blender-character-video - SKILL.md Agent Skill

name: blender-character-video description: "Build and iterate Blender music videos that feature a person/humanoid character: imported low-poly/game rigs, face/eyes, hair, clothing, accessories, body proportions, pose approval, and final-video promotion. Use when the brief includes a person, humanoid, witch, dancer, character model, face, eyes, hair, outfit, body shape, hands, arms, or a character riding/flying/performing in a Blender video." version: 1.0.2 author: Hermes Agent license: MIT metadata: hermes: tags: [blender, character, humanoid, rigging, low-poly, music-video] related_skills: [blender-video-iteration, blender-animation-rigging, blender-modeling-modifiers, blender-style-ps1]

Blender Character Video

Overview

Use this when a Blender music video contains a recognizable person/humanoid character, especially a low-poly or PS1-style imported game model. Character videos are a different class from environment-only videos: the user judges face, pose, outfit, body proportions, hair, hands, accessories, and silhouette first, then motion and environment.

The durable lesson from the first full person-model video: do not treat a character as generic scenery. Build a visible approval loop around one canonical still, preserve the exact approved rig/texture path, and only promote it into a full render after the still and a motion probe prove the character holds together.

References

references/character-asset-audition.md — how to download, inspect, preview, and report character model candidates so the user can review actual local artifacts.
references/imported-character-clothing-fit.md — how to debug imported character bounds and place fitted clothing/accessories on the actual visible body instead of helper/preview mesh extents.
references/anime-vroid-fitted-clothing-lookdev.md — how to turn an anime/VRoid-style candidate with a bad/free-test outfit into fitted low-poly wardrobe options without presenting cardboard proxies as final clothing.
references/anime-character-wardrobe-lookdev.md — how to prove and present quick hair/outfit/accessory changes for an imported anime/VRoid-style character without launching a full render or creating folder sprawl.
references/witch-character-read-and-fit-lookdev.md — session-tested loop for making an imported anime/VRoid girl read as a witch: body-derived fitted clothing, hat/face framing, cloak/broom/moon cues, and broom-pose checks.

When to Use

Use when:

The user asks for a person, humanoid, witch, dancer, rider, singer, operator, or other character-centered scene.
The work involves an imported low-poly/game asset with armature, UV atlas, eye textures, hair, hat, clothing, hands, or accessories.
The user gives art direction about pose, attractiveness, outfit, body proportions, eye color/blinks, hair clipping, arm symmetry, or hand contact.
A character still was approved and must become the final MP4.

Do not use this for environment-only videos where character fidelity is not part of the brief.

Required Companion Skills

Load these as needed:

blender-video-iteration — still/contact-sheet/segment/final mux loop and Gyre reference docs.
blender-animation-rigging — armatures, IK, constraints, bone rolls, pose controls.
blender-modeling-modifiers — real mesh/accessory/proportion edits.
blender-style-ps1 — crunchy low-poly render/upscale/look constraints.

