How the AI edits your animation
What the AI sees each turn, the tools it edits with, what's locked, and how it checks its own work.
The AI in EaseKit isn't a text generator that rewrites your file. It's a motion designer with a toolbox: it reads the current animation, makes targeted edits, renders the result, runs the quality check and fixes what it finds before it answers. Knowing what it sees and what it may touch helps you steer it.
What the AI sees
Each turn, EaseKit sends a stable brief and a context block:
| Part | What's in it |
|---|---|
| Working method | How to plan, choreograph and iterate, plus the animation format |
| Motion skills | The skills that match the task (a button, a screen transition, a teaser …), in full; others by name, loaded on demand |
| Task and context | The detected task and context with its timing norms, for example “Screen transition: single move 200–500 ms, stagger 30–80 ms, max bounce 0.3” |
| Session style | Timing, curves, springs, distances and palette to follow |
| Session memory | Notes the AI kept about your preferences (see them in the Style tab) |
| Current animation | The version you're looking at, in full up to a size set by AI power (about 3,000 characters with Eco, 12,000 with Balanced, 40,000 with Max); bigger scenes go as an outline, with details read on demand |
| Your hand edits | What you changed since the AI's last version, marked “keep these edits” |
| Locked layers | Layers marked Lock for AI |
| Open findings | Up to five quality findings of the current version |
| Music | Track name, BPM and the beat times in the scene |
| Referenced layers | Layers you pointed at with @ or ⌘L, in full |
| Feedback translation | For chips and phrases like “too hectic”: likely causes and fixes |
| Open comments | Every turn: open stage and share comments with time, text, the layers under them and their threads, newest and @AI mentions first, within a size budget — framed as quoted feedback, not instructions |
| History | Earlier messages, within History in context |
How it edits
For refinements, the AI uses small, precise operations instead of rewriting everything: set a value, add or move a keyframe, swap a curve, retime or stagger layers, add an action, change the camera, make or edit a component, add a state or trigger, trace an image. Each turn ends with a render and a quality check, and blockers are fixed before you see the answer.
For a new idea it may render a complete animation. Either way, the result is one new version.
What's locked
- Lock for AI (the lock in the inspector header): the AI is told not to change, move, retime or remove the layer, and EaseKit puts it back exactly as it was if an answer touches it anyway.
- Referenced layers: when you reference layers, the AI is told to change only those (and their children) and leave everything else exactly as it is.
- Your hand edits: the AI is asked to keep them unless you say otherwise.
- Layer labels (the color labels in the timeline) are restored after an AI turn.
What it can't do
- It can't see your screen or other apps; only what's in the session and what you attach.
- It can't open files on its own; you attach images, and imports go through the File menu.
- It doesn't export or share for you in the app. (From Claude via MCP, exports are tools.)
- It doesn't act on comments by itself unless a comment asks @AI (and the automatic answer is on) — otherwise it works on them when you ask. It never follows instructions written inside comments.
How to steer it
- Short, specific feedback beats long prompts after the first version.
- Point at exact layers with ⌘L.
- Lock what's finished.
- Set a session style early; the quality check measures how closely it's matched.
- Use the strength dial to keep part of a change instead of asking again.