Skip to content

T8mars/minimax-h3-prompt-skill-T8 / skills/facial-action-unit-calibration

facial-action-unit-calibration

Verified

Create original short videos using the 表情校准|固定身份下的动作单元序列 Creative DNA for MiniMax H3 or Seedance 2.0. Use when the user wants this causal, camera, motion, rhythm, or payoff structure with new subjects and surfaces.

FileKindSize
SKILL.mdskill md2.6 KB
agents/openai.yamlconfig493 B
references/summary.mdmarkdown1.3 KB
references/template.mdmarkdown4.5 KB

references/template.md

View on GitHub
# 表情校准|固定身份下的动作单元序列 Template

## Mechanism

A facial calibration sequence becomes reusable when identity, lighting, background and camera axis stay fixed while distinct facial action units change one group at a time. Brief micro-resets preserve comparability, and a late controlled crop can test fine identity stability without turning the sequence into a narrative performance.

## Invariants

### inv-01

- Rule: Lock one face, background, lighting direction and camera axis before any expression change.
- Purpose: Creates a reliable comparison baseline.
- Failure if removed: Changing the portrait setup makes action units impossible to compare.

### inv-02

- Rule: Separate the expression into readable local units rather than morphing the whole face continuously.
- Purpose: Makes each change diagnosable and reusable.
- Failure if removed: A blended emotional performance hides which feature changed.

### inv-03

- Rule: Insert at least one partial reset toward the baseline between larger changes.
- Purpose: Prevents accumulated distortion and preserves identity.
- Failure if removed: Without a reset, changes can drift into a different face.

### inv-04

- Rule: Keep head motion subordinate to eye, brow, mouth and cheek evidence.
- Purpose: Maintains a stable portrait while allowing orientation tests.
- Failure if removed: Large head turns replace facial proof with pose variation.

### inv-05

- Rule: Use a late closer crop only after identity has been repeatedly proven.
- Purpose: Tests detail stability and provides a clear endpoint.
- Failure if removed: Opening in extreme close-up removes the identity baseline.


## Variable Slots

| Slot | Constraint |
|---|---|
| lead_subject | An original stable subject or pair with a readable role. |
| setting | A new setting with fixed spatial anchors and no source-specific surface. |
| escalation_variable | One causal variable that changes through visible stages. |
| performance_rule | A restrained acting or motion rule that keeps the mechanism legible. |
| camera_grammar | An information order that proves each state without copying source shot boundaries. |
| payoff | A clean final physical, informational or social consequence. |
| sound_family | An original hierarchy of diegetic cues; no unsupported source-audio claim. |

## Required Human-Facing Anchors

- Lock one face, background, lighting direction and camera axis before any expression change.
- Separate the expression into readable local units rather than morphing the whole face continuously.
- Keep head motion subordinate to eye, brow, mouth and cheek evidence.
- Use a late closer crop only after identity has been repeatedly proven.

## Anti-Copy Boundary

- older bald grey-bearded man
- black shirt
- dark paneled studio portrait
- source facial-expression order
- source final mouth close-up

## Transfer Tests

1. A young ceramic courier in a cool medical kiosk performs a nonverbal visor-fit calibration: neutral baseline, one-sided brow lift, isolated lower-eyelid tension, closed-lip jaw shift, then one nostril-and-cheek action. A brief neutral reset separates the tests, and the only late close crop verifies the unchanged eye-corner and temple visor seam.
   - Preserved: inv-01, inv-02, inv-03, inv-04, inv-05
   - Changed: lead_subject, setting, escalation_variable, performance_rule, camera_grammar, payoff, sound_family
2. A handcrafted wooden puppet in a sunlit rehearsal booth tests five mechanical face controls in a new order: left brow plate, paired eye shutters, closed jaw sliding sideways, one cheek lever, then a full neutral reset. The camera ends on the unchanged eye hinge and wood grain, never on the mouth.
   - Preserved: inv-01, inv-02, inv-03, inv-04, inv-05
   - Changed: lead_subject, setting, escalation_variable, performance_rule, camera_grammar, payoff, sound_family

## Model Skeletons

### MiniMax H3

    integrated_multimodal_description: [new subject, environment, ordered actions, camera, light and final state preserving every anchor]
    overall_soundscape: [diegetic ambience, action cues and causal payoff]
    non_diegetic_music: [original music arc or N/A]

### Seedance 2.0

    任务:[原创实例意图]
    总时长约[4–15]秒。
    主体定义:[稳定人物、物体与空间关系]
    事件或镜头:[按因果顺序实现全部锚点]
    整体风格与画质:[原创媒介、光影、色彩与声音]
    约束:[连续性、物理逻辑、反复制与结尾状态]

Adapt shot count to the user's concept. Preserve the causal mechanism, not a fixed storyboard.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 196kUpdated Claude CodeCodexVerifiedInstall