Product Scene Studio
A reference-led product concept workflow with an explicit preserve/change contract. Approve the reference first, create scene variants from that reference, and inspect component counts, shape and visible details before delivery.
A little less setup
Make this part of your toolkit.
Give your agent this prompt. You’ll approve access in your browser before it installs a Skill.
Read https://skilly.uk/install/skills/product-scene-studio.md and follow the instructions to install this Skill for my agent.REAL SAMPLE / RELEASED PACKAGE
RELEASED v1.0.0A closer look at the actual output.
5 selected original PNG examples from pilot and independent new-brief executions. Prompts and rejected attempts are retained in the package; the images are not deterministic templates.
GENERATED EXAMPLES
Explore the actual images.
Browse original generated images, example details and review notes. Choose Open for original images and source links. This gallery does not generate images; use the Skill in Codex with native image access.
Responsive gallery · original image proportions preservedUse the preview’s own scroll and keyboard controls.
Inside this gallery
5 images- Lantern — concept reference
- Lantern — studio
- Lantern — evening
- Bottle — concept reference
- Bottle — morning desk
- Tested runtime
- Codex native image generation/editing · macOS arm64
- Pinned dependency
- OpenAI guidance (Apache-2.0); original Skilly scenario worksheet and approval gates
- Checks completed
- Saved prompt and reference identities, byte-preserved PNGs, per-feature visual review, independent new briefs and offline archive integrity checks.
Actual native-tool images, including revised attempts; independent new-brief testing with disclosed failures.
For concept imagery, not manufacturing geometry, certified advertising claims or pixel-exact editing. All sample products are fictional. A rounded-square radio brief initially produced a wider rectangle; that attempt is retained as a failure. Backend model ID was not returned; these are not verified GPT Image 2 or 2.5 benchmarks. A targeted radio repair still missed strict square geometry, so its scene generation was stopped.
From a good brief to a better result.
A reference-led product concept workflow with an explicit preserve/change contract. Approve the reference first, create scene variants from that reference, and inspect component counts, shape and visible details before delivery. The published package has its own version, tested environments, and included outputs listed alongside this overview. Any earlier draft samples below retain their original verification scope.
What you bring
- Authorized reference image, or a request for a fictional concept
- Scene, allowed changes and features that must stay fixed
What you’ll make
- Original PNG concept references and scene examples
- Editable preserve/change worksheet and scene prompts
- Per-feature review, hashes and rejected attempts
The intended workflow
- Inspect an authorized product reference and lock visible invariants
- Pass the reference gate before creating independent scene variants
- Compare components, shape and details; disclose or reject drift