your generator (Midjourney, GPT Image/DALL·E, Gemini/Nano Banana, FLUX, Stable Diffusion, Ideogram, Firefly). It's the art director, not the generator: it writes the best possible prompt (plus the negative prompt, parameters and a reusable style kit) and explains every choice. Founding rule: **model-aware, not flowery.** Prompt syntax doesn't port — `(word:1.3)` is Stable-Diffusion-only, Midjourney uses `--no`/`--sref`, GPT/Gemini/FLUX take plain natural language; generic bots that bolt adjectives onto an idea are why AI images all look the same. So it asks the model **and version** first, then emits the right dialect. It fills **every pillar** (subject → action → environment → composition → camera → lighting → color → style → quality — a missing pillar means generic slop), translates vibes into physical descriptors ("premium" → matte metal, soft studio light, negative space), directs the camera (shot/angle/focal length/depth of field) and **names the light** (golden hour, Rembrandt). For **series** it solves the #1 pain — consistency — with a fixed style block plus a **character bible**, seed discipline, and the right reference technique per model (`--oref`/`--sref`, FLUX Kontext, Gemini/Nano Banana, IP-Adapter). It handles **negative prompts** honestly (real only in the SD family; elsewhere phrase positively), **routes legible text** to the models that can render it (Ideogram/GPT Image/Recraft, not Midjourney), sets the **aspect ratio** to the destination (9:16 story, 4:5 feed, 16:9 thumbnail) and debugs **one variable at a time** (hands, faces, garbled text, prompt bleeding, the "AI look"). And it flags **rights and labeling** honestly: commercial license differs per model, a prompt alone doesn't confer copyright, don't clone living artists, composite exact logos/colors — and published AI images may need a label (hand off to the **KI-Transparenz** skill). All verify-live, because the models change monthly.

Examples

Tested with Claude Code

matte-black insulated water bottle on a seamless charcoal studio surface, single soft key light from upper left with a subtle rim light separating it from the background, faint contact shadow, condensation droplets, minimalist premium product photography, brushed-metal cap, shallow depth of field, 85mm, ultra sharp --ar 4:5 --style raw --stylize 150

What it does not do

• **Doesn't generate** and doesn't upscale/inpaint or check the image — it writes the prompt; you generate in your own tool
• **Guarantees no** exact logos, exact text, exact brand hex or pixel-exact character identity — gives the reliable path (compositing/reference) instead
• **Not legal advice** — flags license/copyright/trademark/labeling; verify the model's live terms; labeling → KI-Transparenz
• **Verify-live** — models change monthly; it dates its claims rather than hardcoding "the latest is vX"

Compatibility & tech

Tested (internal)

3 scenarios

Recommended runtime

Claude Opus or Sonnet; with web access to verify current model versions/parameters and "best model for X" live. The two scripts need Node (no extra packages); without Node the skill produces the prompt and lint by hand from the references.

Modes

Write an image prompt (model-aware) · Consistency kit for series (style/character bible) · Troubleshooting / iteration · Rights & labeling check

Inputs

idea/subject + target model + use-case · brand (colors, mood, style) + one-off or series · an existing prompt (to improve) or a failed result (to fix)

Output format

A Markdown deliverable per mode — the model-correct prompt (plus negative prompt/parameters) in clean copy-paste blocks, optionally the reusable style/character bible, plus a brief why, a rights flag, and one next iteration.

Subcategory

AI image generation, prompt direction & image consistency

License

Proprietär

Security profile

What you get

• image-prompt-director-1.0.0/10 files
  • .claude-plugin/marketplace.json
  • plugins/image-prompt-director/skills/image-prompt-director/10 files

    • SKILL.md
    • manifest.json
    • references/6 files

      • image-prompt-director-anatomy.md
      • image-prompt-director-consistency.md
      • image-prompt-director-models.md
      • image-prompt-director-rights.md
      • image-prompt-director-troubleshooting.md
      • image-prompt-director-usecases.md
    • scripts/2 files

      • prompt-forge.mjs
      • prompt-lint.mjs
  • .agents/skills/image-prompt-director/→ universal — same content (Codex, Cursor, Copilot, Gemini, Windsurf, Cline)
  • LICENSE.txt

Installation

Also works as a chat prompt

Note

Image Prompt Director writes prompts — it **doesn't generate, upscale or inpaint** and doesn't check the finished image; you generate in your own tool. AI image models change monthly: with web access the skill verifies current versions/parameters live and dates its claims. It's a **first orientation, not legal advice** — verify commercial license, copyright, trademark and likeness per model; it doesn't clone living artists and can't guarantee exact logos/text/brand colors (composite instead). For the labeling duty of published AI images (EU AI Act Art. 50), the KI-Transparenz skill is responsible. The bundled scripts need Node (no extra packages).

Changelog

• v1.0.021.08.2026Initial release: turns an idea + brand into a precise, model-aware image prompt. Asks the target model (Midjourney, GPT Image/DALL·E, Gemini/Nano Banana, FLUX, Stable Diffusion, Ideogram, Firefly) and emits the correct syntax (--ar/--sref/--no vs (word:1.3)+negative field vs plain language), fills every pillar (subject, composition, camera, lighting, color, style), builds a reusable style/character bible for consistent series (the #1 pain), routes legible text to the right model, gives negative-prompt/aspect-ratio/iteration guidance, and flags rights/labeling (license, copyright, trademark, EU AI Act → hand off to KI-Transparenz). Two Node scripts (prompt-forge, prompt-lint), 6 references. It doesn't generate — verify-live.

Frequently asked questions