Imported from Sunwood-ai-labs/sakaki-tesshin-vrm (AGENTS.md). Install upstream with npx skills add Sunwood-ai-labs/sakaki-tesshin-vrm. Copyright stays with the author.

Sakaki Tesshin motion authoring rules

Screenshot-gated VRMA visual QA

VRMA schema validation, a successful build, and successful browser playback are not sufficient evidence that a generated motion is visually correct.

Before publishing any new or changed motion:

• Record the generation provenance explicitly. Distinguish ARDY/motion-model output, LLM-authored keyframes, and manually authored JSON keyframes. Running spec2vrma.mjs only converts a supplied spec to VRMA; it is not proof of AI motion generation. Never label a motion "AI-generated" unless the actual model/engine invocation and its output for that motion are preserved.
• When a remote multi-GPU host is used, preserve the nvidia-smi physical index/PCI-bus map, the CUDA ordering variables, a health response, the exact request JSON, and the raw returned spec. A model merely loading on cuda:0 is not proof that the intended physical GPU was selected.
• Never mark raw ARDY sequences as seamless loops unless their first and last hips positions and rotations were explicitly made continuous and visually checked. Prefer one-shot playback over a visible end-to-start snap.
• For a motion presented as full-body, require intentional lower-body tracks and visible weight transfer. Arm, torso, head, and hips-offset tracks alone are an upper-body keyframe prototype and must not be reported as a generated full-body motion, even if its static silhouette differs from neutral.
• Capture every motion at 25%, 50%, and 75% of its duration from the normal viewer angle.
• Capture the highest-risk hand pose from both the normal and an oblique angle.
• Inspect hands, wrists, elbows, and shoulders for body/clothing penetration, left/right-hand overlap, floating planar palms, hyperextended joints, and a gesture that does not match the motion label.
• Capture the reset/neutral pose from the same camera as the comparison baseline. Every motion must visibly depart from that baseline in at least one audited frame through an intentional elbow bend, wrist displacement, or arm silhouette change. A set of differently labeled screenshots that still read as the same A-pose fails the audit.
• Do not fix hand intersections by simply returning all arms to the neutral pose. The retry must satisfy both gates at once: natural hands and a motion-specific silhouette that is distinguishable in a static overview.
• Treat any visibly awkward frame as a failure even when the VRMA validator and browser playback pass. Revise the source spec, regenerate the VRMA, and repeat the screenshot audit.
• Preserve compact before/after and timeline overview images under artifacts/hand-audit-* as the completion evidence. Raw per-frame captures may remain local and ignored.

Do not report a motion finished until this screenshot-based audit passes for every in-scope motion.