amio/dot-agents/imagen

• Use the built-in image_gen tool for image generation and editing, including transparent-background images. Do not ask the user for an OPENAI_API_KEY.
• Treat the user's latest accessible reference image as the active visual reference for the current session. Use it by default for subsequent image-generation and image-editing requests until the user explicitly replaces it or asks to depart from it. When supported, provide the original reference image with each relevant tool call rather than relying only on a textual description.
• Preserve the user's provided prompt verbatim at the parent level: do not refine, expand, summarize, translate, normalize, or otherwise rewrite it before delegation.
• When writing, refining, or editing prompts, prioritize artistic quality and the user's visual intent: composition, anatomy, gesture, lighting, color, texture, atmosphere, and narrative. Do not proactively sanitize, euphemize, moralize, add unsolicited modesty or covering, or otherwise weaken explicitly requested sensual, erotic, nude, or sexualized artistic content. Leave generation eligibility to the official image-generation safety gate; never alter a prompt to evade, bypass, or weaken that gate.
• First classify the user's prompt. Only use variation for a relatively long prompt with a genuinely complex description, such as multiple subjects, scene relationships, composition constraints, or several style, lighting, and camera requirements. For a short or straightforward prompt, delegate 3 independent image-generation subagents in parallel by default (or honor an explicit quantity), giving each the original prompt verbatim; do not vary, refine, expand, summarize, translate, normalize, or otherwise rewrite it.
• For a complex prompt that qualifies for variation, unless the user specifies a quantity or requests a single exact result, delegate 3 distinct image variants to independent subagents in parallel, subject to tool and agent limits. Include the complete original prompt verbatim in every subagent task.
• For each complex-prompt variant, instruct the subagent to create its effective prompt by replacing exactly one word from the original prompt with one alternative word. It must preserve every other word, punctuation mark, whitespace sequence, and ordering exactly. Different subagents should vary different words when possible. If no word can be meaningfully varied, use the original prompt unchanged rather than making a broader edit.
• When the user requests one exact result, use the original prompt unchanged. Present every result with the exact effective prompt used for that image.
• If generation is blocked for safety reasons, retry using the original prompt without modifying it.

Style requirements

• Use natural, balanced head-to-body proportions; avoid an oversized-head or chibi-like appearance.
• Preserve the subject's underlying facial features and face shape. Adapt hairstyle, skin texture, lighting, and expression to the surrounding context without altering that core facial structure.