All-Model Image Prompt Compiler

Problem: prompts drift across models

The hard part of image generation is not “not knowing what to describe”; it is making the same idea work across models with different syntax, parameters, compliance rules, and strengths. Midjourney flags like --ar and --no do not map cleanly to FLUX, DALL·E 3, or domestic models. E-commerce images may need domestic platform banned-word handling; posters with text need layout control; and face, product, or content references can conflict if the camera facts are not consistent.

How it works

This skill turns user intent into copy-ready, model-native prompts. It routes by image type: e-commerce product shots to Jimeng/Kling/FLUX/SeaArt, text-heavy posters to Ideogram/Recraft/Canva, and anime/illustration to PixAI/NovelAI. The workflow includes:
- 8-layer normalization: subject, scene, pose, style, composition, lighting, tone, and technical texture;
- reference classification: face, product, contents, style, and scene, with constraints against incompatible visual facts;
- compliance scan: banned words, diversion signals, and on-image text risks, with replacement suggestions;
- anti-AI-flavor rewriting: replacing vague terms like “confident smile” and “cinematic” with concrete facial states, lens choices, lighting, and film simulation;
- iteration loop: fixing face drift, composition, color, AI smell, or text artifacts by adjusting weights, negative prompts, references, or model routing.

Boundaries

It targets still-image prompts only, not direct image generation, video, camera motion, frame rate, or motion blur. Compliance is a prompt-level first pass, not a guarantee against model-side filters or human review. High fidelity depends on reference images, ControlNet/IP-Adapter, locked seeds, and multiple iterations; text-only prompts cannot promise 90% similarity.