Skip to main content
Every sequence is created with an agent that writes its messages. Each email, LinkedIn note and reply comes from the step’s prompt, the contact’s row, your business context and the thread so far. The AI writing tab, after Builder, shows what the agent writes beside how it writes.
The AI writing tab writing for Satya Nadella at Microsoft: the Opener email the agent wrote, marked Waits for approval, with a What should change box and Looks good and Needs work buttons under it. On the right, How it writes lists Goal, call to action and replies, What it writes from, Personality, Advanced: effort and research, and Feedback and learning, all shut.

The AI writing tab, with one contact's thread beside How it writes.

​
Open the AI writing tab

Click AI writing on a sequence, or Check the writing on its Builder, which also writes the first contact. How it writes, on the right, starts with every section shut:
  • Goal, call to action and replies sets what the whole sequence is for, the link it points at, the language, and the prompt behind its replies.
  • What it writes from holds Based on, which opens your business context beside the sequence, read only. Edit in Settings opens it in a new tab, so nothing you have not applied is lost.
  • Write from, in the same section, lists the research and AI columns the agent uses as angles, a different one in each email. A contact with one filled in skips web research. Untick a column to leave it out.
  • Personality picks the voice: Direct, Consultative, Bold, or Custom in your own words.
  • Advanced: effort and research picks how much thinking each message gets, and whether the first message may read the web, the contact’s LinkedIn profile, or both.
  • Feedback and learning turns saved feedback into new instructions through Teach the agent. Feedback from the inbox collects here.
  • The Agent actions menu, beside How it writes, holds Version history, Copy setup from another sequence and Teach the agent.

​
Tailor a campaign before you create the sequence

  • Every sequence runs one campaign from a library of tested campaigns. Tailoring picks one for your table and rewrites each step’s prompt for your company
  • On Sequences, under Start with a campaign, click Best fit to let tailoring choose, or click a campaign
  • See all campaigns opens the whole library, with every step each campaign writes
  • Choose the table and Who sends it, then click Tailor campaign
  • In about a minute it shows the campaign, why it fits, the Goal and each step’s prompt. Change the goal if you want, then click Create sequence

​
Change the campaign later

  • The campaign’s name sits on the Builder canvas, beside Fit. Click it to see why it fits. On a draft, click Change to pick another, or Tailor again in its menu
  • Tailoring again rewrites every step’s prompt, so your edits to them are replaced. A launched sequence keeps its steps and can’t be tailored
  • A draft started without tailoring stays on the standard campaign. Click its name, then Tailor to the audience
  • A sequence whose steps you wrote yourself reads Your own steps there instead
  • While it tailors, the steps are read-only and a launch waits
  • Once it finishes, Thought for under the result opens each step it took: what it read, the campaign it chose and the prompts it rewrote

​
Write a prompt for a step

A prompt says what the message is for. Leave the wording to the agent.
  • The cadence starts with three prompts written for you: a specific opener, a new angle, and a short break-up. Edit those rather than starting from nothing
  • On an email that opens a thread, Subject has two choices. Written by the agent writes a fresh subject for every contact. Written by you is one every contact gets, and the agent writes a body that goes out under it
  • A step that replies into an earlier email’s thread takes that thread’s subject. See Email steps

​
What the agent writes, and when

When a contact’s first message falls due, the agent drafts the rest of that contact’s planned messages in the same pass. Each draft reads the ones before it, so a follow-up builds on the opener instead of repeating it.
  • A step that only runs on a condition is left for its own turn
  • Editing or regenerating a message drops the drafts written after it, so the follow-ups answer your version
  • A draft is charged when its own step falls due, not when it is written

​
Check the writing for one contact

The AI writing tab writes one contact’s whole sequence, in order, and sends nothing.
  • With no audience yet, it asks for one first. Opened from the tab itself, nothing is written until you click Write the emails
  • We pick the contact: the row with the most on it. Change who it writes for in Writing for, then click Write the emails
  • Only the agent’s messages are here, each named by its step and marked Written by the agent. Steps you typed stay on the Builder tab
  • Contact details, Run activity and What it learned sit shut under the buttons: the contact’s row, what the agent did while writing, and the lessons it kept
  • Every contact you check stays on the tab, up to ten. Pick one again in Writing for and their thread is still there, not written again
  • Emails Chat wrote after building the sequence are here too. A contact’s link in Chat opens this tab on those same emails
  • A warning above the thread says when their LinkedIn profile names a different company from their row, which usually means they changed jobs
  • Where a message offers the sender’s booking link and that sender has none, a No booking link chip stands in its place
  • A message that waits for your approval in Tasks says so under it, with why: what the agent was unsure of, or the setting that holds it, with Change to open that setting
  • Edit in Builder on each message opens its step’s Preview for the same contact
  • A message given Looks good on its own, on a step’s Preview in the Builder, carries a Looks good mark until the thread gets its verdict
  • Read the thread, then click Looks good or Needs work in the bar pinned under it, or press G or N
  • Looks good makes that thread what sends to the contact. It is written again only if their row or the writing settings change, and a lesson you keep does not count
  • The bar then asks whether to keep training. Next contact shows the next contact in Writing for that nobody has judged yet, and writes them if needed. Stop here takes you back to the Builder tab and writes nothing more. When nobody is left, it says so

