An Icebreakly Voice is a reusable profile that tells the AI how you write comments: your tone, how long you run, what you know, a handful of comments you would really post, and the lines you will not cross. You train it once in a four-step editor, test it on real posts in a built-in playground, and attach it to any campaign. After working through this guide you will have a Voice whose drafts need little or no editing.
Where Voices live
Open Voices in the sidebar. The page lists every Voice as a card showing how many campaigns use it, with two buttons: Edit and Test Voice. Small icons on each card duplicate or delete it.
Click New Voice. A window titled Create a new Voice asks for a name and nothing else, with the note that you will train its tone, samples and guardrails next. Name it after the person and the audience, such as "Maya, CFO audience", because you may end up with more than one.
The editor opens with four numbered steps across the top (Voice & Tone, Strategy, Sample Comments, Guardrails) and a separate Test playground button that is available at any point. Steps open in order the first time through. Once the Voice is complete you can jump between them freely.
Step 1: Voice & Tone
Tone is required. Eight preset chips are offered: insightful, consultative, professional, warm, direct, candid, analytical and curious. Below them is a text field where you can describe a custom tone instead. The editor calls tone the biggest lever on how you come across, and it is right.
Pick the chip that matches how you talk to a peer, not how you wish you sounded. If you are blunt in meetings, choose direct. A custom tone is worth writing when one word undersells you. For example:
dry, practical, slightly skeptical of hype, generous with specifics
Three dropdowns follow.
- Length: a few words, one punchy line, 1-2 sentences, 2-3 sentences, or 3-4 sentences (substantive). The default is 1-2 sentences. Longer is not more impressive. If you are unsure, this guide to comment length explains when each length fits.
- Emoji: none, rarely or sparingly. The default is rarely. There is no "often", on purpose.
- Language: the default, "Auto — match the post", writes in whatever language the post is in. Or fix one language from the list.
The last field is How you write (style notes). It is optional and passed to the model as extra guidance. Use it for habits, not adjectives:
I start with the specific thing I am reacting to, never with praise. I use plain verbs. I do not use exclamation marks. If I disagree I say so in the first sentence.
Step 2: Strategy
Your perspective is required, and it is the field people rush. It tells the AI who is speaking, so comments carry real knowledge instead of general agreement. The editor is explicit that this is used for substance, never to brag or name-drop.
A weak perspective:
Sales leader with 10 years of experience.
A useful one:
I run implementation for a warehouse management software company. I have been through about forty go-lives in food distribution and 3PL. I know what breaks in week one: label printing, cycle counts, and floor staff who were never asked. I care about adoption more than features.
The second version gives the AI something to say when a lead posts about a painful rollout.
How to engage is a set of six chips. Select the moves the AI is allowed to use, and it varies them from post to post:
- Ask a sharp question
- Add a data point
- Share an experience
- Respectfully challenge
- Amplify & extend
- Fresh angle / reframe
Select at least three so comments do not all take the same shape. Leave out any move you would not make yourself. If you never cite numbers in conversation, do not select "Add a data point". Leave out "Respectfully challenge" when the audience is senior and you have no relationship yet, and add it later.
Questions controls endings: "Sometimes — when it fits", "Always end with a question" or "Never end with a question". Sometimes is the default and the right choice for most people. A question on every comment becomes a visible tic by the third time a lead sees it.
Step 3: Sample Comments
Paste comments you would really write. The editor asks for 3 to 10, requires at least one, and caps the list at 10. It also makes a promise worth knowing: the AI mimics your voice and never copies the samples.
The best source is your own LinkedIn activity. Scroll back through comments you have posted and copy the ones you are still happy with. If you have none, write five now against real posts in your feed. Aim for variety:
The week-one dip is real. We plan for pick rates to fall by a fifth and tell the client before go-live, so nobody panics when it happens.
Curious how you handled the night shift. In my experience they are trained last and blamed first.
Disagree slightly. Barcode discipline matters more than the software you pick. I have seen a great system fail on bad labels.
Those three differ in length, in move (experience, question, challenge) and in opening. That spread is what stops drafts from converging on one pattern. Avoid samples that open with a compliment unless you want every draft to open with one. For more raw material, these LinkedIn comment examples are sorted by type of post.
