The editor is organized into Personality, Knowledge, Conversation, Goals and History. It autosaves and keeps restorable versions.
Personality — the voice sliders actually steer generation
Formality, warmth, humor, empathy, assertiveness and urgency (0-10 each) aren't decorative: every turn, they're rendered into a literal style description injected in the system prompt ("very formal (9/10)", etc.) with an instruction that every message must match it exactly.
Urgency does double duty. It's not just described in the prompt — it mathematically shrinks how long the agent can stay in small talk before it's forced back to business (see Conversation below), and separately drives dedicated closing-pace guidance in the final stretch of a sale.
Qualification sensitivity (1-10) is a real gate, not a style knob: the higher it is, the more proof of fit the agent needs before it starts positioning your offer.
Conversation — pacing has a hard ceiling, not just a suggestion
Conversational freedom (1-10) sets the max number of turns the agent can go without steering back to business — once that ceiling is hit, the agent is issued a hard rule, not a soft nudge, to redirect. Urgency shrinks that ceiling further: at max urgency, the agent gets one off-topic turn regardless of the freedom setting.
Response delay (min/max) makes the agent wait a randomized time in that range before replying, so it doesn't feel instant — set the same value in both fields for a fixed delay.
Human intervention pause (default 30 min, adjustable 1 minute to 24 hours) is the mechanism behind "a manual reply pauses the AI": if you or a teammate write to a lead directly — from the inbox composer or straight from the IG/Messenger app — the agent stays silent for this exact window, then resumes automatically on its own, no action needed. Turning AI off manually from Inbox instead pauses it indefinitely until someone turns it back on.
Experiments and routing (Experimentos y enrutamiento)
Needs at least one saved preset. Two independent tools, both apply only to newly created leads, never to existing ones:
- Source rules — up to 20 rules, evaluated top to bottom, first match wins. Each rule can match entry type (ad, story reply/mention, post, link, direct DM) and/or a substring in the campaign name. A rule with no conditions never matches anything.
- A/B test — sends a random percentage (5-90%) of new leads to a second preset; it's a fresh coin flip per lead, not a sticky assignment tied to the lead's id. Only runs when no source rule matched first.
Goals
Main goal — meeting, payment link, qualification-only, or a free-text custom goal — changes what the agent actually pushes toward. If the goal is "book a meeting" but no booking method is configured in Meetings, the agent is told to promise a teammate will follow up instead of trying to schedule live.
Memory works two ways: a raw message-history window the agent reads each turn (10-100 messages, default 50, regardless of what you set outside that range), and a separate structured long-term memory (facts + a running summary) that persists per conversation and gets reinjected every turn.
Forbidden topics and pricing rules are prompt instructions, not a backend filter — there's no separate step that blocks a reply after it's generated; the model is instructed every turn to deflect or redirect, same as everything else in Personality and Conversation.
Use the simulator for realistic tests and History to restore an earlier version.