How to Train AI to Write in Your Voice on LinkedIn

The most common complaint about AI writing tools isn't that they write badly. It's that they write like everyone else.

Generic hooks. Corporate phrasing. Posts that could have been written by any professional in any industry. Your audience can feel it — and so can you, when you read the draft back.

The fix isn't a better prompt. It's training: giving the model your actual writing and your actual expertise, so it stops generating from an average of the internet and starts generating from you.

This guide covers both halves of that. First, how to describe your own voice, which you need whether you use ChatGPT, Claude or a dedicated tool. Then the specific mechanics of doing it inside ContentIn's AI ghostwriter — what to upload, where the upload actually lives, what happens to your files, and how to tell whether it worked.

Why Training Beats Prompting

An untrained model pulls from a vast average of everything ever written online. That average is competent, inoffensive and completely indistinct — the opposite of what makes a LinkedIn post worth reading.

Better prompts don't solve this. You can describe your voice in a paragraph and the model will approximate it for one draft, then drift back to the average on the next. What changes the output durably is material: your posts, your frameworks, your vocabulary, sitting in the context the model writes from.

The practical difference is in how much you rewrite. A trained ghostwriter produces first drafts that need editing. An untrained one produces drafts that need replacing — which defeats the point of using it.

Step 1: Describe Your Own Voice First

This step is unavoidable and most people skip it. You cannot check whether an AI sounds like you until you can say what "like you" means.

"Before you can train ChatGPT on your writing style, you need to be able to describe your writing style." — Matt Giaro, content creator, mattgiaro.com

Read your own posts like a stranger

Take three to five of your best LinkedIn posts — the ones that sound like you, not just the ones that performed. Then:

1. Copy them into one document.
2. Strip out anything that identifies you.
3. Read them as though someone else wrote them.
4. Write down what stands out.

You are looking for three things. Tone — formal, blunt, warm, dry? Sentence shape — short and clipped, or long and qualified? Vocabulary — the words you reach for, and just as importantly the ones you never use.

Most people find at least one pattern here they had never noticed.

Write the rules down

A page of notes is enough. Cover:

• Tone in one word. Authoritative, friendly, provocative — pick, don't hedge.
• Three to five traits you want the writing to carry.
• Do's and don'ts. The don'ts matter more. "Never opens with a rhetorical question" is a more useful instruction than any positive one.
• Sentence structure. Punchy or developed?
• Transitions. How you move between ideas — "here's the thing", "the problem is", a line break.
• Your tics. The construction you always reach for. These are the fingerprints.

Step 2: Choose What to Feed It

The best training material captures two different things, and they are usually in two different documents. Separating them is the single most useful idea in this guide.

For voice: your best LinkedIn posts

Put five to ten of your strongest LinkedIn posts into a single text file. These are the clearest signal of how you write for this platform — your hook style, your rhythm, how you open and close, how much you lean on stories versus data.

If you're unsure which to pick, use this test: choose the ones that got comments along the lines of "this is so you", or "I knew this was yours before I saw the name". Engagement is a weaker signal than recognition.

For expertise: your long-form work

Newsletters, talk transcripts, case studies, articles, internal frameworks — anything where you went deep. This is what stops the output being surface-level. When the system has your actual thinking on a subject, the ideas it generates are grounded in what you know rather than in generic industry commentary.

Strong candidates: a newsletter edition where you explained a framework you use, the blog post laying out your contrarian view on something in your field, a transcript of a talk where you had to defend a position.

What doesn't work

• Content you didn't write. If the voice isn't yours, the system learns the wrong patterns — confidently.
• Short, context-free snippets. Not enough signal to generalise from.
• Heavily formatted documents. Dense tables and slide decks are mostly structure; the writing is buried.
• Outdated material. If it no longer reflects how you write or what you think, it is teaching the wrong thing.

Step 3: Where the Upload Actually Lives

In ContentIn, expertise training sits inside your profile settings rather than your account settings — a distinction worth knowing, because the account settings page has nothing to do with it.

1. Open your profile menu. Click your avatar in the top-right corner and choose Profile & AI settings. On older accounts the same entry reads AI & Content Settings. Either way it takes you to the profiles page.
2. Open the AI Training tab. Then scroll to the section headed Expertise training files.
3. Add your material. Add File takes plain-text (.txt) and PDF uploads. There are also two other ways in that most people miss: paste a URL and the article behind it gets pulled in and extracted, or paste raw text straight into the Transcript tab — which is the fastest route for a talk transcript or a newsletter you only have as an email.
4. Watch the status. A new file shows as draft. Once knowledge extraction has run, it flips to extracted. Files you retire show as archived, and there is a rework action that pushes a processed file back to draft so it gets read again.

