Storyflow has one AI assistant for the whole app. Not a separate one for documents, another for the canvas and a third for Tactics: the same assistant, which knows where you are and what you are looking at.

That matters more than it sounds. Because it is one assistant with one memory of the conversation, you can research a topic on the dashboard, move to a board and ask it to build from what you just discussed, then open a document and have it write the same material up, without re-explaining anything.

Where to open it

Click the AI button in the top navigation. The panel opens beside your work and stays with you as you move between the dashboard, boards and documents.

What it can see

The assistant is not guessing about your workspace. On any request it knows:

• Where you are: the project, the board or the document currently open.
• What you have selected: elements selected on the canvas, or text highlighted in a document, become the subject of the request.
• What you attached: anything you added to the message with the + button.
• Who you are: your AI settings, meaning what you have told it about yourself and any knowledge items you have saved.

Everything else is closed until it opens it. A file on your board is just a filename until the assistant reads it, and a link is just a URL until it fetches it. This is deliberate: it looks things up rather than assuming, which is why it can say "that PDF says X" instead of inventing something plausible.

How a request is carried out

1. It reads first. For anything about your own content, it fetches the current state of the board or document rather than relying on what it saw earlier in the conversation.
2. It researches when the answer depends on the world. Prices, tools, events and anything with a date benefit from a live web search, and it cites the sources it used.
3. It works step by step, applying as it goes. Cards appear on your board one at a time as they are made. There is no preview step and nothing to approve: the work is real content from the moment it lands.
4. It asks when the request is genuinely ambiguous, using a short form in the chat rather than guessing and building the wrong thing.

Because the work is applied as it goes, undo is how you reverse it. One undo reverses a turn, and anything deleted stays in the board's trash for 7 days.

Steering it

Select first. Whatever is selected when you send a message is what "this", "these" and "it" mean. Selecting three cards and asking to expand them edits those three rather than creating three new ones alongside.

Attach what it should work from. The + button in the composer adds context or a Tactic to your message. Attaching the brief, the transcript or the reference material is the single biggest lever on output quality.

Use Tactics as frameworks. Attaching a Tactic gives the assistant the framework's structure, so asking it to work "using this" produces output shaped by that method rather than a generic answer.

Stop or queue. The send button becomes a stop button while it works. You can also type your next message mid-run and it will be picked up in turn.

Chat history

Conversations are saved per project, so you can reopen an earlier thread and carry on with its context intact rather than starting over. Long-running work keeps its thread instead of scattering across one-off chats.

Personalization

In AI settings you can tell the assistant about yourself, how you want to be addressed and how you like answers written, and save knowledge items it should treat as reference facts. These apply to every conversation, so you stop repeating the same preamble.

Best practices

Give it context before you give it instructions. A board holding the real material beats a long prompt describing it.

Ask for research on anything factual. Built from memory, a factual board reads generic and dated. Built from research, it carries names, numbers and specifics.

Say what good looks like. Length, tone, audience and format up front save a round trip.

Correct the thing, do not restate the request. "Shorten this to two sentences" beats retyping the original prompt with one word changed.

Iterate in one thread. The conversation is context. Refining in place is better than a fresh chat per idea.

Privacy

Your projects are private by default and your content is not used to train AI models. See Data Privacy & AI Training for the detail.

Plans and usage

AI usage counts against your plan's allowance. What a higher tier buys you is more generation, not access to different features.