Turn a photo's own gesture, silhouette and materials into playful portrait art.
中文说明 · Skill instructions · Prompt template
A person gives two thumbs up: their sleeves become boosters. Pressed palms and a pleated dress become the structure of a leaf. A tilted smile and black shirt become a curling vinyl record.
The idea starts with what is visible in the photo. The subject becomes part of the composition.
- A focused Agent Skill for photo-led surreal portraits.
- An open choice of mechanisms, with boundaries learned through iterative feedback.
- A complete prompt template and behavior test requests.
- No model weights, API keys, automatic uploads or bundled portrait photos.
Use an agent that can inspect images and load skill instructions. Creating images also requires an available image generation/editing tool. Without that tool, the skill can provide a prompt, not a finished picture. Image-provider access and any charges are separate from this repository.
The development workflow was exercised in Codex Desktop with its built-in image tool. Other agents and image providers have not been independently tested. This public package removes private workstation dependencies; that packaging change is not proof of equivalent image quality on every host.
From your project's root, put the repository in a skill directory:
git clone https://github.com/jay-yangPY/portrait-visual-play.git .agents/skills/portrait-visual-playIf that destination already exists, review it instead of overwriting it. Reload the client if the skill is not yet listed. This repository contains no installer or startup script.
For another compatible agent, follow its skill-directory convention. You can also read SKILL.md and use the prompt contract manually.
Attach a photo, then ask:
Use $portrait-visual-play on this photo. Find the most interesting gesture or shape and turn it into one playful portrait poster. Keep the person recognizable and the mood friendly.
For a prompt without image generation:
Use $portrait-visual-play to inspect this photo and give me only the complete image prompt.
- Observe the actual photo.
- Consider three different mechanisms internally and choose one.
- Define the structural change and the identity cues to preserve.
- Test whether the concept still works without text or decoration.
- Generate with the host's real tool, then inspect the final image.
The workflow distinguishes a clear mechanism from a pretty effect. A new background or a flower frame may be a valid edit, but it is outside this skill's purpose.
Run the lightweight package check with Python 3:
python scripts/validate.pyIt checks required files, local links and accidental private-path/credential patterns. It does not execute an image model or certify security, likeness or aesthetics.
Development included real image generations and independent visual review for selected outputs. These are limited case studies, not a benchmark. Image models may alter faces, teeth, skin, hands or text, especially from small source photos. Inspect each result before use.
Original instructions and repository code are released under MIT. This license does not grant rights to source photos, depicted people, third-party reference artwork or external model outputs. Development photos and private review logs are not distributed.
The method was developed through iterative portrait-editing experiments and study of photographic surrealism, material continuity and negative-space composition. No reference artwork, artist signature or third-party skill source is bundled.
The file layout follows the Agent Skills specification. The skill itself is a set of instructions, not a trained image model.
Open an issue with the request, intended relationship and observed failure. Share photos only when you can share them; avoid private identities or credentials. Describe the actual result separately from the prompt. A small reproducible failure is more useful than an unsupported quality score.