Every comment and DM Extrovert drafts is written with a voice. Without one filled in, drafts read generic. This article covers what to write in each field of a voice, and what to leave out.
For how voices work as a whole - the library, how many you can have, duplicating and deleting, and which campaign uses which - see How voices work.
To open the one you want to edit: Settings → Voice library, then click it. Or from a campaign, Campaign → Voice.
What affects what
Both comments and DMs: Background.
Comments only: Comment tone, Approved examples, Talking points.
DMs only: DM tone, DM templates, Selling points.
Keep each part doing its own job. Facts about you and the company go in Background. Opinions and angles for comments go in Talking points. Sales material goes in Selling points and DM templates. Rules about how the writing should sound go in the tone fields. Mixing them is the most common reason drafts feel off.
Background: about you and your company
Where: open a voice → Background. Two fields, up to 2,500 characters each.
Both are optional and both make every draft better. This is the fastest single improvement you can make to a new voice.
About you: your role, what you are known for, your background and experience, the beliefs you hold strongly, and a few personal details that make you relatable. Include anything you might naturally mention in a comment or a DM.
About your company: what it does, who it serves, the problem it solves, what makes it different, real results, and recent milestones.
Write in paragraphs separated by blank lines. Keep it factual and specific rather than filling it with marketing lines.
Comment tone
Where: open a voice → Comments → Comment tone instructions. Up to 2,000 characters.
Describe how your comments should sound: tone, preferred length, sentence structure, punctuation, emoji use, which language you write in, phrases you use, phrases you avoid, and any other signature pattern in how you write. Also say whether you want to mention the post author by name - if the draft contains their name, they are tagged automatically when you post it.
Works best as separate, clear rules in a list, one rule per line.
Writing rules that change the draft
Every line in a tone field does one of two jobs. It gives the AI a fact it needs, or it sets a rule about how to write. A line that does neither takes attention away from the lines that do.
Rules work better when they say what to do, not only what to avoid:
Instead of "don't write long comments", write "two or three sentences, one point each".
Instead of "don't be salesy", write "close with a question about their work, not an offer".
Instead of "avoid generic openers", write "open with the specific claim you are responding to".
A ban on its own leaves the AI to decide what goes there instead, and that decision often brings back the thing you were trying to remove.
Say each rule once. Repeating it in stronger words does not make it stronger. If a rule is still ignored after two tries, the rule is usually not the problem. Check first that the thing you are trying to fix actually comes from the draft text, because tagging, spacing and how a comment renders once posted are not controlled from here. Then replace the rule with a positive instruction, or add an approved example that shows the behaviour you want. Examples carry more weight than descriptions of them.
Keep one rule per line. A line holding three rules gets followed in part.
Be concrete where it matters. "Two or three sentences" changes the draft. "Professional but friendly" does not, and it is the kind of line that fills the field without doing anything.
Don't repeat yourself across fields. Each field is read for its own job, so the same fact in Background, tone and talking points does not get three times the weight. It only costs you room.
Why a "skip this one" rule doesn't work
The tone fields control how a comment or DM is written. They cannot control whether one is written. The assistant always returns something for the post or conversation it is given, so a rule like "skip this post" or "only message after a real exchange" has nothing to act on.
What happens instead is one of two things, and neither is what you wanted:
It writes its own reasoning out as the draft. Comment drafts that read "No comment", "(skip)" or "Skip and like", and DM drafts that read "No DM recommended yet" followed by an explanation, are this. In Manual mode you can reject them. In Auto mode comments publish without review, and a DM draft sits in the composer with Send next to it.
It follows whatever fallback your instructions give it. A rule like "if there is no neutral angle, write one short sentence about the author's work" produces a polite, harmless comment that is still attached to the post you did not want to touch.
Wording worth removing from your tone instructions:
"Skip the post", "skip it rather than comment"
"Do not comment on ..."
"Only message after ...", "only after a real exchange"
"No comment is better than a thin one"
"If there is no good angle, don't comment"
"Silence is better than a generic comment"
"If the post is in another language, do not comment at all"
One more thing instructions do not reach: the like that goes out alongside a comment. That is part of the action on the draft, not part of the text, so no wording in a tone field changes it.
To decide which posts come up in the first place, use the campaign rather than the voice. Post categories on the campaign's Commenting tab excludes whole categories such as Politics and Extra sensitive, Only my custom topics narrows the feed to subjects you define, and the prospect list decides whose posts you see at all. All three are in Commenting settings.
For DMs the equivalent controls are on the campaign's Outreach tab. Under Delay outreach, warm-up targets decide how many comments have to happen before a prospect reaches the DM stage at all. Two settings there release a prospect early even when the comments never happened: Warm-up time limit starts outreach once the wait is up, and Go straight to outreach when there are no recent posts starts it immediately for anyone who has not posted in 30 days. Turn both off if you only want to message people you have actually engaged with. The first outreach wait and the follow-up limit are on the same tab.
