Andrew McGuire · GTM Engineer → Agent Operator

Setup

Four steps. Ten minutes.

Create a Claude project

Go to claude.ai → new project → paste the Voice Profile Builder into the project instructions.

Feed it your posts

Export your LinkedIn posts as CSV, or paste 15–30 posts as text. Include engagement numbers. More data = sharper profile.

Save the output

Claude produces a performance analysis and a voice profile. Save the voice profile — you'll paste it into every future pattern.

Use it with patterns

Each Friday, Shipped includes a new pattern. Paste it alongside your voice profile and get a draft in your voice using a proven structure.

Before you start

Export your LinkedIn posts.

The builder works best with a full export. Here's how to pull your data from LinkedIn.

Option A · LinkedIn data export

The official route. Takes 24–72 hours but gives you everything.

1.Go to LinkedIn → Settings & Privacy → Data privacy → Get a copy of your data
2.Select "Posts" (you can deselect everything else to speed it up)
3.Click "Request archive" — LinkedIn emails you when it's ready
4.Download the ZIP, find the Shares.csv file inside
5.Upload that CSV directly to the Claude project alongside the Voice Profile Builder

Note: The LinkedIn CSV doesn't include engagement numbers (likes, comments). The builder will analyze voice and structure without them, but it can't score performance. Pair it with Option B for the full picture. If you get stuck, LinkedIn's data request page walks you through it.

Option B · Copy and paste

Faster. Works right now. Best for founders who post weekly.

1.Go to your LinkedIn profile → Activity → Posts
2.Open each post, copy the full text
3.Paste into a text file or directly into the Claude conversation
4.Include the date and engagement numbers (likes, comments, reposts) for each — visible at the bottom of every post
5.15–30 posts is the sweet spot. Prioritize recent posts (last 6 months)

Format tip: Paste each post as: date, then post text, then engagement numbers on a new line. The builder handles messy formatting — just make sure each post is clearly separated.

Inputs

What to give it.

Required — one of these

A CSV export of your LinkedIn posts (post_url, date, likes, comments, reposts, post_text). Even 50 posts is enough to score.

Or 15–30 posts pasted as text with dates and engagement numbers if you have them.

Strongly recommended

Company name + what it does

Your ICP — who you write for

Your thesis — the one belief that drives most of your content

Optional

A post you think represents your best work

A post that felt off or underperformed

Words or phrases you want banned from any AI draft

Phase 1

Performance analysis.

The numbers come first. You need to know what works before you analyze why.

1A · Baseline metrics

Total posts, date range, posts/month trend, average likes/comments/reposts, comment-to-like ratio (the quality signal — high ratio means real conversations, not just scroll-stops), and 6-month engagement trend.

1B · Content categorization

Every post tagged with one primary category:

Product/companyMarket takeBuilder/tacticalCustomer storyPersonal/founderData/researchEngagement playAmplification

1C · Per-category performance

For each category: post count, % of total output, avg likes/comments/reposts, avg comment-to-like ratio, and the best-performing post. Sorted by engagement rate — shows which categories your audience actually responds to.

1D · Post scoring (0–100)

Every post scored using five weighted signals:

Engagement deviation 30%
Comment quality 25%
Reach proxy 20%
Recency 15%
Repost signal 10%

Top 10%

The formula for 'do more of this'

Bottom 10%

What to stop or rethink

Hidden gems

High comment ratio, mid reach — the audience engaged deeply

Viral flukes

High reach, low comment ratio — got seen but didn't start conversations

Phase 2

Voice analysis.

Focused on the top 25% of posts by score. That's where the real voice lives — the version of you your audience responds to most.

2A

Sentence architecture.

→Average sentence length in top posts
→Deliberate length variation vs. uniform rhythm
→Fragments and one-word sentences — frequency and intent
→Paragraph length patterns
→3–4 representative sentences quoted from the data

2B

Vocabulary fingerprint.

→5–10 signature words/phrases from the data
→Technical jargon vs. plain language
→Contraction patterns (always, sometimes, never)
→Verbal tics or signature phrases
→Words they never use

2C

Opening patterns.

→How top posts start (statement, number, story, provocation)
→2–3 most common opening structures with examples
→Hooks vs. mid-thought starts

2D

Structural patterns.

→How they build a post (linear argument, story→lesson, data→interpretation)
→Lists vs. paragraphs
→Transitions between ideas
→Word count range for top performers

2E

Closing patterns.

→How they end (CTA, question, statement, punchline, trail off)
→Engagement asks vs. letting the post stand

2F

Tone and stance.

→Confident vs. tentative
→Wins vs. failures and mistakes
→Generous to competitors or combative
→Role: teacher, peer, builder, analyst, or provocateur

2G

Thesis clarity.

The core belief that runs through multiple posts. How explicitly they state it vs. letting it emerge. What percentage of posts connect back to it.

2H · The most important section

The unscriptable part.

What about this voice depends on the person's actual lived experience, role, access, or position — the thing a writing tool can approximate structurally but can't make real. This section determines whether the whole system works.

Output

What you get back.

A structured voice profile you save once and paste into every future pattern conversation. The section names are the API that patterns call.

# [Name] - Voice Profile
Built from [X] posts ([date range])
## Identity
Role · ICP · Thesis
## Performance snapshot
Posts analyzed · Posts/month · Avg engagement
Comment-to-like ratio · Top category · The number that matters
## How they sound
2–3 sentences distinguishing this voice from any other founder
## Sentence architecture
## Vocabulary fingerprint
## Opening patterns
## Structural patterns
## Closing patterns
## Tone and stance
## Thesis
## Format and rhythm
## The unscriptable part
What depends on lived experience. The section patterns
reference when they need the real tension.

The system

How patterns plug in.

The pattern provides

→The structural pattern (setup, body, close)
→Why the pattern works
→What signal it creates for GTM
→A ready-to-use prompt with blanks you fill in

Your voice profile provides

→Sentence architecture rules
→Vocabulary and tone
→Opening and closing patterns
→The unscriptable context you supply

The prompt inside each pattern references sections of your voice profile by name. That's why the structure can't change — it's the API the patterns call.

Rules

How the analysis runs.

Every claim in the performance analysis needs the number behind it. No unsupported assertions.

Quote actual words when identifying voice patterns. Don't paraphrase style — show it.

If a pattern is based on fewer than 10 posts, flag the sample size.

Don't prescribe improvements. The voice profile is a diagnostic, not a coaching session.

The voice description must be specific enough to distinguish this founder from any other founder in their space.

The "unscriptable part" section determines whether the whole system works. It gets the most attention.

Voice Profile Builder.