Not read at runtime. How the skill was built and why it makes its choices.
\nDistilled from the Blender 2D animation exploration on scenario.com: a 7.7 second side-scroller shot in three looks (a red panda in a golden-hour jungle, a black cat on a moonlit Halloween street, a snow leopard cub in a winter valley), ten Scenario generations per look, built and rendered by one Blender script. The page offers the skill as a download; this directory is that skill brought to the repository's authoring contract.
\nThe exploration used one text-to-image model for every picture and one image-to-video model at 768p for the run clips. The body names neither and discovers both lanes with recommend (txt2img, img2video), because availability differs per team and both generations will be superseded. The page's credit figures are left out for the same reason: the body prices with dry_run.
Every paid step (ten generations per look) is a Scenario MCP call; Blender only lays out and renders, locally. The expert tier exempts its skills from the MCP rules, which are exactly what this skill must be graded on, and the Blender family's specialists share the lead's toolkit and provenance, neither of which applies here. scenario-patina-retexture set the precedent: MCP generation, then bundled Blender scripts run headless. The name keeps the download's, so the exploration page and the repository agree.
search to recommend; payload field names hedged as schema examples.scenario-side-view-game-kit and scenario-sprite-animation named as the sprite-strip alternatives. The download's other related skills are not named: the model-family ones point at a model, and the workflow hands nothing to the rest, whose bodies would only push the fan-out past the re-attach pool.build_scene.py --live used to empty whatever file was open. It now refuses unless the file is new (nothing beyond the startup cube, light and camera) or the script's own output. bpy.data.is_dirty cannot make that call: Blender reports a new file as dirty.run_loop.py now prints run_canvas_px and build_scene.py reads it with the jump still's own height.references/pipeline.md already records them for anyone changing the scene script.reframe_stills.py is a byte-for-byte copy of the one in scenario-side-view-game-kit, kept so each skill installs alone; tests/scenario-blender-2d-animation/test_reframe_stills.py fails when the two drift, so a fix to one reaches the other. run_loop.py shares that kit's loop-window idea (seam over in-window motion) but adds the border key and writes a Blender image sequence instead of a pixel strip.
The suite runs the five helpers on synthetic inputs. test_build_scene.py builds, bakes and saves a synthetic set in a real Blender and is skipped unless BLENDER names the binary or blender is on the PATH, which CI does not provide; it passed on Blender 5.0.1.