🤖 The AI Assistant That Builds Your Graph

In the Text tab of the AI Generation window (Cmd+Shift+G), ask the AI to build your story — not to tell it to you. It works in the graph you have in front of you: blocks appear and wire themselves together while it answers.

⚠️ The Model You Pick Must Be Able to Use Tools

Not all of them can. When yours cannot, a banner warns you: it cannot touch the graph and will just answer in the conversation.

🔺 Switch model before sending your request.

🧱 Building the Story

• Adding, removing and renaming blocks, within the limits of your subscription.
• Connecting two points, flow (in / out) as well as values, and disconnecting them.
• Writing the content of an input: the text of a Narrator or an Alert, a number.
• Setting the options of a Choice block: each one becomes an output, wired up in the same move.
• Creating, renaming and removing characters, then writing the lines of a Dialogue or Choice with dialogue block. A line with no character becomes narration.
• Creating, renaming and removing variables (Text, Number, Boolean) and pointing an input at one, so the story remembers.
• Writing the page of an HTML5 block (mini-game, form, simulation), if your plan allows it.
• Configuring an integration block: on a Baserow block, it offers the databases, then the tables, and wires the block to the one you name.

🗂️ Organising and Navigating

• Moving a block, grouping blocks into a named group, editing or removing that group (the blocks stay).
• Showing a block: it centres and selects it to point it out to you.
• Listing the project's graphs, working on a different one, or creating a module with its graph and its start block.
• Reading back the graph and the content of an input, then checking the project: what would stop the story from running, outputs that lead nowhere, inputs that read nothing, HTML5 pages with no way out.
• Saving the project to the server. Before that, its work only lives in your browser.

🔺 It cannot create a connection the editor would refuse: the same checks apply to it.

🔎 Everything It Does Is Shown

Every action is written on a line in the conversation, in plain English: "Adding the character Alex", "Setting 3 choices", "Saving the project". Unfold it for the detail; a failure shows in the error colour. This is deliberate: you must be able to read what was done.

✍️ An Example Request

Create the beginning of my story: the hero arrives in front of a locked door, Nora speaks to him, then a three-way choice — knock, pick the lock, walk away. Each branch leads to a short scene. Put the whole thing in a group called "Prologue".

No need to specify the graph: it is the one that is open.

↩️ Undoing a Whole Turn

Under the message that started the work, an Undo this turn button puts the project back as it was before: the unit being undone is the complete turn, not each individual change. Redo this turn puts it back as the assistant left it.

🔺 That restore also overwrites whatever you changed by hand since that message, and empties your usual undo history: your earlier steps are gone.

🚧 Known Limits

• Layout stays imperfect. It is supposed to lay its graph out from left to right, but nothing guarantees it: blocks can overlap. Ask it to move them.
• Progress during the turn. Blocks appear in the graph before the conversation comments on them: do not conclude too early that nothing is happening.
• It does not know the text already written. It knows which blocks exist, not what they contain: it reads an input back before changing it, and a rewrite replaces the entire content.
• Turns are bounded. After twelve rounds of tool calls, it stops; what it has built stays in the graph.
• Characters have no image: a name, a biography, nothing more.

→ Next step: read the Choice block page again, then ask the assistant to complete a branch you left open.

Updated on: 23/09/2026