4 · Learn from every chat
What every conversation compounds into — verbatim signals and the personas behind them, on a board your coding agent can read too. The layer that runs whether or not you build anything.
The other four layers do something for a user in the moment. This one does something for you, afterwards.
Every conversation Sarah has gets read once more after it ends: what did this person actually want, what did they call it in their own words, and does it match what other people have been saying? That reading is the whole layer. It doesn't move money on its own — it decides how well the other four land. Who deserves an offer, how big it should be, which bug is worth your Tuesday.
It is also the layer with nothing to install. Every piece below runs from the moment your first conversation completes.
From one conversation to one signal
Here's an ordinary exchange at a pricing page — nothing dramatic, the user doesn't buy:
What that becomes, without you doing anything:
Note the shape. Not "user had a pricing question" — the useful part is the sentence they said, kept in their words, attached to a named competitor and a real use case. A summary would have thrown all three away.
What lands
Signals
Every finished conversation — whether the user wrapped up or walked away mid-thread — is mined for six kinds: a problem, a wish, a persona, a use case, a pricing signal, or a competitor mention.
Two rules make them trustworthy. Quotes are verbatim — a guard rejects anything the user didn't literally type, typos included, so you're never reading a paraphrase that drifted. And they're deduped — the twelfth person to complain about the same export bug strengthens one signal instead of creating a twelfth ticket.
Personas
Signals accumulate into people. Not segments you defined in advance and then went looking for — the groupings that actually showed up, described in the language they used about themselves.
The Signals board
Everything above lands on one board in your dashboard, sorted by how many distinct people said it. "This week" is the seven-day view — what users said and what Sarah did about it, side by side. Every card opens into the conversations underneath it, so a count is always one click from the sentence a real person typed.
Your own agent can read all of it
Everything above is also exposed to your coding agent over MCP, read-only. So instead of you relaying what users said into a prompt, the agent asks directly — what's the context on this project, what has anyone said about exports, show me that one conversation.
This is the opposite direction from Actions. There, UserSay calls out to a server you built so Sarah can do things. Here, your agent calls in to UserSay to read what she learned. Two MCP servers, pointing opposite ways, easy to confuse — worth keeping straight.
What zero code catches
All of it. This layer needs no trigger() calls, no MCP server, and no configuration — it runs on whatever conversations happen. If you install the script tag and never write another line, you still get signals and personas.
The only thing that changes its output is how many conversations Sarah gets, which is what the other four layers are for. A quiet install produces a thin board. That's the honest dependency: this layer is free, but it isn't automatic — it eats what the other four feed it.
Next: The user journey for the map, or Actions if you're ready to give her hands.
3 · Grow the account
The only upward money — conversion of the users on the fence, expansion for the ones who already outgrew their plan. Two of its four states cost you no code at all.
Install
From nothing installed to fully working — the script tag, identify, triggers, React, your knowledge base, and exactly what happens to your users' data.