How to Make AI Write Like You: writing-style-skill for Claude Code

By Beau Johnson··8 min read

Quick answer

writing-style-skill is a free, open-source GitHub repo that learns your voice from your edits. You edit an AI draft, observe.py records the before and after, and improve.py pulls rules from the changes and ranks them by confidence. Those rules get written back into a SKILL.md file that works with Claude Code or OpenClaw.

In this video I walk through a free, open-source GitHub repo called writing-style-skill. It teaches your AI to write in your voice by learning from the edits you already make to its drafts. If you use AI for scripts, captions, newsletters, blog posts, or cold outreach and the output still sounds like a robot, this one is for you.

I recorded this in March 2026. Tools like this change fast, so check the repo's current README before you set it up.

What I cover in this video

  • 1:32 Why AI writing sounds like AI writing, and why pasting in examples isn't enough
  • 3:03 The five-step feedback loop behind writing-style-skill
  • 5:20 Under the hood: SKILL.md, observe.py, and improve.py
  • 6:05 Confidence-ranked rules (P0, P1, P2) and built-in rollback
  • 7:36 A real example: the voice profile I built by hand for my own content
  • 8:22 Setup for Claude Code and OpenClaw
  • 9:54 The bigger picture: using AI to amplify your voice instead of replacing it
  • 12:57 Limitations and my honest take

Why AI writing doesn't sound like you

The biggest complaint I hear in our community is that AI writing sounds like AI writing. You give Claude or ChatGPT a prompt and get back something technically correct that reads like a corporate training manual: "delve into this," "leverage that," "navigate the landscape." Nobody talks like that.

You can write huge system prompts, paste in examples of your writing, or build custom instructions. It still drifts, and you end up reminding it over and over. AI doesn't actually understand your voice. It's pattern matching, and every new conversation starts from zero. All the context about how you write is gone.

The other problem is that your edits get thrown away. You fix the draft, move on, and the AI makes the same mistake next time. writing-style-skill keeps those fixes.

How the feedback loop works

The concept is simple. A developer who goes by JZOCB on GitHub built it, and it runs in five steps:

  1. The AI writes a first draft using whatever rules are in your skill markdown file.
  2. You edit the draft until it sounds like you.
  3. You run a script that compares the original AI version to your edited version.
  4. It pulls rules out of that difference: what you changed, what patterns you added, what you removed.
  5. Those rules get added back into the skill file, so the next draft follows them.

You write, you edit, it learns, and it repeats. Each round gets a little closer to your voice.

The three files

  • SKILL.md is the brain. All your writing rules live here, things like "always use contractions," "never use em dashes," "start sentences with and or but," or "keep paragraphs short." The template it ships with covers voice and tone, sentence structure, vocabulary preferences, formatting conventions, and common anti-patterns to avoid. You can start with the template and let the auto-learning fill in the gaps, or customize it right away.
  • observe.py is a zero-dependency Python script, with no pip installs. It records two things: the AI's original draft and your edited version. It doesn't care how many rounds of editing happened in between. Because it's pure Python, it runs anywhere: your laptop, a server, or a Raspberry Pi.
  • improve.py is the smart one. It diffs the two versions, uses an LLM to work out what changed and why, and turns those changes into writing rules.

Confidence ranking and rollback

This is the clever part. improve.py ranks rules by confidence. P0 rules are high confidence, meaning changes you made consistently across multiple samples, and those get applied automatically. P1 and P2 rules are lower confidence, like a one-off edit, and those get flagged for review. So it isn't blindly absorbing every change. It tries to tell your real style apart from something like a typo fix.

It also backs up your SKILL.md before every update, with one-click rollback, so you won't accidentally wipe out your style rules.

Setting it up with Claude Code or OpenClaw

The easiest route is to copy the repo link, give it to your AI agent, and say:

Help me implement this.

If you want to do it yourself:

  • Claude Code: clone the repo and copy it into your skills folder. That's it.
  • OpenClaw: there's a one-line ClawHub install command.

improve.py automatically detects which LLM CLI you have. It will use Claude Code if it's installed, or the llm pip package, or you can point it at something else with an environment variable. It works with pretty much any setup.

