Installable skillMIT

Team Motion Design

Unified team skill for motion design. Animation token systems, scroll choreography, GPU-accelerated transforms, reduced-motion fallback. Uses team-worker agent architecture. Triggers on "team motion design", "animation system".

BY catlog22UPDATED Oct 27, 2026
INTENDED FOR◈OpenClawInstall in a listed agent.

instructions.md

---
name: team-motion-design
description: Unified team skill for motion design. Animation token systems, scroll choreography, GPU-accelerated transforms, reduced-motion fallback. Uses team-worker agent architecture. Triggers on "team motion design", "animation system".
allowed-tools: Agent, AskUserQuestion, Read, Write, Edit, Bash, Glob, Grep, TaskList, TaskGet, TaskUpdate, TaskCreate, TeamCreate, TeamDelete, SendMessage, mcp__ccw-tools__read_file, mcp__ccw-tools__write_file, mcp__ccw-tools__edit_file, mcp__ccw-tools__team_msg, mcp__chrome-devtools__performance_start_trace, mcp__chrome-devtools__performance_stop_trace, mcp__chrome-devtools__performance_analyze_insight, mcp__chrome-devtools__evaluate_script, mcp__chrome-devtools__take_screenshot
---

# Team Motion Design

Systematic motion design pipeline: research -> choreography -> animation -> performance testing. Built on **team-worker agent architecture** -- all worker roles share a single agent definition with role-specific Phase 2-4 loaded from `roles/<role>/role.md`.

## Architecture

```
Skill(skill="team-motion-design", args="task description")
                    |
         SKILL.md (this file) = Router
                    |
     +--------------+--------------+
     |                             |
  no --role flag              --role <name>
     |                             |
  Coordinator                  Worker
  roles/coordinator/role.md    roles/<name>/role.md
     |
     +-- analyze -> dispatch -> spawn workers -> STOP
                                    |
                    +-------+-------+-------+-------+
                    v       v       v       v
           [team-worker agents, each loads roles/<role>/role.md]
      motion-researcher  choreographer  animator  motion-tester
```

## Role Registry

| Role | Path | Prefix | Inner Loop |
|------|------|--------|------------|
| coordinator | [roles/coordinator/role.md](roles/coordinator/role.md) | -- | -- |
| motion-researcher | [roles/motion-researcher/role.md](roles/motion-researcher/role.md) | MRESEARCH-* | false |
| choreographer | [roles/choreographer/role.md](roles/choreographer/role.md) | CHOREO-* | false |
| animator | [roles/animator/role.md](roles/animator/role.md) | ANIM-* | true |
| motion-tester | [roles/motion-tester/role.md](roles/motion-tester/role.md) | MTEST-* | false |

## Role Router

Parse `$ARGUMENTS`:
- Has `--role <name>` -> Read `roles/<name>/role.md`, execute Phase 2-4
- No `--role` -> `@roles/coordinator/role.md`, execute entry router

## Shared Constants

- **Session prefix**: `MD`
- **Session path**: `.workflow/.team/MD-<slug>-<date>/`
- **CLI tools**: `ccw cli --mode analysis` (read-only), `ccw cli --mode write` (modifications)
- **Message bus**: `mcp__ccw-tools__team_msg(session_id=<session-id>, ...)`
- **Max GC rounds**: 2

## Worker Spawn Template

Coordinator spawns workers using this template:

```
Agent({
  subagent_type: "team-worker",
  description: "Spawn <role> worker for <task-id>",
  team_name: "motion-design",
  name: "<role>",
  run_in_background: true,
  prompt: `## Role Assignment
role: <role>
role_spec: <skill_root>/roles/<role>/role.md
session: <session-folder>
session_id: <session-id>
team_name: motion-design
requirement: <task-description>
inner_loop: <true|false>

## Progress Milestones
session_id: <session-id>
Report progress via team_msg at natural phase boundaries (context loaded -> core work done -> verification).
Report blockers immediately via team_msg type="blocker".
Report completion via team_msg type="task_complete" after final SendMessage.

Read role_spec file (@<skill_root>/roles/<role>/role.md) to load Phase 2-4 domain instructions.
Execute built-in Phase 1 (task discovery) -> role Phase 2-4 -> built-in Phase 5 (report).`
})
```

## User Commands

| Command | Action |
|---------|--------|
| `check` / `status` | View execution status graph |
| `resume` / `continue` | Advance to next step |

## Specs Reference

- [specs/pipelines.md](specs/pipelines.md) -- Pipeline definitions and task registry
- [specs/motion-tokens.md](specs/motion-tokens.md) -- Animation token schema
- [specs/gpu-constraints.md](specs/gpu-constraints.md) -- Compositor-only animation rules
- [specs/reduced-motion.md](specs/reduced-motion.md) -- Accessibility motion preferences

## Session Directory

```
.workflow/.team/MD-<slug>-<date>/
+-- .msg/
|   +-- messages.jsonl         # Team message bus
|   +-- meta.json              # Pipeline config + GC state
+-- research/                  # Motion researcher output
|   +-- perf-traces/           # Chrome DevTools performance traces
|   +-- animation-inventory.json
|   +-- performance-baseline.json
|   +-- easing-catalog.json
+-- choreography/              # Choreographer output
|   +-- motion-tokens.json
|   +-- sequences/             # Scroll choreography sequences
+-- animations/                # Animator output
|   +-- keyframes/             # CSS @keyframes files
|   +-- orchestrators/         # JS animation orchestrators
+-- testing/                   # Motion tester output
|   +-- traces/                # Performance trace data
|   +-- reports/               # Performance reports
+-- wisdom/                    # Cross-task knowledge
```

## Error Handling

| Scenario | Resolution |
|----------|------------|
| Unknown command | Error with available command list |
| Role not found | Error with role registry |
| Session corruption | Attempt recovery, fallback to manual |
| Fast-advance conflict | Coordinator reconciles on next callback |
| Completion action fails | Default to Keep Active |
| GC loop stuck > 2 rounds | Escalate to user: accept / retry / terminate |
COMPATIBILITY & REFERENCES

Platform guidance

Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.