​
Say what should change

  • Click Needs work, type it in What should change? and click Rewrite with this. The agent works out the change to its instructions, then writes the contact’s whole thread again with it
  • Needs work on a step’s Preview in the Builder opens this tab on that contact, with the note box open and starting with the step’s number
  • One note covers every message. To aim a line at one message, start it with the step number, like “Step 3: shorter”
  • While it rewrites, the messages shimmer in place
  • The rewrite shows the new words first, marked where they changed. Show original opens the words they replace
  • Looks good keeps the rewrite. The agent takes the lesson as its next version, and the rewrite is what sends to that contact. The bar then asks whether to keep training
  • Needs work on the rewrite refines it with another note
  • Discard rewrite throws the rewrite away and keeps the thread as it was
  • Nothing about the agent changes until you keep a rewrite. On a running sequence, Looks good first says which steps change and who gets the new writing. Use the new writing applies it, and Not yet leaves the rewrite waiting
  • A note about one contact’s own facts gets the agent’s answer that its instructions need no change, and nothing changes

​
Research

Two kinds of research can run before the first message to a contact. A new sequence starts with both on.
  • With Web research, the agent reads the web about a person and their company. It skips this for a contact whose Write from column is filled in, because that research is already done
  • With Person research, it reads the contact’s LinkedIn profile: their headline, summary and work history. This still runs when web research is skipped
  • What either one finds is kept for the rest of that conversation, so a contact is researched once and never again
  • One line under the switches prices the first message to a contact and each one after it. Switch either off to spend less

​
What sends, and what waits for you

The sender’s send mode decides whether a message goes on its own. On Send when confident, the agent sends what it is confident in and hands you the rest in Tasks. When LinkedIn names a different company from the row, the message waits there too, with a warning naming both. A message on the AI writing tab that will wait for you says so, and why. Some things always wait for a person, whatever the mode:
  • A placeholder that never filled in, or an empty message
  • A claim the agent could not ground in the row, the research or your context
  • Too little known about the contact to write something honest
  • A message that failed one of the safety checks
  • Banned jargon, such as “system of record”, or a sentence copied word for word from your business context, still there after one rewrite
  • Stray formatting a reader would see as it is, such as half a link or a note to the reviewer, still there after one rewrite

​
Teach it

The agent learns from what needs to change:
  • A rewrite on an approval card or an inbox draft becomes an example of what you wanted instead
  • Needs work on an approval card, with a comment, says why
  • A note on the AI writing tab teaches it once you keep the rewrite it wrote. What it learned, under the buttons, lists each lesson kept
  • Looks good on a first draft teaches nothing, here or anywhere else. It marks what sends to that contact, and it never changes how confident the agent is

​
Good to know

Type / in a step’s prompt, or in the goal, and pick a column of the audience table. The agent reads that column’s value for each contact.
Each step the agent writes carries a sparkle on its card. The AI writing card on the Builder names them, with how many contacts you have checked. To write one yourself, open it and click Written by you.When every step is typed, the card and the AI writing tab offer Let the agent write the first email. It hands that step over and opens it, and Undo in the message that follows takes it back.
Open a step the agent writes. It opens on Preview, on the contact in Writing for: their message written and waiting, sent, or the agent’s draft. Edit, then Edit for Satya only, pins your words for that one contact, and Let the agent write it again hands the contact back. Edit for everyone opens the step’s prompt.
On a draft, click the campaign’s name on the Builder canvas, then Use standard instructions in its menu. The steps go back to the standard campaign with its own prompts, and nothing tailors the sequence again until you ask.
A step’s default is the library’s own prompt for that step, before tailoring rewrote it for your company. Start from the default puts it back on that one step.
No. Tailoring is free, up to a daily number per workspace. When the day’s tailorings are used up, the card says so and the sequence keeps its standard campaign.
Tailoring rewrites the prompts from your Business context, so it needs one written first. Once it is, click the campaign’s name on the Builder canvas, then Tailor to the audience.It also stops when the sequence changes while it runs, such as a step’s prompt or the goal, so it never writes over your change. The card says so, and on a draft you can start it over from there.
Your company’s own wording, such as spelling and what you call your product, is Voice in Business context. How bold or warm the emails sound is Personality. Each email is also written as the person who sends it, so a founder and a sales rep on one sequence each sound like themselves.
Person research needs the audience’s LinkedIn profile URL column, set on the Audience tab. See Choose the audience. A contact whose row has no profile URL, or whose profile cannot be read, still gets a message. A read that never runs is never charged, so the price line quotes what you will actually pay.
You do not have to. The AI writing tab writes the contact’s whole thread, so each follow-up is read under the message it answers. A follow-up only makes sense against what came before it, so the message above it is the other half of the judgement.
Only if you say so. Feedback on an approval card names the message you gave it on, so it changes that step’s instructions alone. Tick Teach all N steps, not just this one to teach every step the same lesson.A note on the AI writing tab is different. The agent decides which instructions it changes, and you see the whole rewritten thread before you keep it.
At most ten contacts’ messages: written and waiting to send, waiting in Tasks, or being written. When more than ten of the sequence’s mailboxes can still send today, it keeps one per mailbox instead.Emails and LinkedIn steps are counted apart, ten each. The other contacts read Queued to write until a message ahead of them is sent or rejected in Tasks.While its messages wait in Tasks, the sequence says how many, such as “Waiting for approval: 10”. The Tasks count in the sidebar includes them too. Nothing else notifies you.
A line above the thread says which failed. Click Try again on it to write those again and leave the rest. A thread you are reading picks up where you left off after a refresh.When the whole write fails, Run activity names what was attempted, “Writing the messages”, with “Could not write this contact’s messages.” under it.
Yes. It reads the web and their LinkedIn profile the way the agent is set up to, so what you read is what would send.
A training run saves its messages, context, and reviews. Older agent versions are marked; failed generation keeps the previous sample.
Choose one of the five Goal presets, or choose Custom to enter your own goal. A step’s prompt opens when you click the step on the canvas; the reply prompt opens through Show prompt. Step edits autosave on drafts. On a running sequence they wait for Review changes in the header, and Check the writing opens that review first.The review lists each change beside the current version. An edited message reads as the contact sees it, with each variable as a named pill, such as Greeting, and the change marked in place.Tailoring picks one of the presets for you. It never replaces a goal you chose yourself.
The agent writes them again. Change a step’s prompt, or how the agent writes, on a running sequence, and every message it wrote that nobody has sent, edited or approved is written again, as it would be for a new contact.Review changes counts them and shows the most it costs before you make the changes live. A message waiting in Tasks comes back with its new words. One you edited or approved stays as it is, and so do the follow-ups approved with an email’s thread. So does anything already sent. A lesson from a note writes nothing again.The agent writes a contact’s follow-ups ahead, with their first message. If their row changes before a follow-up is due, that follow-up is written again from what the row says then.
No. Looks good on the AI writing tab decides what the agent sends that contact later. A message already written for them, or waiting in Tasks or the Inbox, stays as it is, and you approve it there.Nor does it make a message the agent was unsure of send on its own. Its confidence changes only with its instructions and the contact’s data.
No. Threads marked Looks good keep the words you approved after the agent takes a lesson from a note. Only a change to their row or to the writing settings writes them again.
Another contact’s rewrite is waiting. The agent answers one note at a time, so keep or discard that rewrite first.
Yes. Each contact written counts against the workspace’s free previews for the day. Once you judge one contact, the contact after the one on screen is written ahead, so the next press does not wait. Only one is written ahead.
No. Each sequence has its own, so tuning one never changes another. Copy setup from another sequence, in the Agent actions menu, copies a setup across as a starting point.
Yes. Every change is a version. Version history, in the Agent actions menu, shows them and puts an earlier one back.
No. The body ends with the ask, and your signature signs it. When the ask is a meeting, it offers the link under Goal, call to action and replies, or else the sender’s booking link. With neither, it asks for a reply.
Live messages read bounded row context. Training retains complete row values and available research runs, including earlier results. Large context is summarized for generation, while its retained originals remain inspectable.
It is told not to, and a second check reads every message against what it was allowed to use. Anything it cannot ground waits for a person instead of sending.
It stays. Pinned words are yours, and the agent does not write over them until you hand the contact back with Let the agent write it again.

​
Next steps

Email steps

Switch a step to the agent and write its prompt.

Senders

Decide what sends without you.

Tasks

Approve, edit or reject what waits for a person.

Business context

Tell the agent about your company, in one place for every sequence.