Describe what you want in plain words. PRISM refracts the request into its component pieces, recombines them into an engineered prompt for the model you choose — then refine it in plain language until it's right.
One or two sentences is enough. Rough is fine — the decomposition step will surface anything missing.
Everything below feeds the final prompt. Edit any value, remove what doesn't apply, add what's missing. Dashed cards were inferred, not stated.
Answer what you can. Anything left blank uses the suggested default and is marked as an assumption in the prompt.
The prompt is written in that model family's native conventions — XML tags for Claude, markdown role framing for GPT, descriptor strings for image models.
Tell PRISM what to change and it rewrites the prompt. Each pass keeps everything above it in mind.