| name | gpt-image-prompt-engineer |
| description | Build model-free multilingual GPT Image prompt bundles for generation and editing, with user-locked parameters, deliverable-specific auto choices, reference-image fidelity rules, final safety/quality review, and optional review-module reports compatible with subagent orchestration. |
GPT Image Prompt Engineer
Version 5.5.0. Use the Python entrypoint python/skill.py; the deterministic core lives in src/gpt_image_prompt_engineer.
Runtime metadata and defaults are recorded in skill.lock.json. Use scripts/run_prompt_flow.py for package-level execution and scripts/verify.sh for the bundled verification workflow when a shell environment supports it.
Key guarantees:
- The skill is model-free; callers add the image model externally.
response_mode=text_prompt returns only final-reviewed prompt text.
response_mode=json_bundle returns prompts, locked_user_params, auto decision trace, safety review, final review, review-module reports, conflict resolution, reference preflight, and render parameters.
- Deliverable-specific profiles cover posters, social/story posts, presentation slides, product and packaging mockups, storyboards, and other supported image formats.
- Explicit user-supplied values are recorded as locked parameters and must not be overridden by
auto decisions.
final_review repairs unsafe/high-risk wording before output, reinforces reference fidelity and storyboard continuity, and reports missing inputs plus clarifying questions.
- Real product/place workflows require factual reference grounding through
verified_reference_facts and reference_sources; those references may supplement but never replace user-provided details.
- The native runtime is deterministic and model-free; review-module reports do not call an LLM unless a host application deliberately replaces them with external agents-as-tools.