Character-First Workflow

Pick the character representation deliberately. Prefer a real imported rig/mesh when the user asked for a person. Procedural primitive approximations are acceptable for silhouettes, but not for a face/body-centered brief unless the user explicitly wants abstraction.
Build an asset-audition packet before asking for approval. For downloaded character candidates, save the source asset, render front / 3/4 / side previews plus a contact sheet, record import/rig/mesh notes, and show the contact sheet to the user. Be explicit about which candidates were actually downloaded versus only researched. See references/character-asset-audition.md.
Make one canonical approval still. Use a stable path such as renders/<song>/one_frame.png. Keep overwriting it during still approval instead of creating probe sprawl.
Get the global read before details. Confirm face visibility, pose, body silhouette, outfit palette, hair/hat/accessory placement, and camera framing before polishing small texture details.
Patch one visual axis at a time. Separate pose, outfit/texture, body proportion, accessory placement, face/eyes, environment, and mux/export changes.
Use real geometry/material changes. Body proportions, hat fit, hair clipping, shoulder/sleeve silhouette, and hand contact must be changed in mesh/rig/material space and rerendered. Do not use 2D overlays or verbal claims.
Inspect with the approval camera. Low-poly bodies and faces are camera-dependent. A fix that is mathematically symmetric or correct in viewport can still read wrong in the rendered frame.
Promote the approved path into the final generator. Before full render, confirm generate.py imports/uses the same rig, texture atlas, eye overlays, pose helpers, and accessory placement that produced the approved still.
Render a motion probe before final. Sample start/quarter/mid/three-quarter/end frames with the full timeline mapping. Check that hands, hair, eyes, hat, clothing masks, and accessories remain attached during bob/sway/fly cycles.
When the user likes an anime/VRoid character but asks to change the outfit, make a still-only wardrobe option sheet first. Use the actual imported model, keep probes in the existing candidate folder, and label/caveat rough garment overlays as lookdev proxies rather than final fitted clothing. See references/anime-character-wardrobe-lookdev.md.
If the character must read as a witch, verify identity cues before video — but distinguish rejection types. A black fitted outfit is not enough: the approval still must show face/eyes, a readable witch hat, cloak/cape silhouette, broom relation, and moon/night backdrop. If the user says “not witch enough,” treat it as a global silhouette/prop/pose failure and use references/witch-character-read-and-fit-lookdev.md. If the user says the character/scene “will not be a witch” or clarifies they want a cute anime girl instead, stop adding witch props immediately; preserve the liked character and switch the concept back to cute/anime/moonlit styling.
Then render video-only and mux last. Keep audio out of visual debugging; mux source/louder WAV only after visual approval and verify final streams.
Clean after approval. Once final is accepted, preserve deliverables/source/assets/README and remove obsolete probes/frame sequences/options with before/after directory and size checks.

Person-Model Checks

For each approval still or motion probe, explicitly inspect:

Face: eyes visible, not hidden by hat/hair/camera angle; expression/blink state intentional.
Pose: torso/head/limbs support the action; no detached hands, broken wrists, spike/fin limbs, or odd asymmetry.
Silhouette: the whole body reads from the camera; outfit and body shape are visible without becoming blobby or pasted-on.
Hair/accessories: hats, hair, broom/rail/props follow the posed head/body and do not float or clip badly.
Texture stack: main atlas, eye overlays, generated candidates, and environment assets are scoped separately.
Environment separation: moon/star/sky/NASA/public-domain assets are never relinked to the character atlas by broad material loops.

Implementation Patterns

Imported rig first: preserve the chosen mesh, vertex groups, textures, and silhouette. If the imported control layer is cursed, build a clean armature/control rig around the original weighted mesh instead of replacing the character with procedural limbs.
IK/control motion: animate armature controls, IK targets, and accessories together. Do not animate only skinned mesh vertices for hands/arms unless the requested edit is literally a fingertip-only local deformation.
Eye overlays: low-poly characters may use separate open/half/closed textures. When changing eye color/style, update every overlay path (new_bg_eye01/02/03.png and compatibility copies when present), then verify an eye-focused sheet.
Atlas edits: inspect real UV layout before repainting. Generate labeled overlays/cut packs when using image editing, preserve atlas layout, and install only approved texture candidates.
Object/model-space masks: if an atlas-painted neckline/cutout creates W-shapes, seams, diamonds, or panel splits, restore the atlas base and use object/model-space material coordinates or texture-paint strokes for continuous costume shapes.
Anime/VRoid fitted wardrobe lookdev: when the character face/hair is right but the outfit is wrong or a free-test underwear/bikini is baked in, do not show large cardboard panels as “clothes.” Neutralize the base body material/texture if needed, then build body-aware low-poly bodice/skirt/stocking shells or texture-paint fitted garments; keep bad proxy sheets private/cleaned. See references/anime-vroid-fitted-clothing-lookdev.md.
Body proportions: make bounded geometry/rig edits and stop at the visible topology limit — before low-poly geometry turns blobby, clips arms/props, or reads like a separate attached shape.
Fitted clothing on baked-test assets: if an imported anime/VRoid model has baked underwear/test clothing, clean or replace the body material/texture first. Do not make the outfit taller and taller to hide the old texture; that creates poncho/lampshade silhouettes. Add proportionate fitted top/skirt/corset shells only after the base body is clean.
Accessory follow-through: recompute hats/hair/props from the posed head/body, not world coordinates. Moving a hat can reveal hair clipping; pair it with a targeted hair-tuck probe.

