How to Build a LinkedIn Content System in Claude Code

Build a LinkedIn content system in Claude Code that turns your own stories into drafts in your voice: six files, one check script, three drafts a week.

This is a guide to building a LinkedIn content system in Claude Code: a few files and one script that turn your own material into drafts that sound like you.

The most useful part is a filter that kills ideas before a single draft gets written. Mine asks one question: would I have wanted to learn this two years ago?

Claude Code is Anthropic’s AI assistant that works with the files on your computer. Claude, the AI model behind it, writes the drafts.

We’ll build the system for one example: Maya, a solo consultant in Denver who sets up CRMs and pipeline reports for B2B companies with 20 to 200 employees. She posts when she remembers, and every draft from a plain prompt sounds like the same consultant.

By the end you have six files, one check script and one skill. Once the files are filled, say one hour on Monday gives Maya three checked drafts for the week.

This guide is for operators, founders and consultants who post about their own work on LinkedIn. Build it in order. If you already have a voice file, start at step 2, the story bank.

How the system works

Six small files and one script do the work. A skill holds the order. A skill is a file of instructions that Claude Code follows when you type its name, for example /linkedin-post.

Claude reads the files, writes inside the rules, runs the script and hands you a draft plus a list of warnings. Here’s the folder:

content/
├── CLAUDE.md                     # what this is, which file to read when
├── voice/voice-profile.md        # your real lines, sorted by what they do
├── stories/story-bank.md         # one entry per story
├── ideas/topic-bank.md           # raw ideas, unfiltered
├── log/posts.md                  # every published post, tagged
├── log/killed.md                 # ideas the filter killed, and why
├── tools/check_post.py           # the checks every draft must pass
└── .claude/skills/linkedin-post/SKILL.md

CLAUDE.md is the file Claude Code reads at the start of every session in that folder. Keep it short: what the folder is for and which file to read for which job.

The skill runs these steps:

  1. Take an idea from the topic bank or a story from the story bank.
  2. Run the filter. If the idea fails, log the reason in killed.md and stop.
  3. If the post touches something you sell, answer the four value questions first.
  4. Draft from the voice profile and the story, using only facts from the story bank.
  5. Run check_post.py, fix every error, keep the warnings.
  6. Hand you the draft with the warnings listed.
  7. After you post, add a row to log/posts.md.

There’s a separate guide on setting up Claude Code itself (CLAUDE.md, skills, memory). If you’ve never used it: create the folder, open the Terminal, go to the folder and type claude. Then paste:

Set up a LinkedIn content system in this folder with the structure
I'll paste below. Create every file with a short header that says
what goes in it. Write CLAUDE.md in under 15 lines: what this folder
is for and which file to read for which job. Then write the skill
.claude/skills/linkedin-post/SKILL.md with the seven steps I'll paste.
Don't write any post yet.

Paste the folder tree and the seven steps from above. The files stay empty until steps 1 to 3 fill them.

1. Build the voice profile from lines you wrote

A voice profile is a file that tells Claude how you write. If it’s made of adjectives, it gives the model nothing to copy. “Direct, warm, a bit funny” describes half the people on LinkedIn. Thirty lines you wrote describe you.

Collect them from places where you weren’t trying to sound like a post:

  • Posts that got real replies, beyond likes
  • Emails where you explained something to a client
  • Messages where you disagreed with an idea
  • Comments you left on other people’s posts
  • Voice notes you transcribed, in your own words only

Aim for 30 to 50 lines. Then let Claude sort them by what each line does, with a prompt that forbids rewording:

Here are 40 lines I wrote myself: posts, emails, messages, comments.
Sort them into these groups: how I open, how I state an opinion,
how I soften a claim, how I disagree, how I close, words I use often.
Quote every line exactly as written. Never reword.
Never add a line I didn't write.
Then name 5 patterns you see, each proven by two of my lines.
Don't describe my voice with adjectives.

The patterns show you your own habits. The quoted lines are what the model copies.

The second half comes from your edits. Every time you rewrite a draft, tell Claude what you changed and why, and it adds a line to the Never section. Maya always replaces “stakeholders” with the actual role, and she deletes every closing question. After a month, in my experience, that section holds more of your voice than the quoted lines.

Here’s the template:

# Voice profile: [name]
Updated: [date] · Built from: [which posts, emails, notes]

## Who I write for
One reader in one line: [role, company size, what they're stuck on]

## How I open (real lines, quoted exactly)
- "..."
- "..."

## How I state an opinion
- "..."

## How I soften a claim
- "..."

## How I disagree
- "..."

## How I close
- "..."

## Words I use a lot
[list]

## Never (one line per edit I keep making)
- [what I cut] -> [what I write instead] · why

## Shape
[words per post, line length, emoji rule, links]

2. Fill a story bank with a number, a scene and a cost

In my experience, drafts turn generic when the model has nothing specific to tell. You can tune the prompt for weeks. The fix is in the input: a file of real moments the model may use, and nothing outside it.

