> ## Content Index > Fetch the complete content index at: https://blog.annabyang.com/llms.txt > Use this file to discover other available public pages before exploring further. # How to Use Generative AI for Writing Assistance Without Creating Slop - URL: https://blog.annabyang.com/ai-writing-voice-system/ - Published: 2026-09-14T15:15:37.000Z - Updated: 2026-09-14T15:15:36.000Z - Description: A style guide alone isn't enough. - Author: Anna Burgess Yang - Tags: Writing You have a whole bunch of notes from a meeting and think they would make great LinkedIn posts. After all, the notes reflect *your* ideas. You feed your notes to a generative AI tool and ask it to turn your notes into four LinkedIn posts. The result? Super polished, AI-sounding posts. The ideas are there, but your voice and tone are not. Most solopreneurs know that the next step is to give AI a bunch of past examples of their LinkedIn content and hope that fixes it. You try again. Second pass still includes a word you'd never say, or the rhythm still feels “off” in a way you can’t exactly put your finger on. My AI-assisted writing uses three separate reference files. Each has *one* job. If you set up a good system, any AI output you receive will need a lot less editing (key words being "less editing" not "no editing." ## How I use AI for writing AI is an assistant in my writing process, never the writer. It helps me with things like outlines and research, or turning something I've already published into social posts. For things landing page copy or email copy, I get a first version, following a specific template I’ve created, and rewrite from there. Every one of those uses [starts from my ideas](https://blog.annabyang.com/maintain-authentic-writing-voice-with-generative-ai/). The raw material is mine. What AI does is speed up the mechanical parts of getting it out into the world, with a final editorial pass by me. But anywhere in that process, AI might try to sound like me. It proposes the H2s for a post on this blog (which I’ll either keep, edit, or remove). It writes the first version of the FAQs at the bottom of an article. Without clear writing guidelines, what comes back is slop: technically correct but completely generic. What I want instead is output that cuts down the amount of time I spend on the repetitive parts of writing so I can spend *my* time on narrative and personality. That only happens when I give my tools something concrete to work from. ## My 3-file system for AI-assisted writing When I first started using AI for writing assistance, I felt like I was correcting the same mistakes over and over. Eventually, I came up with my system of three reference files that my AI tool (Claude) uses every time. ### Voice and tone guide The first file is a voice and tone guide. It describes how I already write, such as “encouraging,” “educational,” and “occasionally humorous.” ![Screenshot of my LinkedIn voice and tone guide](https://storage.ghost.io/c/fe/e9/fee9a30d-4c70-4bd2-af95-fe6abc8afd29/content/images/2026/09/annabyang-voice-and-tone-guide.png) I fed ChatGPT (my AI tool at the time) a bunch of my blog posts and asked it to create a style guide for me. Then I reviewed the output and made some tweaks and clarifications. It also includes some structural information, such as title case for blog titles and sentence case for H2s and H3s when writing outlines. Other than the structural items, nothing in it is a rule. It's more like a baseline of how I sound. [How to Use AI to Create a Style GuideA tip for writers.![](https://storage.ghost.io/c/fe/e9/fee9a30d-4c70-4bd2-af95-fe6abc8afd29/content/images/icon/1-f740ee83-d6d7-4dd2-b860-d7522da25be3.png)Tinkering With IdeasAnna Burgess Yang![](https://storage.ghost.io/c/fe/e9/fee9a30d-4c70-4bd2-af95-fe6abc8afd29/content/images/thumbnail/tinkering-with-ideas-process-hero-image-3-831fa9bc-ba17-4c35-a7d6-e1791d7be3f7.png)](https://tinkeringwithideas.io/style-guide-ai/?ref=blog.annabyang.com) ### My “hard rules” file The second file is prescriptive. It's a short, explicit list of bans: words, phrases, and writing patterns I never want to see. No exceptions. For example, I almost never use "leverage. That's just not how I talk, so it's banned outright. If I didn’t ban it, I’d have to keep removing the word “leverage” every time I generate the first draft of a landing page. The file also covers the obvious AI tells and a few constructions I'd never write myself. Some are obvious — I’ve seen them in AI writing, such as the “It’s not X, it’s Y” sentence structure. I also pulled from Wikipedia’s [Signs of AI Writing](https://en.wikipedia.org/wiki/Wikipedia:Signs%5Fof%5FAI%5Fwriting?ref=blog.annabyang.com). ### My "learnings" file My third file is a self-updating “learnings” file. Claude compares my edits against the original version, whether it’s a social post or the first version of an email. If I’ve corrected the same *type* of thing over and over, Claude writes that observation to the learnings file. ### How the files are applied In actual generation, Claude does the following: 1. Reviews the voice and tone guide 2. Looks at the learnings file to make sure it hasn’t introduced something that I would correct later 3. Generates the content 4. Checks any generation against the hard rules file and makes corrections All three live in [Notion](https://affiliate.notion.so/annabyang?ref=blog.annabyang.com) *\[affiliate link\*\]*, so Claude can read and update them without me copying anything between apps. While the learnings file is self-updating, I’ll occasionally tell Claude to add new things to my hard rules file. Every model version seems to introduce new “favorite words” or phrases that Claude will start to use, so I need to make sure they’re banned. The voice and tone guide is mostly static, since I’ve long ago established how I sound on The Internet. ## How to build your own version Building this system isn't complicated, but you need to spend some time and start with the right raw material. ### 1\. Start