The script you already wrote, only it runs

From a document to a tree in one run

Call-script documents are written for people, not machines: they have headings, parenthetical notes, long lists and sometimes contradictions. Generation reads the document and assembles a graph, then runs a deterministic repair pass that makes sure the result is runnable at all — that there is exactly one starting point, that every branch leads somewhere, that every field the script asks to capture is actually declared, and that no dead end leaves the caller hanging. Layout on the canvas is automatic, so what opens in front of you is a readable diagram rather than a pile of nodes. The result is a draft you approve — not something that goes live on its own.

• The system classifies the kind of script document and picks the matching adherence mode
• Fields the script collects are declared automatically as conversation variables
• Long checklists collapse into one readable lane instead of dozens of nodes
• A validation pass runs before the tree is opened for editing

Guided vs strict: two levels of adherence

In strict mode the agent says the lines as written, word for word, and branching is decided by the engine — the right choice when the wording itself is the requirement: disclosures, mandated phrasing, a script signed off by a regulator or legal. In guided mode the agent keeps the structure and the order of the stages but phrases naturally and responds to nuance — the right choice when the goal is a conversation that sounds human rather than read out. Even in guided mode the decision about where the conversation goes stays with the engine: guided phrasing, strict decision tree.

• Strict: verbatim lines, zero improvisation in the wording
• Guided: natural phrasing, the same stage order and the same branches
• In both modes branching belongs to the engine, not the model
• You can switch mode after generation and test both

What the flow can do mid-call

A node in the tree is not just a line. It can capture a field and store it on the conversation, call an external tool or a data table and branch on the answer, transfer the caller to a person — dialling a specific extension included — or end the call with a defined status. That is what turns the tree from a speaking script into a business process: "check whether the customer exists in the system; if so continue to booking, if not collect details and open a ticket". You can also see in real time which node a live call is on, which is the single most useful tool for finding where a script breaks.

Edit, test, and a version you can go back to

After generation the tree is yours: drag nodes, fix wording, add the branch the document missed. Testing happens on a browser call — no dialling required — and you watch the cursor move between nodes as you speak. Every content-changing save mints a full version, including who saved it and when, and rolling back is one click that mints a new version rather than deleting history. In practice that means you can try a change to the script without fearing you will have to reconstruct it from memory.

Limits worth knowing before you start

Three. First, the quality of the tree is the quality of the document: a script with internal contradictions, or one that says "use your judgement" instead of stating a rule, produces a confused tree — which is a good signal that the script itself needs sharpening. Second, there is a technical dependency the system enforces for you — guided mode combined with a speech-to-speech engine only works with one model family, and an incompatible choice is blocked with a one-click fix offered, rather than leaving you with an agent that speaks beautifully and ignores the nodes. Third, a strict script always runs on the cascaded pipeline (STT→LLM→TTS) rather than speech-to-speech, because verbatim lines cannot be guaranteed otherwise — the trade-off is slightly longer response time.