How long it takes. The text itself is pulled out of your file the moment you upload. The slower step is knowledge extraction, which runs on a schedule roughly every four hours — so a file uploaded just after a run waits a few hours, not a day. If a file sits in draft far longer than that, the usual cause is that your profile has no content pillars defined yet: extraction works pillar by pillar, so with none set there is nothing for it to extract against.

Two limits to know about. A profile holds up to 30 training documents, so treat the slots as finite and curate rather than dump. And expertise training files sit on the Pro plan — see the pricing section for what's on which plan. Voice learning from your own posts is not gated this way; it starts from the posts on your connected profile.

Step 4: What the System Does With It

Knowledge extraction

Your documents are read against your content pillars, and the frameworks, positions and specifics tied to each pillar are pulled out. This is what makes generated ideas feel like they came from your work rather than from a search result.

Voice pattern learning

From your posts specifically, the system learns structural habits — paragraph length, how you open, how you transition, how you land an ending — along with the vocabulary and recurring constructions that make writing recognisable.

Idea generation

The extracted knowledge feeds the idea generator directly. Ask for post ideas and the suggestions are drawn from what the system knows about your actual niche, rather than from broad topics adjacent to your industry.

Step 5: How to Tell Whether It Worked

Generate a few posts and read them critically. The right question is not "is this a good LinkedIn post?" — it's "would I have written this?"

Check four things specifically:

• The hook. Does it open the way you'd open? Too formal, too casual, or about right?
• The examples. Does it reach for the kinds of things you'd reach for, or for industry clichés?
• The structure. Is the rhythm yours — short and punchy, or longer and developed?
• The vocabulary. Any words you'd never use? Any that land exactly right?

When something is off, name it. "This is too formal" or "I'd never open with a question" is far more actionable, to you and to the system, than a general sense of wrongness.

If you're still getting generic output after uploading, the usual cause is the files: too short, not actually written by you, or not representative. Go back to Step 2 and audit them against it.

Step 6: Keeping It Current

Refresh every quarter

Your writing moves. Every three months, add two or three recent posts that represent how you write now, and retire files that no longer do. With 30 slots, this is a swap rather than an accumulation.

Add depth when you go deep

Write a long article, record a podcast, give a talk on something you want to post about — upload the transcript. Richer expertise material produces more specific ideas, which is where most of the value is.

Watch what you keep editing

The most useful signal you have is your own editing pattern. If you rewrite the same element every time — openings, closings, calls to action — that is not a one-off; it is a gap in what the system has learned. Add examples of how you do that thing well, or say so directly in your style description.

"The AI gets better at your unique style with each piece of content you create together. It's a team effort that improves over time." — Sebastian Kinzlinger, founder of ContentIn

What Good Output Looks Like

You'll know it's working when you read a draft and think "I'd have said it almost exactly like that". Not perfect — you'll still edit — but the gap between draft and published post shrinks from a rewrite to a pass.

Concretely, good output:

• Reaches for examples from your actual work, not generic ones
• Opens in your register — not more formal, not more clickbaity
• Uses your vocabulary, including the field-specific terms you genuinely use
• Has your structural rhythm
• Suggests ideas you actually want to write about

See It In Practice

FAQs

Can ChatGPT copy your writing style?

Partly, and temporarily. Give ChatGPT three to five samples of your writing plus a written description of your style, and it will approximate you for a few drafts. The limitation is persistence: unless the material is in the context every time, the output drifts back toward the model's default register. That's the difference between prompting a general-purpose model and using a tool that keeps your writing in front of it on every generation.

What file types can I upload to ContentIn?

Plain text (.txt) and PDF. You can also paste a URL, which pulls in and extracts the article behind it, or paste raw text into the Transcript tab — useful when the source is an email or a talk you only have as a transcript.

How long does training take?

Text is extracted from your file straight away. Knowledge extraction runs on a schedule roughly every four hours, so a file typically moves from draft to extracted within a few hours of upload. If it stays in draft much longer, check that your profile has content pillars defined.

How many training files can I add?

Up to 30 per profile. That is enough to be genuinely selective and not enough to dump an archive into, which is the right constraint — quality of material matters far more than volume.

What's the difference between voice training and expertise training?

Voice training is about how you write and comes from your own LinkedIn posts. Expertise training is about what you know and comes from long-form material you upload. A tool with only the first writes fluent posts with nothing in them; one with only the second writes accurate posts that sound like somebody else.

Related posts

• LinkedIn About Me Examples
• The LinkedIn Algorithm and Content Format Guide
• 30 LinkedIn Hook Examples