There is no filter for the language a post is written in. You can set which language you write in, but you cannot use a tone instruction to leave a post in another language alone. If that matters, keep those prospects out of the campaign.
Things that are not coming from your instructions
Before you add a rule about how a comment looks, check that what you are seeing is part of the text at all. Two things regularly get mistaken for the AI's writing:
A tagged name looks bolder than the rest. When a draft contains the post author's name, Extrovert tags them, and a tag is rendered as a link rather than as plain words. Nothing in a tone field changes how that renders. If you would rather not have it, say in Comment tone that you do not address the author by name.
Drafts you are looking at were written earlier. A change to a voice applies to drafts generated after you save it. Existing drafts in the feed keep the wording they were created with, so judge a change on the next batch, not the current one.
Approved examples: your writing DNA
Where: open a voice → Comments → Approved examples. Up to 20 per voice, 1,250 characters each.
These are real comments you would send. The AI mimics them, so they matter more than any other single setting.
Include:
Comments you actually wrote, long and short, so the range is visible
Ones that show a real opinion or a concrete point, not just agreement
Exclude:
One-word replies and generic praise
Near-duplicates of an example you already added
Text an AI wrote and presented as yours
If you don't have good examples yet, leave this empty rather than adding weak ones. Every future draft follows the patterns it finds here, so a bad example does real damage. You can add better ones later.
Auto-learn is a switch at the top of the same section, on by default. While it's on, every comment you edit and approve in the feed is saved here automatically, so the voice keeps improving without you doing anything. Turn it off if you only want examples you added yourself.
For comment patterns that work for different campaign goals, see Comment copy tips.
Talking points
Where: open a voice → Comments → Talking points. Up to 40 per voice, 750 characters each.
Ideas, facts, and points of view the AI can weave into a comment when they fit the post. Each one should hold a single clear thought and show real expertise. Three to five good ones already change how the drafts read.
Because they are only used when they fit, they need to match the subjects your campaign actually surfaces. If you comment on posts about hiring but every talking point is about your manufacturing process, none of them will reach a draft.
Include:
Industry observations based on your own experience
Counterintuitive findings, or lessons from something that failed
Methods or frameworks you developed
Personal anecdotes that illustrate a broader principle
Quantified results or data points from your work
Your take on an emerging trend
Practical tips from hands-on experience
Exclude:
Common knowledge everyone already agrees on
Surface-level observations with no specifics
Plain facts about your company - those belong in Background
Pitches, pricing, booking links, and outreach templates - those belong in Selling points and DM templates
Good sources: your own posts, content pillars, articles you wrote, webinar or podcast transcripts. Frame company expertise as your personal opinion rather than a product claim.
Decide your DM strategy first
The DM fields are easier to fill once you have answered four questions. Answer them before you open the form.
Who receives this. People whose post you commented on, people who saw your comment on someone else's post, people you have never interacted with, and people who replied once and went quiet are four different situations. Most voices need messages for more than one.
What the first message does. Start a conversation, or make an offer. Both work. Mixing them into one message usually does neither.
Whether you pitch, and when. If you pitch in the first message, say so in DM tone. If you never pitch before a reply, say that instead. This decision changes every draft more than any other line in the field.
What the second message does. A follow-up that repeats the first one reads as pressure. Decide whether yours adds a new reason to reply or simply resurfaces the original.
How many follow-ups go out, and how far apart, is not part of the voice. That lives on the campaign's Outreach tab. Keep the two consistent: if your DM tone says one follow-up and then stop, set the follow-up limit to match.
DM tone
Where: open a voice → DMs → DM tone instructions. Up to 2,000 characters.
How your direct messages should sound: tone, how quickly you get to the point, how direct or salesy to be, length, punctuation, emoji use.
Cover how you open a cold conversation, how you follow up when someone never replied, and how you continue an existing thread. Good DM tone instructions capture how you actually message people privately, which is usually not how you write in public. Works best as a list of separate rules.
The same limit applies here as in Comment tone: an instruction describes how a DM is written, never whether one is written. See why a "skip this one" rule doesn't work.
DM templates
Where: open a voice → DMs → DM templates. Up to 40 per voice.
Example messages the AI can pick from and adapt for a matching prospect. It decides which one fits and how far to adapt it, so each template has three fields. All three are required - a template won't save with any of them empty.
The message (up to 750 characters) - write it the way you would actually send it.
When to use (up to 250 characters) - the situation it fits. For example: first touch after they engaged with a comment.
Adaptation (up to 250 characters) - how much freedom the AI has. For example: reference their concrete post and keep the ask low-pressure.
Keep the "how it should sound" rules out of the template itself. That belongs in DM tone.
