Talking to the AI
Nadia
@AI turn this thread into tickets, one per thing that has to happen.
👀 1
WeMachinesAI
Filed T-135, T-136 and T-137 - one per decision. @Priya owns the first two.
Most advice about prompting is written for a chatbot that knows nothing about you. This is the opposite problem. An AI teammate here has already read the channel, the board, the docs and - if you connected a repo - the code. So the useful skill is not describing context. It is knowing what to hand over, and how much.
Mention it where the work is
@-mention it in a channel, in a task comment, in a comment on a page. Where you ask matters more than how you phrase it: a mention in a task comment is already about that task, and a comment thread anchored to a passage of a document is already about that sentence. “@AI is this still true?” on a highlighted paragraph is a complete question.
A 👀 from it means it has seen you and will get to it. Tag it again while it is still working and it answers both messages together rather than one at a time.
Hand over jobs, not questions
The thing people underuse is scale. A mention can carry a real piece of work, not just a query:
@AI go through every task in this topic and re-prioritise them against the launch date.
@AI triage the backlog: anything that is really a duplicate, say so and name the pair.
@AI read the last two weeks of #support and write a page grouping what people complain about.
It pages through the whole board and applies changes in batches rather than doing a handful and stopping. If a job is too big to finish in one go, it says where it stopped - and mentioning it again picks up the rest. It never reports a half-done sweep as finished, which is the failure mode that would make the whole idea useless.
Ask in chat or in a comment for anything like this. Voice is for short things, because somebody is standing there waiting.
It can ask you back
When what you asked could reasonably mean two things, you get a small card with the question and two to four options to click. Picking one carries on where it left off, and posts nothing in the room in your name - the card itself becomes the record of what was chosen, and who chose it.
Question for Nadia
Which launch date should I re-prioritise against?
It looks the answer up in the workspace first, so a question means the workspace genuinely does not settle it. And asking ends its turn rather than holding something open: a question you never answer just sits there, the way a person’s would.
Show it the screenshot
Paste a screenshot into chat, attach a photo, embed a diagram in a task or a page. It looks at the actual picture when it matters, so “what’s wrong in this screenshot?” gets an answer about the screenshot rather than about its filename. This is the fastest way to get a useful answer about a broken UI, a chart nobody can read, or a design that does not match the build.
It joins in without being asked - a bit
Every message in chat is checked by a fast classifier that decides one of three things: reply, react with a single emoji, or stay out of it. Staying out is the common case, and agents deliberately keep out of ordinary teammate-to-teammate talk. A question nobody picked up, a decision going in circles, a wrong fact worth correcting - those earn a reply.
Being addressed by name always earns one. Talking about an agent does not: “I already asked Eve about that” is not a summons. An admin can switch the whole behaviour off in Settings → Agents if a team wants agents that only ever answer mentions.
The conflict cards
Every message and comment a person writes is checked against the workspace records most related to it - tasks, decisions, pages, earlier messages - and against general knowledge. One that contradicts them gets a small “flagged a possible conflict” card underneath, naming what it disagrees with.
Leo
Launch is still the 12th, so I am booking the webinar for the 14th.
WeMachinesflagged a possible conflict
P-27 Launch plan moved the launch to 26 October on Tuesday, and T-131 was re-dated with it.
Announcing a change is an update, not a conflict, and most messages get no card at all. The point is the sentence said in passing - a date that moved, a number from an old spreadsheet, an owner who left - that everyone in the room believes because nobody remembers the ticket that says otherwise.
Talking to it out loud
Press ⌘O and the workspace’s primary agent answers, out loud, with its own name and face on screen (the voice assistant is on every plan, the trial included). It already has the context, so “what’s left before Friday?” is a complete sentence. Spoken answers stay short by design. Two things are chat-only and it will say so: the skills your team keeps in its repo, and tools from an outside MCP server.
⌘O is the only way to talk to the AI out loud, and it is private: one person and the primary agent. Agents do not join huddles. What the AI does for a call comes after it, when it reads the transcript and writes the summary on the Transcripts page.
The habit that makes the difference
Write the thing down where it belongs. Almost every complaint that starts “the AI didn’t know” ends with the fact living in somebody’s head or a DM in another tool. Decisions in pages, work in tasks, conversation in public channels - and the room answers back better every week. That is not a WeMachines quirk; it is the entire mechanism.
Next: the same teammate, tuned into several. Give each agent a job.