Chat and context

Chat is where you define the outcome and the evidence Marketer should use. A strong first turn gives the Marketing Agent enough information to produce something you can verify—and makes clear whether it may only analyze, may create a draft, or may prepare an action for approval.

Marketer Home with the Ask anything or at-sign to add context composer, agent selector, send control, workspace metrics, and active work — Start in the Home composer, then use @ or the plus control to attach the records that define the request.

Start in the right workspace

Open Home.
Confirm the workspace name in the upper-left selector.
Check the date range above Overview if you plan to refer to a metric visible on Home.
Select the composer labeled Ask anything or @ to add context.

The workspace controls which connected accounts, campaigns, products, briefs, creatives, Knowledge, Skills, and prior conversations the agent can use. Switching workspaces changes that scope.

Attach the exact record

Type @ in the composer or open the plus menu and select Add context. The picker groups current workspace records under:

All
Campaigns
Briefs
Creatives
Products

Choose the record that should define the task. It appears as an attachment in the composer before you send.

Choose the attachment path

Starting point	Use	What gets attached
You know the record	`@` or Add context	A campaign, brief, creative, or product selected from the workspace picker
You are already viewing it	Pick from page when offered	The exact tagged item on the current product page
The evidence is a document or image	Plus menu → Add files or photos	The selected file, alongside any workspace records
You want material from the asset library	Plus menu → Creative library	A creative-library reference for the request

Attach the smallest set that uniquely establishes scope. The correct campaign plus a clear period is stronger context than every recent campaign with no decision attached.

The attachment identifies the record; it does not supply every analytical choice. State the provider, account, date range, attribution view, currency, or comparison you need in the request.

Build the first turn

Write the request in this order so each part can be checked later:

Outcome: what decision or artifact should exist at the end.
Scope: the attached record, provider, account, channel, or audience.
Evidence: the sources and metrics the answer must use.
Window: exact dates or complete-period logic.
Constraints: facts, guardrails, exclusions, and permitted actions.
Completion criteria: what a reviewer must be able to see.

For example:

Using the attached prospecting campaign, compare the last 7 complete days
with the previous 7 in Meta Ads. Show spend, purchases, CPA, and ROAS, and
identify the largest material change. Name the account, dates, timezone, and
attribution view. Separate measured facts from hypotheses and recommend one
test. Analyze only; do not edit or launch anything.

This request tells Marketer what to calculate, which record and source to use, how to frame time, how to distinguish evidence from interpretation, and where to stop.

Choose Auto or a specialist

The agent control in the composer starts in Auto, described in the interface as Routes each request. Leave it in Auto when you want the Marketing Agent to choose the appropriate perspective for each turn. If the menu shows a specialist whose perspective you want to keep, select that specialist before sending. The specialist remains visible in the composer for that conversation. Return to Auto when you want the Marketing Agent to route the next request again. Choosing a specialist changes the perspective applied to the conversation. It does not by itself start a longer task or grant permission to change a connected account.

Read the first response state

After you send, the conversation can take one of four paths.

Direct answer

The agent answers in the thread without creating longer work. Check the source, dates, scope, and assumptions in the answer. Ask a follow-up in the same thread when it depends on the same attachments and decisions.

Clarifying question

Marketer asks for a missing source, record, constraint, or decision. Answer in the current thread so the original context stays attached. If the missing input changes the objective substantially, start a new conversation instead.

Proposed agent or pipeline run

Marketer shows the work it wants to perform before starting the longer run.

Marketer chat showing a proposed Creative Pipeline with seven steps and buttons to Dismiss, Edit, or Approve and Run — The proposal is a review point: the steps and controls are visible before the run begins.

Review the proposal against your original request:

Are the required context-gathering steps present?
Does the plan produce the requested artifact?
Does it add a write or generation step you did not request?
Are important review points or constraints missing?

Then select Dismiss, Edit, or Approve & Run. Edit returns the plan for revision; it is the correct branch when the goal is right but the execution plan is not.

Waiting for input or approval

A run that already started can pause in Waiting. Open its detail to see whether it needs a conversational answer or a recorded decision. Answer questions in the thread; handle approval items through Awaiting action when that is where the interface routes them.

Continue without losing the boundary

Keep the current conversation when the next turn relies on the same record, definitions, or decision. Use follow-ups that name the delta from the prior result:

Keep the same campaign, dates, source, and read-only boundary. Recalculate
the CPA change, show the previous and current numerator and denominator,
and flag whether either period is partial.

Start a new conversation when you change any of these materially:

The objective or deliverable
The workspace area or provider account
The authority boundary
The source-of-truth record
The intended reviewer or handoff

This prevents old attachments or constraints from silently shaping unrelated work.

Review the result in context

When the response or task completes:

Compare the output with the completion criteria in the first turn.
Confirm that attached records still represent the intended scope.
Identify the source, account, date range, timezone, currency, attribution view, and freshness behind each decision-driving figure.
Separate observed facts, interpretations, generated artifacts, and proposed actions.
Follow output links to the owning surface instead of relying on a chat summary alone.
If an action is proposed, preserve it as a separate review decision.

Marketer conversation with a completed launch plan, three finished task steps, a final result, suggested follow-ups, and the composer — A completed task returns its final step state and focused continuation options to the conversation.

Recover from the wrong path

The wrong record was attached

Do not continue from an answer grounded in the wrong object. Start a clean turn, attach the intended record, and ask Marketer to name it and its provider before analyzing.

The answer is general despite an attachment

Restate the decision, date window, source, metrics, and expected output. An attachment narrows identity; it does not define the full measurement frame.

The proposal contains too much work

Select Edit and remove unnecessary steps, reduce the output count, or separate research from generation. Approve only the plan you intend to run.

The task became long-running

Open its task card or Activity. The initial response may acknowledge the request before the full output exists.

An old thread keeps influencing the request

Start a new conversation and attach only the current records. State the new authority boundary and completion criteria in full.

Next step

Follow the task

Track queued, working, waiting, completed, partial, failed, and canceled work.

Inspect specialist work

Open Agents to see who owns each part of coordinated work.

Review a proposed action

Confirm the exact scope and effect before recording a decision.

Reopen a conversation

Open Chats and tasks → Chats when the next turn depends on earlier context.

Get Started

Essential Guidelines

Working with Marketer

Product Guides

Integrations

Workspace and Admin

Templates and Use Cases

Release Notes

Start in the right workspace

Attach the exact record

Choose the attachment path

Build the first turn

Choose Auto or a specialist

Read the first response state

Direct answer

Clarifying question

Proposed agent or pipeline run

Waiting for input or approval

Continue without losing the boundary

Review the result in context

Recover from the wrong path

Next step

Follow the task

Inspect specialist work

Review a proposed action

Reopen a conversation

​Start in the right workspace

​Attach the exact record

​Choose the attachment path

​Build the first turn

​Choose Auto or a specialist

​Read the first response state

​Direct answer

​Clarifying question

​Proposed agent or pipeline run

​Waiting for input or approval

​Continue without losing the boundary

​Review the result in context

​Recover from the wrong path

​Next step

Follow the task

Inspect specialist work

Review a proposed action

Reopen a conversation

Start in the right workspace

Attach the exact record

Choose the attachment path

Build the first turn

Choose Auto or a specialist

Read the first response state

Direct answer

Clarifying question

Proposed agent or pipeline run

Waiting for input or approval

Continue without losing the boundary

Review the result in context

Recover from the wrong path

Next step