AI Page Assembly¶

Overview¶

AI page assembly generates page layouts from natural language descriptions. Instead of manually configuring sections, blocks, and data sources, you describe what you need and the AI builds it — referencing your real collections, triggers, functions, and orchestrations.

There are two flows: Guided Assembly for creating new pages, and Improve This Page for modifying existing pages from the editor.

Guided Assembly (New Pages)¶

From the Pages list, click Guided Assembly in the top-right corner to open the wizard.

Step 1: Select Collections¶

Choose one or more collections that the generated pages should work with. The AI uses these collections — including their field names and types — to build appropriate data sources, table columns, form fields, and filters.

Step 2: Describe Your Goal¶

Choose a mode:

• AI Assembly — describe what you need in plain language. Examples: "A customer management dashboard with an approval workflow", "CRUD pages for managing orders", "A detail page with related line items."
• Quick Templates — generates standard CRUD pages immediately (list + form per collection, plus a detail page if your goal mentions "detail" or "view", and a dashboard if you selected 2 or more collections). No AI involved, no polling — results are instant.

Step 3: Review Proposals¶

Each proposal is shown as a card with:

• Page name and type (list, detail, form, or dashboard)
• AI badge indicating the page was AI-generated
• Rationale — a brief explanation of what the page does and why it was generated
• Accept button — click to create the page as a draft

You can accept individual proposals — you don't have to accept all of them.

Iterating on Proposals¶

After proposals appear, a text input lets you describe changes:

"Add a kanban board grouped by status" "Include a related orders table on the detail page" "Remove the form page and add a dashboard instead"

Click Regenerate and the AI produces additional proposals using the full conversation history. New proposals are appended to the existing list — previous proposals aren't replaced, and already-accepted pages aren't regenerated.

You can iterate as many times as you need within the same session.

Improve This Page (Editor)¶

For pages that already exist, open the page in the editor and switch to the AI tab.

Submitting Instructions¶

Enter what you want to change in the text area (up to 5,000 characters). Press Cmd+Enter (Mac) or Ctrl+Enter (Windows/Linux) to submit, or click the button.

Examples:

• "Add a chart showing orders by month"
• "Change the table to a kanban board"
• "Add a delete confirmation modal"

The panel keeps a history of your last 5 instructions — click any to reuse it.

Reviewing Changes¶

After generation, the panel shows:

• Summary of what changed
• Before/after comparison — section count and block count
• Full definition — expandable JSON view of the complete modified page

Click Accept to apply the changes to your draft, or Discard to reject them. Either way, the panel resets so you can submit another instruction.

What the AI Knows¶

The AI has access to your workspace context when generating pages:

This means the AI can generate a data table backed by a real collection, a button that invokes an actual HTTP trigger, and navigation actions that link to existing pages — without hallucinating references.

Block Types¶

The AI can use any of the 18 block types available in Pages:

Validation and Auto-Fix¶

Every AI-generated page definition goes through automatic validation before being presented to you:

Checks:

• All data source references point to real collections or smart queries in your workspace
• Field references match actual fields on their collections
• Action targets (triggers, orchestrations, pages) exist
• Navigate-to-page actions reference existing or proposed pages

Auto-fixes:

• Regenerates all block IDs to ensure uniqueness
• Injects missing recordSlug into data sources (looks up the correct slug from the collection ID)
• Adds data. prefix to sort and filter fields that reference user-defined data (system fields like createdAt are left unprefixed)

If validation fails, the system retries automatically — up to 3 attempts — feeding the specific errors back to the AI so it can self-correct. You only see the final result. If some proposals still fail after retries, they're discarded and you'll see a toast notification with the count.

After Accepting¶

When you accept a proposal:

1. A new page is created with the name, slug, and type from the proposal
2. The AI-generated definition is saved as the initial draft — it is not published automatically
3. You can edit the draft further in the page editor before publishing

This means you always have a chance to review and adjust before anything goes live.