with your existing writing Compile a document (or several documents) of your own writing without any AI assistance— emails, old blog posts, etc. The more you have, the better the baseline. A few paragraphs won't be enough. If you don’t have a lot of pre-AI writing — like you’re trying to ramp up your content *with* AI — it’s going to be really hard to get output that doesn’t sound generic. Your AI tool won’t know what **you** sound like. ### 2\. Create your voice and tone guide Have Claude or a similar tool analyze the samples and describe your tone and rhythm back to you. This file describes how you sound in plain terms, the way you might explain your voice to someone who has never read your writing before. Read the result before you start using it, because some of what the tool notices will be right and some of it won't be. I actually have multiple voice and tone files. There are slight variations between my file for social copy and my file for email copy. I even have different voice and tone guides for different social platforms. ### 3\. Write your hard rules file Keep this file short and explicit: specific words and phrases you **never** want to see. Mine started with a handful of entries and grows every time I catch something in output that I would never write. If it’s a “sometimes” or “it depends” — it doesn’t belong in the hard rules file. You’ll need to apply your editorial judgment. ### 4\. Create a self-learning process This process takes a bit of finagling to be fully automated. When Claude generates a landing page for me, for example, it actually generates two copies: one in my working folder that I edit, and a pristine “Claude-only” folder that exists on my Google Drive. Once a week, Claude compares the working versions with my edits to the pristine versions. That’s how it comes up with the learnings. I actually have a whole approval process where Claude presents its learnings to me before updating the learnings file, but that’s an entirely separate blog post. You could start with a voice and tone guide and a hard rules file, and that gets you about 80% of the way there. The learnings cover things that are less prescriptive. ### 5\. Make sure your AI tool checks your files Finally, you’ll make sure that AI is always referencing your files. I use Claude skills so that any content generation follows the same process each time. My voice and tone guide, hard rules, and learnings files are referenced within the skill, with explicit instructions that Claude needs to review the files. ## The human step: editing AI output always goes through me as a final step. No matter how good the three files get, there will always be editorial judgment that AI can’t replicate. And I don’t expect it to. The goal is to cut down on editing, not eliminate it. Most of the time, the drafts (like for social posts) are close, thanks to the voice guide, but not *exactly* how I would say something. AI works well as [a writing partner](https://blog.annabyang.com/artificial-intelligence-editing/), but **partner** is the operative word. Not a replacement. AI output will always be a regurgitation of everything else on The Internet. That’s how the models work: predictive text of what “should” come next. Katie Parrott, staff writer and AI editorial lead at Every, puts it this way on the *AI & I* [podcast](https://open.spotify.com/episode/2P2YrktwHWW4Hy402uAk7J?ref=blog.annabyang.com): > *"What the model already knows is commoditized information. That's not useful or interesting."* What makes writing stand out is having opinions and [a voice that AI can't replicate](https://blog.annabyang.com/ai-writing-perspective-voice/) because it's specific to *you*. ## Update the files continuously These files aren't something you write once. Without continuous updates, the output “drifts” over time, and you're back to making a lot of manual edits. The easiest way to keep them current is to make updating them part of your human editing. When you catch something in the output you never want to see again, add it to the hard rules file. I use Notion for my files because it’s so easy for Claude to make the updates for me. Markdown files (stored locally on a computer) are also a common choice. Google Docs — at least, right now — can’t be updated once they’re created. Start with creating your voice and tone guide this week, because it’s the easiest to create. Pull together a bunch of samples and let your AI tool describe your voice back to you. Now you have a baseline to build the rest of this system. [![CTA Image](https://storage.ghost.io/c/fe/e9/fee9a30d-4c70-4bd2-af95-fe6abc8afd29/content/images/2026/05/ecourse-authentic-online-writing.png)](https://links.annabyang.com/aby-online-writing?ref=blog.annabyang.com) **Not sure where to start with posting on social media?* **These 30 prompts are designed to spark some ideas* [Get the prompts ](https://links.annabyang.com/aby-online-writing?ref=blog.annabyang.com) ## FAQs #### What's the difference between a voice guide and a style guide? A voice and tone guide is descriptive. It’s often full of adjectives that describe a person’s writing. A style guide is more prescriptive and includes grammar preferences, such as using contractions, average sentence length, or using title case versus sentence case. Many people use a single guide that covers both voice and style preferences. #### How do I create a voice guide from my own writing? Compile a few examples of writing you did before you used AI, then have a tool like Claude or ChatGPT analyze it and describe your tone and rhythm. The result is a reference file you can for future AI-assisted writing. #### Can I use the same rules file across different platforms or AI tools? Yes. You can apply the same "writing rules" across all writing that's different from voice and tone guides that are specific to different mediums. A list of explicit bans doesn't change based on the platform, so it can be reused across social posts, email copy, or any other AI-assisted writing.