After that, each cycle is four steps: record the original, make your edits, record the final, and run improve. It takes about 30 seconds.

Why this matters for agents and teams

For my own content I built a voice profile by hand: use "y'all" naturally, no em dashes, start sentences with "so," "and," or "but," keep it conversational. That meant going through hours of old YouTube transcripts and emails I'd written to pull out the patterns. With this repo you skip that work. You just create content, edit it to sound like you, and the style guide builds itself over time.

If you're in the OpenClaw ecosystem, this plugs right in. Every agent that writes content for you reads from the same SKILL.md, so they all improve together. The same idea works for teams: one master skill file captures the brand voice, everyone edits toward the same standard, and nobody has to memorize a 50-page style guide. For more on building these kinds of setups, see my AI agents page.

Limitations

  • It's only as good as the model doing the analysis. If you use a smaller model for the improve step, you may get less accurate rules.
  • It needs volume. Two or three edit pairs isn't enough to separate real patterns from noise. Plan on using it consistently for at least a few weeks before it really starts to work.
  • It's not automatic. You have to remember to run the observe and improve scripts after you edit. For some people that will be a deal breaker.

People have tried to solve this with custom GPTs, fine-tuning, and prompt engineering. In my view those are too expensive, too complex, or too fragile. This one is simple, free, open source, and works with tools most of us already use.

Key takeaways

  • Your edits to AI drafts are training data. writing-style-skill captures them instead of throwing them away.
  • The repo has three parts: SKILL.md (rules), observe.py (records before and after), and improve.py (extracts and ranks rules).
  • High-confidence P0 rules apply automatically. Lower-confidence P1 and P2 rules get flagged for review, and every update is backed up.
  • It works with Claude Code (clone into your skills folder) and OpenClaw (one-line ClawHub install).
  • Give it a few weeks of consistent use. The goal is AI as an amplifier for your voice, not a replacement for it.

FAQ

What is writing-style-skill?

It's a free, open-source GitHub repo by JZOCB that teaches AI to write in your voice. It compares the AI's draft with your edited version, pulls writing rules from what you changed, and adds them to a SKILL.md file the AI follows next time.

Does writing-style-skill work with Claude Code and OpenClaw?

Yes. For Claude Code you clone the repo and copy it into your skills folder, and for OpenClaw there's a one-line ClawHub install command. The improve script auto-detects your LLM CLI, such as Claude Code or the llm pip package, or you can set one with an environment variable.

How long does it take for the AI to learn my writing style?

It needs volume. Two or three edit pairs isn't enough to find real patterns, so plan on using it consistently for at least a few weeks before it really starts to work.

Will it learn from typo fixes and one-off edits?

It tries not to. Rules are ranked by confidence: P0 rules from consistent changes across samples apply automatically, while lower-confidence P1 and P2 rules are flagged for review. Your SKILL.md is also backed up before every update so you can roll back.

Chapters

  1. 0:00The #1 problem with AI writing
  2. 1:45How writing-style-skill works
  3. 3:30Under the hood: observe.py and improve.py
  4. 5:15Why your edits are being wasted right now
  5. 7:00Real example from our content setup
  6. 8:45Technical setup (Claude Code + OpenClaw)
  7. 10:00The bigger picture: AI voice amplification
  8. 11:30Limitations and honest take
  9. 12:45How this works with OpenClaw agents

Transcript

Read the full transcript

0:00One of the biggest problems with AI is getting AI content to actually sound like you. Well, in this video, we're going to be going over GitHub repo that has over 109 stars in just 1 day, and this repo solves that problem. So, I don't mean you like you paste in some examples and hope for the best and hope that it generates some type of AI generalized voice. You know, if you're doing outreach or you're creating marketing material, a lot of it is AI slop. So, this repo solves that problem, and I'm going to walk you step by step through the process so you can stick out when you actually reach out to people and create content that was created with AI, but it sounds in your voice. So, I'm going to break down like, you know, usually when AI writes something, you know, you have to fix it. You have to go

