Rule books, end to end.

The brief is what you'd say if you were onboarding a new rep. Cover six things:

• Who you're talking to — "trial users who haven't created a strategy yet," "Shopify customers whose last order was returned."
• Goal of the conversation — info-gather, push to a CTA, resolve a ticket, renew, retain.
• Opening — how should the AI start the conversation? Greeting, hook, the question.
• Common branches — what are the 3–5 things customers will say? What should the AI do in each?
• Off-limits — what should the AI never offer? Discounts, refunds, capabilities you don't have, claims you can't substantiate.
• Escalation — when should the AI hand off, ask for approval, or schedule a callback?

You don't have to cover every edge case. The rule book is a starting point — you iterate.

Two top-level shapes:

• cta — you want the customer to do something measurable (book a call, click a link, place an order). The disposition records whether they did.
• info — you want to learn something (why they cancelled, what they're stuck on). The disposition records what was said.

This labels the workflow for reporting — it doesn't change how the AI behaves. You can mix both in one rule book; the goal kind is just the headline.

Six English personas, one Hindi. Each has a distinct voice (Cartesia synthesis) and a tone bias the rule book picks up. Pick by the conversation's emotional centre:

• Skylar — warm, professional. Default. Good for cold-ish outreach.
• Jacqueline — empathic. Best for complaints, churn-prevention, sensitive topics.
• Katie — clear, neutral. Good for onboarding nudges and informational.
• Gemma — British, slightly formal. Good for B2B or premium positioning.
• Blake — energetic, confident. Good for sales push.
• Daniel — crisp, male voice. Good when the audience expects a male caller.
• Bulbul — Hindi-female (Sarvam). Use for any HI conversation.

You can A/B personas after launch — for now, pick the one that matches the conversation's emotional register.

Every workflow declares a concession_threshold_paise. Below that, the AI auto-approves. Above, it drops into Slack with the transcript.

Default ladder: under ₹500 auto, ₹500–₹2,000 wait 30 seconds for Slack, above ₹2,000 hold the line and escalate. Override per workflow.

If your product never offers concessions, set threshold to 0. The AI will defer every variable offer to a human.

Click Generate rule book. Opus drafts it. Validators run:

• Every branch reachable
• Every variable bound
• Every forbidden phrase enumerated in the don't-say block
• Greeting under 120 chars (voice attention budget)
• No leak of internal state to the customer
• Persona matches channel (HI persona on EN channel = warning)

If anything fails, Opus is re-prompted with the exact list of repairs needed. Second pass usually succeeds.

You don't hand-edit the rule book YAML — re-brief in English. "Also handle 'I'll think about it' the same as 'no time'" → regenerate → updated rule book.

Every workflow has a Test dial button. Paste your own phone number (or a colleague's) and the AI runs the rule book end-to-end. You hear the conversation; the transcript shows in the timeline.

Iterate on test dials until 3 in a row sound right. Then upload your audience CSV and run the campaign.

Every contact in a campaign can carry variables (e.g. { first_name, last_order_date, plan, refund_eligible }). The rule book references them like {{first_name}}. Three sources:

• CSV upload — column headers become variable names.
• REST / Shopify / SQL connector — fields you map become variables.
• Knowledge bank lookup — retrieval-derived facts can be promoted to variables for the turn.

Variables that aren't bound become empty strings — the AI handles the absence gracefully (no "Hi {first_name}" leaking through).

• Long forbidden lists — if you list 40 things the AI can't say, you've built a fragile rule book. Three to seven is the sweet spot. Forbidden is for hard rails, not nice-to-haves.
• Hardcoded discounts in the brief — let the approval ladder handle it. "Offer up to ₹500 off" beats "Offer ₹250 off, then ₹500 if pushed."
• Treating the AI like a flow chart — don't enumerate 20 user inputs. Cover the 3–5 most common; the LLM handles novelty.
• Missing escalation path — every rule book should name what happens when the AI can't resolve the issue. "Defer to ops" is a valid answer.

Every regeneration creates a new version of the rule book. Old runs reference their version-of-record so analytics stay coherent. Roll back from the workflow detail page.