Step 8 — The Channels

Step 8 · Conversation

The interface is the whole experience. A capable agent reached through an awkward interface is worse than a modest one reached naturally.

My design goal: the household should talk to it the way they talk to a person, in an app they already have open.

Do this

Pick the app your household already lives in for family conversation. Not a new app. Adoption is the whole battle.
Pick a second channel for operational output — summaries, alerts, links, system messages.
Keep one thread per person, and one per group. Not a new conversation per request.
Address it by name in group chats, and make it answer only when addressed.
Text first. Voice is a garnish, not the interface.

Why two channels

Need	What matters	My choice
Family conversation	Everyone already uses it	A messaging app the household lives in
Operational channel	Throughput, links, notifications, logs	A second app, used only for this

The reason is simple: the family channel should stay calm. Summaries, alerts and system chatter belong somewhere a person looks deliberately, not in the middle of a family conversation.

One conversation, one thread

This is the decision that matters most and is least obvious.

The agent keeps one continuing conversation per person, and one per group — not a new chat per request. It remembers the thread of what has been discussed, the way a colleague does.

  One person  ──>  one thread, always       it remembers what was just discussed
  A group     ──>  one thread, always       it can see what the household said
  Me          ──>  one operational thread   summaries, alerts, drafts

What that buys: “move it to Thursday” makes sense to the agent, because it knows what it refers to. Without continuity, every message has to be a complete instruction, and you start writing like a programmer.

Plain language, not commands

Nobody should learn syntax to ask for something.

What I say	What I should not have to say
“Can you move Thursday’s appointment to the same time next week?”	`/calendar move --event 4712 --weeks 1`
“What have I got on this weekend?”	`list events --start fri --end sun`
“That parcel — anything?”	`track parcel --id 88213 --status`

Absorbing ambiguity is most of what an agent is. When a request is genuinely ambiguous, it asks a short question rather than picking one. Guessing is the failure mode that erodes trust fastest.

The rules for group chats

A household group chat is the most useful and most dangerous surface.

It answers only when addressed by name. Otherwise it observes. This is not something I compromise on: an agent that replies to everything becomes the reason people stop using the group.

It routes rather than broadcasts. Something said to one person is not automatically repeated to the group.

It does not send the same thing twice across surfaces. If it exists in the family thread, it does not need to appear in mine.

Matching the surface to the content

Content	Where it goes	Style
Household logistics, family things	The family channel	Warm, brief, plain text
Operational output, summaries, alerts	My channel	Denser, links, tables where useful
Anything sensitive	Never a group. One person only	Direct, minimal
Credentials or secrets	Never in a message log at all	A different mechanism entirely

One rule about formatting: anything the household reads casually is written as plain text, not formatted markup. Nobody wants asterisks around their reminder.

Voice, briefly

Voice works and it is a pleasure, and it is not the primary interface.

I use it sparingly — a spoken note alongside a screen, or a short exchange when typing is inconvenient. Everything spoken is harder to review, harder to search and impossible to skim, so text stays the durable form.

Checkpoint

Every household member can reach it, and it can reach them
In a group it waits to be addressed — tested
It remembers what was said earlier in the same thread
It asks when a request is ambiguous rather than guessing
Nothing sensitive has ever appeared in a group thread

Step 9 — Memory and the wiki.