read

conversation design

systems that say what they will do, then wait

I design what a machine says, and when it stops talking to ask. Two of them are running now: an AI documentation agent that drafts on a pull request and waits for a developer's approval, and the guide on this site, which answers from a script and hands you a human when it should.

discipline
Conversation design, AI agents
span
Developer tooling & portfolio
surface
GitHub, async chat · web
both
Live, and open to test
the pattern

act, confirm, hand off

Both systems run the same three steps, and the difference between them is the stake. Tap one to see what each step does.

The match is clear, so the system states its plan outright and shows its source. On a pull request that means a full draft, posted and waiting. In a chat it means the answer, with the page it came from one tap away.

the cases

one I joined, one I built alone

On Askmanu I own the conversation layer of an engineer's system: the style guide, the tone, the QA gates. The guide on this site is mine end to end, script and tests included. Open either.

op-ed · september 2026 How we look for knowledge The research behind this page: what people used to judge a source on before they read it, what a generated reply stops carrying, and why the decision about what a system may say has to be made before anyone asks. read the piece