Plan, make, and review art assets
Start by describing the asset's job: what it is, how it will be viewed, what needs to move, and where it will run. A small inventory icon, a close-up character, and a distant environment prop have different production and review needs.
Ask the agent to read STYLE.md and the selected engine before proposing tools. A useful starting request is:
Plan a readable environment prop for this scene. Check the project's visual style, selected production pipeline, target budget, and any configured asset library. Show the proposed steps and review points. Do not generate, download, or promote an asset yet.
Choose a style and a production method
STYLE.md combines one or more visual styles with exactly one primary art pipeline. The style describes appearance; the pipeline describes how the asset is produced. New style contracts default to PROOF_TO_MODEL unless another pipeline is explicitly selected.
| Pipeline | What it is for |
|---|---|
CLOUD |
Generated images through a compatible capability available in the session. |
LOCAL |
Direct Bubo generation using explicitly prepared local tools and models. |
PROOF_TO_MODEL |
Reviewed source images and a complete six-view set leading to native reconstruction and finishing. |
CODE |
Three.js models authored as JavaScript/TypeScript source; this path requires THREEJS. |
BLENDER |
Blender 4.5 LTS authoring, validation, renders, and exports under its working-copy controls. |
Choose a pipeline because it suits the requested result and is available. If it fails or lacks prerequisites, the agent must explain that and request direction rather than silently switching methods. A task-specific pipeline override does not rewrite the project's primary choice.
For a newly agreed cloud illustration style, these host-root commands generate and validate the contract:
bash .hephaera/commands/generate-style.sh --pipeline CLOUD CEL_SHADED_COMIC
bash .hephaera/commands/validate-style-contract.sh
That records the choice. It does not authorize generation or install a provider. CLOUD and LOCAL execution also follow MEDIA.md, with explicit input rights, budgets, staging, and review.
Review the right thing at each step
For PROOF_TO_MODEL, original source images establish what is actually known about the subject. Generated views and model renders cannot become evidence of the original anatomy or structure. Review originals together with front, right, back, left, top, and bottom proofs, and approve the exact complete set before reconstruction. Revising one view invalidates that set's approval.
Creating missing proofs through CLOUD or LOCAL is a separate generation request with its own allowance. Reconstruction and finishing need the prepared native runtime and tools; choosing the pipeline does not prepare them automatically. See proof to model.
Keep candidates outside runtime asset folders until review. Check silhouette, materials, topology, scale, performance limits, and the intended game camera. Approval of a picture is not approval of a rig, a motion range, or runtime behavior. Promotion into the project is a separate explicit step for the accepted version.
Reuse existing assets safely
External libraries are optional. If STYLE.md already configures one and the task needs a 3D asset, inspect it before making a replacement. The source library stays read-only. Copy a selected asset or complete bundle into the repository through the existing copy command before loading or editing it, then check its license and modification rights.
No configured library means no extra startup question. A library also does not bypass the selected pipeline's rules.
Add motion as a later handoff
A static model is not automatically ready to animate. Inspect articulation, topology, scale, rights, and any existing rig first. Model-to-animation creates reviewed motion; model-to-action composes clips into sequences. Neither decides combat damage, targeting, or movement authority just because an animation played.
Read visual pipelines and animation and actions for the detailed stages. If your task is about sound instead, use audio, music, and voice.