Demo tool · built by Derek Dorsett

Prompt Console

A composer I built to turn an image brief into a paste-ready prompt tailored to each engine's architecture. It never calls a model — it composes the prompt; you paste it into Gemini or ChatGPT and run it yourself. The point: the same brief produces a different, architecture-correct prompt per engine, and references — not seeds — carry visual consistency.

A clean public demo. The reference library below uses samples of my own design work. The production version I built for a confidential client used a curated program-specific library, a swappable program-pack model, and a brand/voice layer — available to discuss on request.

Compose

Engine

Reference images (the consistency lever — these engines have no usable seed)

Click a thumbnail to add it as a reference, then set its role. Hover (or focus) to preview.

Paste-ready prompt

GEMINI · IMAGE

…

How it composes (the engineering)

Dual rendering by encoder. Gemini/GPT use LLM text encoders, so they get narrative fragments; CLIP-encoder engines (SDXL, Midjourney) get comma-separated tokens. This tool emits the narrative form.
Per-engine dialect. Gemini image prompts are sectioned (§ SUBJECT / § STYLE / § REFERENCES / § CONSTRAINTS) with a hard no-text prohibitions block; GPT image prompts are one declarative scene sentence + compact rule lines.
Negatives are architecture-specific. These engines have no negative-prompt field, so constraints are phrased positively ("clean background," not "no clutter").
References, not seeds. Neither Gemini-native nor GPT image exposes a reliable seed, so reference images carry consistency — each is given an explicit role (identity-lock, style-match, geometry-lock…) and listed in an attach manifest.