0:46back and make changes and make it sound like you wrote it. And it automatically learns from every edit. So, this repo basically attacks that problem. You know, the more that you use it, the better it gets. So, if you're building an AI agent, it's going to learn your editing style, your talking pattern, the way you write. And this thing is called the writing style skill. Like I said, I have all of the sources linked down below in all of these videos, so you can go check them out yourself. You know, so if you've ever like been frustrated by AI writing that sounds like a robot trying to be your friend, this is going to change from here on out if you just go and get this repo, give it to your AI agent, and start making changes. So, I'm going to back up a little bit cuz I know that was a lot to start, but generally

1:32when you use AI to write content for you, scripts, outlines, Instagram captions, reaching out to people on LinkedIn, it just sounds like generic AI. We need to figure out how to humanize the voice profile for your AI agent, whether in OpenClaw Cloud Code, so it actually sounds like you. It writes like you, it talks like you. One of the biggest problems with AI right now and the number one complaint I hear people in our Shipping School community is that AI writing sounds like AI writing. And you know what I'm talking about? Game changer. You know, it's like, come on, man. Like, we know. We know. You know what I'm talking about. You know, you give Claude or ChatGPT a prompt, it comes back with something that's technically correct, but it sounds like it was written by a corporate training manual. You know, delve into this, leverage that, navigate

2:18the landscape. Nobody talks like that. That's crazy, right? You'd be insane. Nobody talks like that. You know, and here's the thing. You can write these massive system prompts trying to get it to sound like you. Or you can basically paste in examples of your writing, too, and you can build out these custom instructions, but it still drifts away from it, and you have to remind it over and over again. And it still ends up sounding off, not quite right. Because here's what people don't realize is, you know, AI doesn't actually understand your voice at all. It's basically just using pattern matching, and every time you start a new conversation, it's starting from scratch, and all that context about how you write is gone. It resets back to zero. That's where this repo comes in. You got to check it out, especially if you're doing a lot of cold outreach and marketing or content. A

3:03developer who goes by JZOCB on GitHub built something that should have existed a long time ago. It's called writing style skill, and the concept is very simple. It's actually kind of beautiful. Here's how it works. The first step is basically the AI writes you a first draft using whatever rules are in like your skill markdown file. The second step, you edit it, you fix it until it sounds like you. The third step is you run a script that compares the original AI version to your edited version. The next step is you automatically have it extract the rules from that differential. Like, what did you change? What patterns did you add? What patterns did you remove? Things like that. Step five would be those rules get added back to the skill markdown file, and now the next time the

3:49AI writes for you, it follows those updated rules. So, it's a great feedback loop. You know, you write, you edit, it learns, and then it repeats. And it gets better every single time. That's the cool part. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We [music] have a full Cloud Code course, a full OpenClaw course, and four live bootcamps every single week where we actually help you get set up from scratch. Like, actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me. We could share screens, and I can help you get Cloud Code or OpenClaw running on your

4:34machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Now, let me break down what's actually happening under the hood because I think it's important to understand this. The repo has three main pieces. First, you have the skill markdown file. That's essentially the brain. This is where all the writing rules live. Things like, always use contractions or, you know, never use em dashes. Start with sentences with and or but. Keep paragraphs short. Whatever, you know, whatever your writing sounds

5:20like, it should sound like your writing. The second thing is you got an observe Python script. This is a zero dependency Python script, basically. Literally no pip installs needed. All it does is record two things. Your AI's original draft and the file you edited. That's it. Two data points. It doesn't care how many rounds of editing you did between. It just needs the before and after. The third thing is you've got The third thing you have is the improve Python file. This is the smart one. It takes those two versions and runs a differential and uses the LLM to analyze what changed and why. Then it extracts the writing rules from those changes. And here's the really clever part of this whole repo is it ranks those rules by confidence. P0 rules are high confidence, things you changed

