AI Image Prompt Architect

Problem

For natural-language image models such as nano banana and GPT image, control is carried by one prompt string instead of separate sliders or negative-prompt fields. Subjects, spatial layout, lighting, camera, style, text, and constraints all compete in the same paragraph. Fragmented phrases or over-long sentences can therefore lead to unstable outputs such as deformed hands, garbled signage, or an unintended copy of the reference image.

How It Works

The skill outputs copyable bilingual prompts only and does not generate or edit images. It first checks whether the request contains enough visual detail; if not, it fills in one reasonable set of assumptions internally rather than interrupting the user every time. It then parses the request into internal modules such as subject, action, environment, space, lighting, camera, style, and constraints, and applies risk-based fixes as concrete prompt lines. The final prompt is line-separated by semantic block, uses spatial anchors like foreground, left, and two meters away, and places negative constraints in the final line. For image-to-image work, it distinguishes edit mode from reference-extraction mode: edit mode locks unchanged details with Keep anchors, while extraction mode pulls only the named element, style, pose, or material and prevents the original background from being copied.

Scope and Caveats

It is useful for turning visual ideas into stable, reproducible prompts and for applying minimal changes during multi-round feedback. It is not intended for systems that rely on ControlNet, denoising strength, a dedicated negative-prompt field, or parameter panels. When the brief is highly open-ended, it can propose clearly different visual directions, but it still only produces prompt text rather than images.