TL;DR. The fix isn't a better prompt. Stop using generic Claude skills and build one from 5 - 15 of your own best-performing posts. Two or three feedback loops gets you to 90%. With this system, eight out of ten posts ship without further edits.
More and more marketing teams are trying Claude to scale up their social media game. The output sounds like AI. Each post needs heavy edits before it ships, and at that point, the AI hasn't saved you anything.
There's one rule that fixes this. I broke it down into a seven-step setup that you can implement today.
Use AI to augment, not replace your content production
Stop using generic Claude skills. Build your own skill from your past posts that actually performed.
The generic skills in Claude's library like "How to write for LinkedIn" or "How to talk to CFOs" are generic. They don't know your audience, your voice, or which of your posts have ever moved the needle for your specific value proposition, or your specific target persona.
What is needed is a skill trained on your best posts, enriched with best practices on top voices. We will run through this below using the example of LinkedIn.
Focus on your personal voice
There are two reasons why generics will fail:
The generic skill produces text in the median voice of LinkedIn. That voice is now the AI-slop voice. Everyone reads it and nobody stops scrolling.
Your voice is the thing your audience signed up for. They didn't follow you to read posts that sound like everyone else's. They followed because something in your past output landed. Train the skill on that something.
Seven easy steps to get there
Run through these seven steps to execute this.
Step 1: Get the best practices for your space
Tell Claude to run a deep research on how the top voices in your space write. Not "how to write LinkedIn posts", but the specific voices your target audience already follows. The goal of this step is to leverage the learnings of top voices in your space.
Prompt: "Run a deep research on how [3–5 named accounts] structure their LinkedIn posts. Look at: hook patterns, post length, paragraph breaks, what they do with the first line, how they close, when they use a hot take vs a story. Give me a structured breakdown. Additionally: Run a deep research on top voices and content creators for [your audience]. Use this to expand the list of accounts we will learn from. Take your time."
Step 2: Train Claude on YOUR posts
Pull 10–20 of your best-performing posts. Don't filter by what you liked. Filter by what your audience responded to: comments, shares, likes.
Feed those to Claude as context.
Step 3: Brainstorm with Claude
This is important to reflect both for yourself why you composed the posts the way you did, and to capture insights on best practices from top performers on the market.
Ask Claude two questions:
"Why did each of my posts perform? What's the through-line in voice, structure, hook?"
"Compare to the patterns from step 1. What do top voices do that I don't? What do I do that they don't?"
Read through what Claude processes for your, and make sure you build your own opinion on what patterns are important to you. This is the baseline to build your skill.
Step 4: Build a skill out of this
Your now have the material. Compile it into a Claude skill: instructions, examples, a do/don't list. The skill should be specific enough that two people reading it could spot a draft that violates it.
Prompt: "Now build a skill out of this and ensure that we can use these instructions to draft posts for LinkedIn in the future. Make sure the learnings from the previous chat above are reflected in that. Before drafting the final skill, run a deep research on the web on best practices on how to formulate skills for Claude and apply them."
Step 5: Let Claude draft a post
Give the skill a topic. Let it produce a draft. It'll be at 60–70%. That's normal. Don't stop here.
Step 6: Refine
Read the draft. Find the parts that sound off. Tell Claude exactly what to change and why: not "make it punchier", but "the opener is too soft, your first line should always be a one-sentence stake-in-the-ground."
Run again. Now you're at 75%. One more pass. Now you're at 90%.
Step 7: Update your skill
When the draft is good, update the skill with the corrections you just made. The next post starts from a better base.
Continue this loop as you produce posts and after a few iterations, Claude will provide posts which require less and less refinement from you.
What happens after a few rounds
I run a podcast as a side hustle. The skill I built off my past posts produces LinkedIn content that ships without further edits eight out of ten times.
That's the number you want. Not "AI doing all the work", but Claude getting to 90% so you only spend time on the last mile.
The mistake most people make: they do this once, see the first draft at 60%, decide AI doesn't work, and walk away. The whole game is the second and third loop. That's where the skill becomes yours.
FAQ
Why doesn't Claude's built-in LinkedIn skill work for me?
It's built on the median voice of LinkedIn, but generic enough to apply to anyone, which means it doesn't sound like anyone in particular. Your audience didn't follow you to read median LinkedIn copy. They followed because your voice landed. Train the skill on that, not on the average.
How many past posts do I need to train Claude on?
10–20 is the sweet spot. Fewer and Claude doesn't have enough signal to extract your voice patterns; more and the skill starts averaging across too much variation. Pick the ones with real engagement (saves + thoughtful comments), not vanity metrics.
What if my voice changes over the next six months?
That's what step 7 is for. Every refined draft is a chance to update the skill with what you just corrected. Six months in, the skill matches your current voice, not where you were when you set it up.
Is this the same thing as fine-tuning Claude?
No. Fine-tuning is model training, expensive, slow, and unnecessary here. A Claude skill is a structured prompt + reference set; you can build it in an hour and update it any time. Save fine-tuning for when prompt-based skills aren't enough (which for content style, they almost always are).