6:05consistently across multiple samples. Those get applied automatically. P1 and P2 rules are more lower confidence, things that might just be one, you know, one or two off edits, right? Like, you know, those get flagged for review. So, it's not just blindly absorbing every change you make, it's actually trying to figure out which changes represent your real writing style versus the ones where you're just trying to fix like a typo, which I think is pretty cool. And it also backs it up to your skill markdown file before every update. One-click rollback if something goes wrong, so you're never going to accidentally destroy your style rules. Safety net built right in. Now, here's where this gets interesting for most people, not just developers. If you think about it this way, you're creating content. If

6:50you're writing blog posts or if you're writing newsletters, you're writing scripts like what I'm doing right now for this YouTube video, you spend a lot time fixing the AI output, a lot. And I don't have time to be going back and reviewing stuff. Like, every fix you make is a learning opportunity with this repo that most tools just throw away. You fix it, you move on, and then the AI makes the same mistake the next time. I don't want that. This repo captures those fixes. It's so easy. Just copy the link of this repo, give it to your AI agent, and say, "Help me implement this." It will turn your editing time into training data. So, over time, you're spending less time editing and more time doing stuff, like making videos. And you don't have to worry about fixing things. You're actually building a permanent style guide that follows you everywhere. Let me give you like a real example. In our setup, specifically, I have a voice profiler

7:36for my content. It's like, use y'all naturally. I say y'all every now and then, right? You know, don't use em dashes in the copy. Geez, don't do that. Start sentences with so or and or I or but or if or whatever. Keep things conversational. And we had to manually write all that, took time, went through hours of transcripts on previous YouTube channels, and pulled out patterns. It wrote all my emails or took all my emails that I had written by hand, um, you know, to my reseller niche email list, and you know, it took a while to build that voice profiler. But with this repo, you can get it done in just minutes. And you get to skip all that, basically, all that manual work. You just start creating content and edit it to sound like you, and then the system builds your style guide automatically as

8:22time goes on. The more content you create, the more accurate it gets. The AI basically becomes a student of your writing. And then now you're the teacher every time you make an edit. I'm going to talk to you about the technical setup because it's not overly complicated. Just like I said, get the link from the repo and give it to your agent and say, "Implement this." But if you're using tools like Cloud Code, you just clone the repo, you copy it into your skills folder, you're done. If you're using OpenClaw, there's actually a claw hub install command, one line, and that's it. The improve script automatically detects what LLM CLI you have, so easy. It'll use Cloud Code if you have it or the LLM pip package or you can point it at whatever you want with an environment variable. Doesn't matter. So, it works with basically any setup. And the skill markdown file template it comes with is

9:09actually pretty good. It covers the majority categories of writing styles, voice and tone, sentence structure, your vocabulary's preferences and formatting conventions, common anti-patterns to avoid as well. You can start with the template and let the auto-learning fill in the gaps, or you can customize it right away and let the system refine it. Whatever you want to do. Doesn't matter. So, I want to talk to you about why this matters beyond just making content sound better because there is a huge picture here. You got to zoom out. We're in this era right now where everyone is using AI to write, and you can know. You I can tell. You can tell if you've been in the AI space, you know, that it's written by AI. How are we going to stand out if everyone's written write using AI to write? I don't know. So, it's all sounds the same. You can scroll through LinkedIn. Like, go to LinkedIn right

9:54now. It's probably 90% AI content or AI generated content. You know, go look at some blog posts and you can tell is instantly written by AI. It has all the same smooth, generic, slightly enthusiastic tone, like really helpful customer service rep who's had way too much coffee, for example. The problem isn't that people are using AI. That's not the issue. The problem is that they're not teaching AI to be like them. They're just accepting whatever the default output is and that default output is designed to be inoffensive and generic and applicable to everyone, which means it sounds like nobody. The repo is the solution to that. Like I said, it's linked down below. You got to check it out. It's the difference between using AI as a replacement for your voice and using AI as an amplifier