Organise templates by situation, not by order
Templates are not a sequence. The AI reads all of them and picks the one whose When to use matches the person in front of it, so the order you add them in means nothing.
That matters because most people organise their real messages by trigger: one message for someone who just raised a round, another for someone who commented back, another for a name they have never spoken to. You can keep that structure. Write the trigger into When to use and the AI will match on it.
Worth covering, one template each:
A first message to someone you have just connected with
A first message to someone you have engaged with through comments
A first message to someone with no prior interaction
A follow-up after no reply
A restarter for a conversation that went cold
Add more when a specific situation keeps coming up and the generic template does not fit it. A template that names its trigger precisely beats three that all say "first outreach".
Placeholders in a template
Write anything you want the AI to fill in as a placeholder in square brackets, in plain words: [their name], [the point they made in their post], [the result we got for a similar customer]. There's no fixed list of variables and nothing to memorise - the AI reads the brackets and replaces them.
It can only fill in what it actually knows. For each prospect it sees:
Their name and headline
Their profile description
Up to 3 of their posts from the last 14 days
Your conversation history with them
Plus everything in this voice: your background, your DM tone, your selling points.
So [their name] is always safe. Something like [their industry] or [their company size] is not a field Extrovert stores - the AI can only infer it from their headline, description, or posts, and if none of those say it, you'll get a vague sentence or a guess. Build your templates around the name, the headline, and whatever is genuinely in their recent posts.
Selling points
Where: open a voice → DMs → Selling points. Up to 40 per voice, 750 characters each.
Facts the AI can use in a DM when they're relevant. Add each one separately rather than in one block.
Include:
Your main value proposition and the specific problems you solve
Results customers typically get, with numbers where you have them
Why you rather than the alternatives
Common objections and how you answer each one
Booking links, trial links, case studies, and anything free you can offer
Recent launches, news, events worth mentioning
Three to five concrete facts already make DM drafts noticeably sharper. A vague selling point produces a vague DM.
To see how these DMs get triggered, reviewed, and sent, read Suggesting relevant DMs to prospects.
Let an AI do the first pass
You don't have to fill any of this from a blank page. Every tab has a Set up with AI card offering three routes: your own AI over MCP, the Extrovert assistant in the side panel, or Copy prompt for your AI, which puts a ready prompt on your clipboard already filled with whatever that section currently holds.
Give it real source material - your website, profile, posts, past comments, sales documents - and then cut anything generic, repeated, unsupported, or unlike you before you save it. The include and exclude lists above are the bar to hold it to.
DM templates and selling points are the one place to be strict about this. They should come from messages you have actually sent and sales material you actually use. An AI writing them from nothing invents your pricing, your results, and your links. If you don't have that material to hand, leave those two sections empty until you do.
If an AI is filling this in for you
These are the rules the assistant should follow, and they are worth knowing so you can tell when it goes wrong:
Work only from material you actually provided. It should never invent results, numbers, pricing, client names, credentials, or links. If a selling point needs a figure you did not give, it should leave that selling point out and tell you what is missing.
Ask before writing when there is nothing to write from. What you do, who you are talking to, and what a good outcome looks like are enough to start. One round of questions, not an interview.
Act when the request is already clear. "Make my comments shorter" needs no clarification. Make the change, show the resulting text, confirm it saved.
Diagnose before writing a rule. If you report a problem, the assistant should establish that it comes from the draft text before adding an instruction about it. Adding a rule for something the AI was never doing fills the field without fixing anything, and the field has a limit.
Check that talking points cover the campaign. Compare them against the topics or people the campaign actually surfaces, and say so when they do not overlap.
Push back when a request won't do what you expect. A rule telling the AI not to comment on something belongs in the campaign, not the voice.
Read the request for what it means. "Write better comments next time" is a request to change the voice. "Stop showing me these posts" is a request to change the campaign settings. They arrive in the same words and go to different places.
Show the exact text before saving it, then confirm what was stored. A change isn't done until it's been checked.
Comments in the wrong language, or mixing languages?
By default a comment follows the language of the post. Comment tone instructions override that.
If comments come out in the wrong language or mix two, open Comments → Comment tone instructions and look for a rule that sets or mixes languages. Remove or rewrite it and save.
Where to start
Fill Background first, from your LinkedIn profile and company website. It feeds both channels.
Then add Comment tone and three or four approved examples, and leave Auto-learn on. That's usually enough to stop drafts reading generic. Add talking points as you go.
When you start using DMs, add DM tone, a few templates, and the selling points you actually use.
Three is the number that matters for the collections: with one or two approved examples or DM templates the AI has too little to pattern-match against, and the Voice library keeps asking for more.
You can change any of this at any time and new drafts use the latest version. To fix one draft after it's been generated, use Improve draft.