Common Pitfalls

Treating a person as scenery. A character-centered brief needs a face/body/pose approval loop, not only a good background.
Procedural overlays instead of the real character. Users can tell when limbs, clothing, or proportions are pasted on. Preserve/import/rig the actual asset when the character is the subject.
Full-rendering from a stale generator. The approved still may come from a helper script while generate.py still uses an older path. Diff/inspect before final.
Fixing local detail before global read. If the user says the person needs a different outfit, skin/hair/eyes, or body shape, do a full palette/outfit/proportion pass rather than defending a neckline-only tweak.
Texture relinks clobbering the sky. Scope atlas relinks to character materials and skip moon/star/sky/environment images.
Still approval mistaken for motion approval. A still can hide detached accessories, bad blink cadence, or arm asymmetry. Render a motion probe.
Ignoring the topology limit. Low-poly body edits have a hard visible limit; push with probes, then stop before the shape breaks.
Candidate research mistaken for downloads. When the user wants to look at models, they need contact sheets or local preview paths. Clearly label downloaded/inspected candidates separately from merely researched links.
Asset-audition folder sprawl. If the user asks to “download a few” or says not to make too many folders, keep all candidate archives, preview stills, the combined contact sheet, and inspect notes in one renders/<project>/assets/candidates/ folder unless extraction/textures force subfolders. Clean exploratory downloads/, models/, probe, or frame folders before reporting.
Whole-asset bounds can lie. Imported character GLB/FBX files may include helper meshes, spheres, floors, cameras, or bounds proxies that are not part of the character. Before fitting clothing/accessories or normalizing scale, inspect per-object bounds and normalize/anchor only the real character meshes (for Theresa: Face/Body/Hair, ignoring Icosphere). If clothes appear to occupy the whole vertical model while the person sits in the middle, stop and fix the anchor/bounds before tweaking garment dimensions.
Fitted outfit is not witch identity. Once clothing fit is fixed, do not present “black fitted bodysuit” as sufficient witch art. Check the global read: face visible, hat readable, cloak/cape silhouette, broom relationship, moon/night backdrop, and non-mannequin pose. If any cue is missing, keep iterating the still.
T-pose can erase character intent. Wardrobe sheets in T-pose are useful for fit only. For broom/flying/witch shots, T-pose arms read as mannequin or costume preview; pose/crop/prop placement must prove the character is riding/flying before video promotion.
Whole-asset bounds mistaken for character bounds. Imported GLB/FBX assets may include helper meshes, preview spheres, collision shapes, or props that inflate the bounding box. If you normalize/place clothes from the whole asset, the character can sit in the middle/top while garments span the whole model. Inspect per-object bounds, ignore non-character meshes, and anchor fitted clothes/accessories to Face/Body/Hair or the actual visible body mesh; see references/imported-character-clothing-fit.md.
Outfit height can become a cover-up bug. When hiding baked-in underwear/test clothing, do not solve it by raising a bodice/dress into a tall panel. Replace/clean the underlying body material first, then build proportionate fitted clothing.

Verification Checklist

Relevant character, rigging/modeling, PS1, and iteration skills loaded.
Downloaded character candidates have source assets, local preview renders/contact sheets, and inspection notes.
For imported characters, per-object bounds were inspected and non-character helper meshes were excluded before camera/scale/clothing placement.
Candidate status is reported clearly: downloaded/inspected, researched only, or blocked.
If the user requested “a few” assets or low folder clutter, candidate files/previews/notes are consolidated under one assets/candidates/ folder and exploratory folders are cleaned before reporting.
Canonical approval still path rendered and inspected.
Face, eyes, pose, hair/accessories, outfit, and body silhouette checked from the approval camera.
Texture/atlas/eye-overlay changes are scoped to character assets and do not touch environment assets.
Any body/accessory/hair/hand/proportion change is a real mesh/rig/material change, not an unverified overlay.
Approved still path is promoted into the final generator.
Motion probe sampled across the song before final render.
Final video-only render and audio mux verified with ffprobe.
README documents the approved character path, final artifacts, known warnings, and cleanup decisions.