10:39for your voice. You're still doing the editing and you're still making the calls about what sounds right and what doesn't. But now those calls are actually compounded over time instead of disappearing into like a void of nothingness and restarting over and over again. Think about where this goes in 6 months from now if you're using this repo today. You've been using this system for half a year. You're editing content every week and the skill markdown file has absorbed hundreds of corrections. At that point, the AI's first drafts are going to be so close to your voice that you barely need to edit them anymore. That's a dream come true for content creators. And that's where this is headed. If you're struggling to keep up with content, well, I'm about to save you about 40 days [music] worth of work. I built something called content machine. It's 10 AI agents that run on the open claw orchestration and they

11:26handle everything. Scripts, thumbnails, X posts, [music] blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake [music] up and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them and I move on with my day. It works for any niche, fitness, [music] finance, real estate, marketing, whatever you are building and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own

12:12thing and it molds it to you. It learns how you talk and it [music] writes so it doesn't sound like AI slop. $97 one time. It's not a subscription. I'll put the link down below and you'll thank me later. Here's another angle that I think's really powerful. If you're running a team, for example, or if you got multiple people writing content for a brand, you can just have one master skill markdown file that captures the brand voice. Everyone edits toward the same standard, the rules coverage, and then you end up with consistent brand voice across your entire content without anyone having to memorize a 50-page style guide, a style guide that essentially builds itself. I also want to point out that something that the architecture that I think is pretty smart here, the observe script is zero dependency. Like it's pure Python. No

12:57installs. That means it runs anywhere, your laptop, a server, a Rise Raspberry Pi, or whatever. The improved script needs an LLM though and the observation and the recording part has zero friction. You're never going to hit a situation where you can't record your edits basically because some package is broken. Let me tell you what I think the limitations are because I do want to be straight up. It's a new repo, 109 stars in just a few days. It's pretty good. The auto extraction is only as good as the LLM that's doing the analysis, too. So, you know, if you're using a smaller model to the improved step, let's say, you might get less accurate rules. Also, it needs volume to work well. If you've only recorded two or three edit pairs, the system doesn't have enough data to find real patterns versus the noise. You

13:42need to use it consistently for a few weeks minimum before it even starts to shine. It does require discipline and patience. You actually have to run the observe and improve scripts after you edit. It's not automatic in the background. You have to actually remember to do it, which for some people is going to be the deal breaker, I know. With the A of a the era of AI agentic work, you know, it takes like 30 seconds. So, come on. Record the original, do your edits, record the final, and run improve. Four commands done. The only thing I want to mention about this is, you know, that it works with open claw skills specifically. So, if you're already in the open claw ecosystem, this plugs right in for you, you know. Your agents can use this right away as part of their system. So, like for example, the Rizza, Inspect Deck,

14:27and any agent that writes my content for me as AI agents and their rules would improve over time as your agents, you know, improve simultaneously because they're all reading from the same skill markdown file, essentially. You know, people have probably already tried to solve this problem in other ways, custom GPTs, fine-tuning, prompt engineering, but those approaches are either too expensive or too complex or too fragile and this one's straight up so simple. It's open source. It's free and it works with the tools that most of us already using. So, if you're someone who creates content, let's say, regularly and you've been frustrated by your AI output that doesn't sound like you, go check out the writing skill on GitHub and I'll put the link again down below. Clone it, try it for a week, and I think you're going to be pretty surprised on how fast it

15:12learns how to build that voice profile like you. And, you know, I will say if you're building with AI and you want to connect with other people who are doing the same, come join us in the community. We have 148 members in the last 14 days. People building all sorts of stuff. We focus on community first, human-to-human interaction. AI is amazing, but we can't use that lose the human element. So, we do six boot camp calls every single week helping you take the next step in your AI journey. I'll put a link to that down below and we just have three spots left before the price goes up. So, get in before it does. If you haven't subscribed already, please do. I release literally four to six videos every single day. And this is me. I'm not an AI avatar. This is me. You can hear my cat screaming in the background. Hit subscribe, hit the notification bell,

15:58and the next video drops in about an hour. Be blessed.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading