Overview
Agent chat answers questions from your own workspace data and can run work for you in the conversation. An explicit command can also publish the selected app's listing when your role and the app's provider prerequisites allow it. Chat is available after you complete Agent setup.
It is bound by the same scopes as the agent. Drafting output lands in the same review inbox; provider actions report their actual outcome in the conversation and workbench.
What it can see
Chat reads your workspace's own data: keyword research, competitors, rankings, listing fields and agent proposals for the app you have open. It is not a general assistant with your data attached, and it cannot see another workspace.
It works in three modes.
- Ask answers a focused question from current Spek data. Read-only.
- Plan proposes a sequence of work. A plan is not execution.
- Action runs work you asked for, inside the same scopes the agent already has.
You do not have to switch modes before giving a clear command. A direct instruction runs as Action for that message without changing your selected mode. Questions, “do not” instructions, conditions, quoted text and requests to make a plan do not run an action. If the action is clear but its exact target or paid scope is missing, Agent asks one short factual question instead of asking whether you are sure.
Where its answers come from
Answers come from a real read of your data, not from a summary the model remembered. Chat reads before it answers, and it says which storefront and locale a number belongs to.
When the data needed to answer is not there yet, it says so rather than estimating. An invented keyword or a plausible-looking rank is worse than an admission, because you cannot tell it apart from a real one.
Volume, rank and difficulty are Spek research or Spek estimates. Chat does not name outside ASO tools, even to say it is not using them.
Replies arrive as one flowing answer. Tool activity, plans and run records stay out of the conversation entirely: no status cards, no operation count, no Details link. When background work finishes, its verified result replaces the earlier queue acknowledgement in that same Agent message.
The Agent mark animates while your turn is being submitted, streamed or run in the background, then stops when the result is ready or needs your attention. Spek keeps it static when your device requests reduced motion.
Commands
Type / in the composer for the command list. Commands cover the things you would otherwise click through to, such as starting a pass or reading the current agent setup.
Anything a command does is still bound by your configured scopes, your role, and the app's current App Store Connect prerequisites. Drafting work still lands in the review inbox. A direct publish command is different: it sends the target drafts and submits the eligible version to App Review. See what the agent will never do.
A clear command is already your authorization for ordinary writes, paid work, provider actions and publishing, so Spek does not ask the same question again. Delete, remove, withdraw and irreversible stop actions show exactly one confirmation with the full scope. Declining that confirmation ends the waiting action immediately; retrying the same decline cannot send anything twice.
Common questions
I do not have it. Is something broken?
Complete Agent setup and let its first analysis finish. Everything else in this section works as documented while that setup is in progress.
Can chat publish something?
Yes, when your role and the app's App Store Connect prerequisites allow it and you give an explicit interactive command. You can name one locale, name an exact set of locales, or publish every pending locale for the selected app. A multi-locale command remains one publish task and one App Store submission. The recurring agent still never publishes on its own.
What is the difference between Retry and Reconcile?
Retry reruns the same message only after a safe transient or read failure. Reconcile checks what the store accepted after an uncertain listing publish; it does not blindly send the publish again. An uncertain operation without a safe automatic checker instead asks you to verify the result in the store yourself. Successful writes show neither action.