How it works
The mechanism, plainly
Aortic runs a loop between two Claude instances. Each has its own system prompt. The reply from one side becomes the user message handed to the other, so the conversation advances on its own — no human in the loop unless you choose to step in.
Turn by turn
Side A opens
Your seed topic is handed to Side A as a user message under its persona prompt. Its reply is stored as turn one.
Side B responds
A's reply is passed to Side B — clearly, this is B hearing what A said — under B's own persona prompt. Its reply is turn two.
Back to A, with memory
From turn three on, each side sees the full running transcript reframed from its own point of view: its past turns as its own words, the other side's as messages received.
How each side remembers
Every turn, a side receives the whole conversation so far — but rewritten from its perspective. Its own earlier turns come through as assistant messages; the other side's turns come through as user messages. That is why the two never collapse into one voice: each genuinely argues from its own seat.
How turns are metered
Before a turn generates, its estimated cost is reserved from your wallet. After the turn completes, the unused reservation is refunded and the exact token cost is recorded. A turn never runs half-paid, and if your balance can't cover the next turn, the dialogue pauses cleanly and asks you to top up.
When it stops
You choose the stop condition when you create a dialogue: a fixed number of turns, a token budget, judged consensus (a lightweight neutral check after each exchange), or manual — you press stop. You can also pause any time to inject a message as either side, then resume.