Step 4: Guardrails
Two columns: Never and Always. The editor defines them plainly. Never is phrases or things to avoid. Always is rules the AI must follow. Click Add rule under either column and type one rule per row.
A starting set that works for most sellers:
| Never | Always |
|---|---|
| Never open with "Great post" or any praise of the post itself | Always refer to one specific detail from the post |
| Never mention our product, pricing or a demo | Always write as a peer, not a vendor |
| Never use the words "journey", "resonates" or "synergy" | Always keep it under three sentences |
| Never tag other people or add hashtags | Always use British spelling |
Write rules as concrete behavior. "Be authentic" gives the model nothing to check. "Never use exclamation marks" does. When you finish, click Save & finish. Guardrails matter most once you stop reading every draft, which is the subject of autopilot and guardrails.
Test it in the playground
Click Test playground. If you have unsaved edits, a prompt offers Save & test, because the playground generates from the saved version of the Voice.
A toggle switches the input between Post content and Post URL. Paste the text of a post, or paste a LinkedIn post URL and Icebreakly fetches the real post the same way a live campaign does. If you have nothing to hand, three sample posts are one click away: Hiring announcement, Product milestone and Contrarian take.
Click Generate comment. The draft appears under "Your voice". Click Regenerate for another attempt at the same post. Your latest changes are saved automatically before each test, so you can adjust a field and try again without leaving the screen.
Test on at least five posts from people on your actual lead list, covering different kinds: a win, an opinion, a hiring post, a question, something personal. One good draft proves little. Five across different post types tells you whether the Voice holds up.
Tuning: read the draft, change one thing
When a draft is off, resist rewriting the whole Voice. Find the symptom and change the field that controls it.
| What you see in drafts | What to change |
|---|---|
| Agreeable but empty | Add detail to Your perspective, and select "Share an experience" |
| Every comment ends with a question | Set Questions to Sometimes or Never |
| Too long | Shorten Length, and replace long samples with shorter ones |
| Opens with praise | Add a Never rule, and remove any sample that opens that way |
| Sounds like a brochure | Change Tone to direct or candid, and add plain-speech habits to style notes |
| A word you would never use keeps appearing | Add it to Never by name |
| All drafts have the same shape | Select more engagement moves and vary your samples |
Change one field, test on the same post, compare. Two or three rounds is usually enough. If a draft still reads as generic after that, the cause is almost always a thin perspective. The reasons generic comments do harm are laid out in why generic AI comments backfire, and the wider craft is covered in AI LinkedIn comments that do not sound like AI.
One Voice or several
Most people need one. Create a second when the audience changes how you speak: a founder who comments on investors' posts and on engineers' posts, or an agency writing for three different clients. The duplicate icon on a Voice card copies everything into a new Voice that is not linked to any campaign, so you can change the tone and samples without touching the original.
In the campaign wizard, the Voice is chosen in the Choose Setup & Voice step. Alongside your trained Voices there is always a Default voice card. It exists so a campaign can run before you have trained anything. A trained Icebreakly Voice will be closer to how you write.
What to do next
Open Voices, create one, and give the perspective field ten honest minutes. Paste five real comments, add four guardrails, and run it against five posts from your lead list. Then attach it to a campaign in Approval mode and keep tuning from the drafts you edit. Every edit you make in the queue is a hint about a field to change here. You can see where Voices sit in the wider product on the features page, or start a free trial and train yours today.
Frequently asked questions
How many sample comments does an Icebreakly Voice need?
At least one to move past the Sample Comments step, and the editor recommends pasting 3 to 10. The maximum is 10. Samples that differ from each other in length and approach give better results than ten variations of the same comment.
Can I use one Voice in several campaigns?
Yes. Voices are reusable. Each campaign selects one Voice in its Setup and Voice step, and the Voices page shows how many campaigns use each one. Editing a Voice changes how future comments are written in every campaign that uses it.
What happens if I delete a Voice that a campaign is using?
The delete dialog lists any active campaigns that use the Voice and warns that deleting it will pause them. You then need to assign another Voice before reactivating those campaigns. Deletion cannot be undone, so duplicate the Voice first if you only want to experiment.
Does the Voice write in languages other than English?
Yes. The Language setting defaults to Auto, which writes each comment in the language of the post it responds to. You can instead fix the language by choosing one from the list, which covers around thirty languages including Spanish, French, German, Portuguese, Japanese and Arabic.