Every entry needs three things before it counts:

  • A number you counted. “4,212 duplicate contacts” reads like something you counted. “Thousands of duplicates” reads like marketing.
  • A scene. A day, a place and a line someone said, in quotes. That line is usually what turns a lesson into a story.
  • A cost. What it cost: hours, a delayed launch, a lost deal, a week of rework. A win without its cost reads like bragging, at least to me.

Opinions without a scene go into the topic bank. They wait there until something real happens that gives them a scene and a cost.

## S07 · The 4,212 duplicates
Number:  4,212 contacts with two owners, found in a CRM migration
Scene:   Day 3 of the migration. The sales lead asks on a call:
         "Why is Jordan calling my accounts?"
Cost:    The migration took 11 days instead of 3. Two reps called
         the same prospect within an hour of each other.
Lesson:  Check contact ownership before you migrate anything.
Reader:  Sales managers at companies with 20 to 200 people
Format:  story
Used in: not yet
Source:  migration notes, client call
Client OK to tell: yes, without the company name

Filling it takes five minutes. At the end of a work session, ask Claude: “Anything from today for the story bank? Number, scene, cost, or nothing.” Most days the honest answer is nothing. A bank of fifteen real entries is worth more than fifty half-remembered ones.

3. Kill ideas with one question before you draft

Once the voice and the stories are in place, Claude writes a decent draft for almost any idea. That’s why you need the filter. A decent post about something your reader already knows costs them time, and it teaches them to skip you.

The question: would I have wanted to learn this two years ago?

It runs as step 2 of the skill, before any drafting. Claude has to say in one sentence what you, two years ago, would have learned from the post, and the answer has to be one of four things:

  • A workflow you can copy
  • A decision rule
  • A mistake to avoid, with its cost
  • A number you didn’t know

If the answer is an attitude (“data quality matters”), the idea dies and goes into killed.md with the reason.

Idea (Maya)What Maya two years ago learnsVerdict
Why clean CRM data mattersNothing she didn’t knowKilled
The 4,212 duplicates and the ownership checkA check to run on day 1 of any migrationPassed
My favorite reporting toolsA list, no decisionKilled
The 31 fields she deleted and the 5 she keptA rule she can copy on MondayPassed
Thoughts on AI in RevOpsAn attitudeKilled, back to the topic bank

Expect it to kill a lot. Say Maya runs ten ideas through it in a week and four survive. That means the filter works.

Why two years: in my experience it puts the reader one real step behind you, and that’s the reader a how-to post helps. One year back, you’re writing for peers who know it already. Five years back, you’re explaining basics.

Read killed.md once a month. Sometimes a dead idea comes back, because a story finally happened that gives it a scene and a cost.

4. Answer four value questions before you name a feature

Posts about something you sell, or something you built, turn into feature lists fast. Answer four questions, in this order, before any feature gets named. They keep the draft on what the reader gains:

  1. What does it cost today, done the usual way? Use the reader’s own units: hours a week, a consultant project, a hire, a quarter of delay.
  2. What does it cost the new way? Use the same unit, so the reader can compare without doing math.
  3. What does one unit of that cost look like? Turn the number into a moment the reader has lived through.
  4. Why can’t it be done by hand? Show the math that proves the manual way doesn’t scale.

Here’s Maya’s CRM cleanup as an example:

  1. Say a sales manager spends 5 hours a week fixing pipeline data before the forecast call. In 50 working weeks, that’s 250 hours.
  2. Maya’s setup takes two weeks, then a 20-minute check every Monday.
  3. One duplicate contact is a rep calling a prospect her colleague spoke to an hour earlier, and the prospect saying so.
  4. 4,212 duplicates at 4 minutes each are 16,848 minutes, or 281 hours. That’s seven working weeks of 40 hours, and new duplicates arrive every day.

Only after that does a feature get one line, if it gets one at all: “an ownership check runs every night.”

In Claude Code this is a block in the skill. If the post mentions your offer, the draft starts as four answers. If Claude can’t answer the first question with a number or a concrete example, the post isn’t ready and goes back to the topic bank.

My take: LinkedIn readers respond to the value version, and the technical details land better on X. I’d write the deep dive there.

5. Let a script check every draft

Every draft runs through a script before you see it. The script catches what can be counted, and the rest is your job.

