The Workbench

Overview

Some node fields hold a sentence. Others hold the heart of your flow — a long prompt, a JSON schema, a shell script, a whole web page. The Workbench gives those fields a real editor: click the ⇲ expand button on the field's label in the inspector and the whole window becomes an authoring surface for that node. Press Esc (or the back chevron) and you're back on the canvas exactly where you left it.

Everything in the Workbench works without AI configured. When you have an AI provider, an assist bar appears as an extra layer — it never becomes a requirement.

Where It Opens From

Each node type opens its own arrangement of the same surface:

Write, Split, Preview

The header offers up to three ways of looking at a document: Write (the editor alone), Split (editor beside the preview — the default), and Preview (the preview alone). ⌘/ toggles between the preview and wherever you came from.

The preview shows the document as it will actually be used: a prompt renders as the interpolated text the model receives; the HTML Editor renders the finished page. A chip above it names where the variable values came from — your Test Values, or the node's last run.

When a node has several texts (a prompt and its system prompt and its schema), they open together as tabs in the header. Each tab keeps its own toolbar, stats, and view modes — a JSON schema doesn't offer a markdown preview.

Variables You Can Trust

Every {{variable}} in the document is painted live: a token that resolves to something upstream wears your system accent; a token nothing provides turns red with a dotted underline — the single most common flow bug, visible while you type instead of at run time. The sidebar's syntax card names the first broken token outright.

• Type {{ (or click the { } toolbar button) and a picker lists every variable this node can reach, with sampled values from the last run.
• The sidebar's variables rail shows the same list as pills — click one to insert it at the caret.
• How data and variables move between nodes is covered in Payloads.

Test Values and ▶ Run This Node

The AI compositions carry a Test Values tab: the input your test runs start from, authored as a simple form (or raw JSON) instead of borrowed from whatever the flow last produced. Fill it from the last run, from the variables the node can reach, or by hand — the preview fills in from it immediately.

The ▶ button on the toolbar (or ⌘↩) runs this node alone, seeded from your Test Values. Nothing downstream fires — no other node runs, no other trigger trips. Press again to stop a long run.

Each run lands in the Node Workbench History panel — the bar along the bottom of the Workbench. It is this node's private ledger: bench runs never appear in the flow's Recent Runs, and a failed experiment here never marks your flow unhealthy.

AI Assist

With a provider configured, an assist bar sits under the editor. Prose documents offer three preset asks — Tighten, Add examples, Restructure — and every document takes free-text requests (“make the table striped”, “generate test values for this”).

The AI sees your other tabs and how the node sits in its flow — upstream, downstream, and the variables in play — so its rewrites fit your flow instead of guessing at it. A rewrite is always proposed, shown in full, and lands only when you press Apply; applying is a single ⌘Z.

Saving Is Automatic

• Your typing reaches the node on a rolling 2-second autosave — the toolbar's ● Saved dot tells you where you stand. ⌘S saves immediately.
• While the Workbench holds a field, that field in the inspector becomes a read-only mirror — one editor per text at a time.
• Closing always saves first, and the whole visit lands as one undo step on the canvas — a single ⌘Z after closing restores the text you opened with.
• JSON and HTML documents offer Prettify — re-indentation that never touches your keys, values, content, or {{tokens}}. The HTML Editor tidies automatically when it opens, so a page an AI emitted on one line opens readable.

Keyboard Reference