ListeningKitChannel texting
The four dials per channel — style, raw snippets, triage, autoreplies — and the exact pick the agent runs before every send.
This page describes an earlier design of ListeningKit, built on an in-browser demo, so parts of it do not match the app that runs today. For what works now, read the Guide.
Channel texting
Global voice compiles into the agent's system prompt. Each channel (facebook, x, reddit) carries its own ChannelProfile on top — how the agent texts there: a style dial, raw snippets it sends, the triage flow it pushes, and toggleable autoreplies. Everything below is what the mock simulateOutbound() actually does, and what the live agent will do with the same inputs.
The four dials
| Dial | Holds | Effect on a send |
|---|---|---|
| Style | casual or standard | casual lowercases the whole picked line before it goes out — like a human typing fast. standard sends it as written. |
| Raw snippets | examples: string[] — chat fragments, no sign-offs | The pick pool. The snippet with the most word overlap against the post the agent detected is sent verbatim as first touch. |
| Triage flow | triage: string[] — ordered asks | The first item is appended to every send. The full list is the order the agent pushes to qualify the deal. |
| Autoreplies | autoreplies: { trigger, reply, enabled }[] | When a post's words match a trigger harder than any snippet, that line sends instead — it's the one you approved for sending. |
Why snippets are fragments, not sentences
Snippets are sent verbatim as the first touch, and the first triage ask rides behind them. A snippet ending in "Best regards, The Team" would ship as-is and read corporate on a Marketplace thread — which is exactly the thing the dial exists to avoid. Practical rule:
- Write it as you'd text it. Fragments, lowercase, no greeting, no sign-off.
anyone around know someone with a van— notHello, do you by chance know a van operator…. - One topic per snippet. The picker matches on word overlap — the post's own words (
van,shed,salthill) have to appear in the snippet you want to trigger it. - Put your safest general opener first. When nothing scores above zero, the first snippet is the fallback send, so slot #1 is your default line.
- Don't end snippets with a question. The triage nudge gets appended as the question.
- Keep triage item #1 to one short ask. It rides along on every send on that channel.
The pick, exactly
Deterministic — the same post always previews (and sends) the same message:
- Score every enabled autoreply by word overlap (words of 4+ chars) against the detected post. The best wins; ties keep the autoreply, because it's the line you approved.
- Score every raw snippet the same way; a snippet that scores strictly higher replaces the current pick.
- Nothing scored above zero → first snippet, else the house fallback line for the channel's style.
- Append triage item #1 (question-marked if the base doesn't already end in one).
casualapplies last, lowercasing the joined line.
These same strings also reach the live agent verbatim — snippets as Example replies to mimic:, enabled autoreplies as Enabled base replies — adapt to their actual words and send:. What the simulator previews is byte-for-byte what the agent is told.
Live simulator labels
The Brand tab's simulator types the detected post and labels the result with one of three lines — they map to the pick above:
| Label under the blue bubble | The pick |
|---|---|
Matched autoreply · <trigger> | An enabled line won on word overlap; the trigger names why. |
Matched a gold example | A snippet won the scoring. |
No close match — style fallback | Nothing scored — first snippet or house line sent, so read it as "this post won't match anything you've written yet." |
If the preview reads wrong, the snippets are wrong — fix the snippets, not the drafts.