Ask Claude Code to write tools/check_post.py from your Never list and the checklist below. A script here is a short Python program. It reads the draft and the last five rows of log/posts.md, prints ERROR or WARN for each rule, and stops with an error if anything failed. The prompt:

Write tools/check_post.py. Input: a draft file. Read the Never list
and the length range from voice/voice-profile.md and the last 5 rows
of log/posts.md. Check every "Script" and "Variety" line of the
checklist I'll paste. Print ERROR or WARN with the rule and the line
it found. Exit with code 1 if there is any ERROR.
Add a test file with one draft that breaks every rule.

The skill fixes errors on its own and shows you the warnings, because some of them need your judgment.

AI patterns, the habits that make text read as machine-written, are the cheapest to catch. A word list and a few text patterns (regular expressions, a short code for “find text that looks like this”) find most of them:

em dash, semicolon, hashtag
"not X, but Y" and "It's not X. It's Y."
three-beat rhythm lines ("Build. Test. Ship.")
a question as the first line
filler: delve, leverage, unlock, game-changing, in today's,
  here's the thing, let that sink in, at the end of the day

Your style rules come from the voice profile: the length range, line length, the Never list and how a post may end.

Variety is checked against what you already posted. Readers see your posts over weeks, so the script compares the draft with the log as well. Each row in log/posts.md holds the date, the format, the opener type, the first six words, the closing type, the story ID and, a week later, the results.

Here’s the full checklist:

PRE-POST CHECKLIST

Filter
[ ] One sentence: what me two years ago learns from this
[ ] It's a workflow, a decision rule, a mistake with its cost, or a number

Script: check_post.py, every rule must pass
[ ] No word or phrase from the Never list
[ ] No em dash, semicolon or hashtag
[ ] No "not X, but Y", no three-beat rhythm lines
[ ] Line 1 is a claim or a number, never a question
[ ] Inside the length range from the voice profile
[ ] Every number is in the story bank or framed as an example

Variety: against the last 5 rows of log/posts.md
[ ] Format differs from the last post
[ ] Opener type not used in the last 2 posts
[ ] First word not used in the last 5 posts
[ ] Same story not used in the last 8 weeks

You
[ ] The scene happened, and the quoted line was said
[ ] You'd say every sentence to a client on a call
[ ] Value questions answered before any feature (posts about your offer)
[ ] Anyone in the story would be fine reading it

6. Post three times a week and keep the rest manual

I’d start with three posts a week and give each day its own job: a how-to post on Monday, a story on Wednesday, an opinion on Friday. Keep that mix for six weeks before you change anything. Otherwise you can’t tell what worked.

One batch session on Monday covers the week. Claude pulls eight to ten candidates from the story and topic banks, the filter kills some, you pick three, Claude drafts and checks, and you edit. Say that takes one hour once the files are filled.

Some parts stay with you on purpose:

  • Picking the story. Claude can rank them. You know which one you can tell honestly this week.
  • The quoted line. Only you know what was said in the room. If you don’t remember the exact words, write what you remember without quotation marks.
  • The final edit. Every change you make becomes a Never line. That’s how the profile gets sharper month by month.
  • Posting. I post by hand. It takes a minute, and it’s one last read. If you’d rather schedule, pick a tool from LinkedIn’s own marketing partner directory, like Buffer, Hootsuite or Sprout Social, and check that it posts to personal profiles.
  • Replies to comments. All of them, yourself. In my opinion a model’s reply under your name is the first thing readers notice.
  • Feeding the story bank. Five minutes after anything that cost or saved real time.

Add the results to the log after seven days. Two more columns in log/posts.md: impressions and comments, copied from LinkedIn a week after the post. Once a month, ask Claude Code to rank each post against your own other posts and print the average per format and per opener type. Compare a post only with your own posts, never with other accounts, because their reach mostly reflects how many followers they have.

Where content systems go wrong

  • Adjectives in the voice file. If it says “authentic and bold”, delete that line and add three real ones.
  • A story bank full of lessons. An entry without a scene is an opinion. Move it to the topic bank.
  • Rounded numbers. “4,212 duplicates” reads like something you counted. A round “4,000” reads like a guess. Use the real number, or frame it clearly as an example.
  • A filter that never kills. If nothing died in a month, the question has turned into a formality. Make Claude write the one sentence, and read it yourself.
  • Other people’s posts in your voice file. Studying how others build a post is fine. Pasting their lines into your profile makes you sound like them.
  • Details that aren’t in the bank. A draft may only use scenes, numbers and quotes from the story bank. Put that rule in the skill, and check every scene against the bank before you post.

Where to start today

You can set up your content system today in three steps:

  1. Paste 30 lines you wrote in the last year and build your voice profile with the prompt from step 1.
  2. Write three story-bank entries from the last month, each with a number, a scene and a cost.
  3. Draft your first post from one entry and let the script check it before you post.

Blog

Keep Reading

Start With a System Check

Start with a free target market map or a short system check. Either way, you leave with something useful.