AI Assistant

Available

The AI Assistant is a chat-based panel that helps you plan, write, and refine your demo content. It can read your sketches, generate planning rows, improve narrative text, and make surgical edits — all through natural conversation.

The AI chat panel alongside the sketch editor

Chat Panel

The chat panel sits alongside your editor and provides a familiar messaging interface:

• Send messages by typing in the input field and pressing Enter
• Streaming responses appear in real-time as the AI generates output
• Message history is preserved for the current session
• Session persistence — chat history is saved per project, so you can pick up where you left off

Built-in Agents

CutReady includes four specialized agents, each tuned for a different task:

Agent	Purpose	Best for
Planner	Sketch planning and structure	Generating planning rows, organizing demo flow
Writer	Narrative refinement	Improving voiceover text, making scripts engaging
Editor	Surgical edits	Precise changes to specific cells or rows
Designer	Elucim visual generation	Creating framing visuals for sketch rows

Switch between agents using the agent selector dropdown at the top of the chat panel. Each agent has a distinct system prompt that shapes its behavior.

Tool Calls

The AI doesn’t just chat — it can read and modify your project files through tool calls. When the assistant needs to interact with your content, it invokes backend tools:

Tool	What it does
list_project_files	Lists sketches, notes, storyboards, and optionally screenshots
read_note / write_note	Reads or writes markdown notes
read_sketch / write_sketch	Reads a sketch or replaces its planning rows
update_planning_row	Updates a single row by index without rewriting the whole sketch
read_storyboard / write_storyboard	Reads, creates, or updates storyboard metadata and sequence
design_plan	Saves the Designer agent’s plain-English visual brief for a row
set_row_visual	Saves or removes an Elucim DSL visual for a row
delegate_to_agent	Hands a focused subtask to another built-in or custom agent
fetch_url	Fetches a URL and appends a structured links section for follow-up browsing
search_web	Searches the public web when Internet Search is enabled
recall_memory / save_memory	Reads or saves workspace memory for reusable project context

Tool calls happen automatically as part of the conversation. You’ll see the results reflected in your editor in real-time.

Activity Panel

The Activity Panel provides a real-time log of everything the AI does behind the scenes:

• Actions are displayed newest-on-top for easy scanning
• Errors and warnings are color-coded for quick identification
• Tool calls, file reads, and modifications are all logged
• Sparkle button actions (silent prompts) appear here instead of in the chat

Tip

Check the Activity Panel when you want to see exactly what the AI changed. Every tool call is logged with its parameters and result.

Reference Autocomplete

Type # in the chat input to trigger autocomplete for project file references. This lets you mention specific sketches, notes, or storyboards inline so the AI knows exactly which document you’re referring to.

References appear as color-coded chips that match the file type:

• Sketches — violet
• Notes — rose/pink
• Storyboards — emerald/green
• Web URLs — blue

Referenced files are shown as footnotes at the end of messages, making it easy to see what context the AI is working with.

Web Context and Search

The assistant can use public web context in two different ways:

• URL references — Paste or reference a URL and the assistant can call fetch_url to read that page. The result includes a “Links on this page” section, so the agent can follow links when you ask it to.
• Internet Search — Enable Settings → AI Provider → Internet Search to expose the search_web tool for current public information. Search is opt-in, and project content is not sent to a public search query unless you explicitly ask for that.

Custom Agents

Beyond the built-in agents, you can define custom agents in Settings with your own system prompts. Custom agents appear in the agent selector alongside the built-in options and are useful for specialized tasks like writing for a particular audience or following specific style guidelines.

Note

The AI assistant requires an API key or OAuth connection configured in Settings → AI Provider (global) or Workspace Settings → AI Provider Override. See the Settings page for setup details.

AI Change Highlighting

When the AI modifies sketch planning rows, the affected rows are highlighted to show exactly what changed:

• Modified rows get a visual highlight with inline diffs showing the before/after text
• Each highlighted change includes an undo button to revert individual modifications
• The “Show AI Changes” button in the toolbar lets you re-view the last set of AI-applied highlights at any time

This makes it easy to review, accept, or reject AI suggestions row by row.

Vision Support

The AI assistant can process images alongside text. When enabled, screenshots from your sketches and notes are sent to the model for visual context:

Mode	Behavior
Off	No images sent to the AI
Notes only	Images embedded in notes are included
Notes + Sketches	Images from both notes and sketch rows are included

Configure vision support in Global Settings → AI Provider or override it per workspace. Vision requires a model that supports image inputs (e.g., gpt-4o).

LLM Compaction

Long conversations can exceed a model’s context window. CutReady handles this with smart conversation compaction:

• When the conversation approaches the context limit, older messages are summarized and condensed automatically
• The AI retains key context and decisions from earlier in the conversation
• Compaction is transparent — the conversation continues seamlessly

CutReady uses model capability metadata when available, including the model’s context length, so compaction can happen at the right point for the selected model.

Responses API

CutReady automatically routes requests to the OpenAI Responses API when using codex or pro-series models. This happens transparently — no configuration needed. For standard models, the regular Chat Completions API is used.