@rules/capability-and-output.md @rules/prompt-compilation.md @references/visual-direction.md @references/prompt-crafting.md @references/runtime-capability-and-drift.md

Image Maker

Job, boundary, and fallback

Turn a one-off create or edit request into a truthful delivered artifact. This skill also owns a one-off prompt-only request. It does not analyze images, build reusable prompt packs, write generic design documents, or implement editable UI or code.

Normal image requests carry fallback_policy=explicitly_allowed: unavailable/unknown relevant completion, or two objective invocation failures using the same immutable brief, saves a prompt-only artifact without another question. A per-request refusal sets explicitly_rejected and blocks fallback. Preserve exact user choices and rendered text; never represent a prompt as an image.

In ChatGPT, a create request is not a prompt-only request: compile the final prompt and invoke the exposed native image-generation tool in the same run. When the runtime exposes a model selector, specify exactly gpt-image-2.5; when it does not (ChatGPT web plugin included), use the runtime default model and do not claim 2.5 — report wording is "runtime default model". Never ask the user to select a model. Model names never enter the prompt string or the brief. Do not choose fallback merely because repository file writing, descriptor-relative writing, or separate retrieval is unavailable; the native returned image is the primary delivery artifact. Fall back only after the native image capability is absent or two objective invocation failures occur.

Contract