Quick Answer
Prompting Claude and training Claude are two completely different things. Prompting is a one-time instruction. Training is building a system Claude carries into every piece of content it writes for you. The system lives inside Claude Projects: a combination of a detailed system prompt, uploaded brand documents, and a set of skill files that tell Claude exactly how to write for your brand. Once it is set up, first drafts stop sounding like AI and start sounding like you.
Who this is for: Marketers using Claude who are still rewriting every first draft because the output sounds generic, corporate, or just not like them.
For three months, I used Claude the same way most marketers do. I opened a new chat. I typed a prompt. I read the output. I rewrote most of it.
The content was technically fine. Correct structure. Reasonable points. But it did not sound like me. It sounded like every other AI article on the internet, and I could not put my finger on exactly why.
Then I realised I was solving the wrong problem. I was trying to write better prompts when what I actually needed was to stop prompting altogether and start training.
What is the difference between prompting Claude and training Claude?
Prompting is a conversation. You ask, Claude answers, and then the conversation ends. The next time you open a new chat, Claude has no memory of what you asked before, how you like to write, or what your brand sounds like. You start from zero every single time.
Training is different. Training means building a persistent system that Claude carries into every piece of content it produces for you. It does not live in the chat window. It lives inside Claude Projects, a feature that lets you store instructions, documents, and context that Claude reads before it writes anything.
The practical difference: a prompted Claude produces a first draft, and you spend 40 minutes rewriting. A trained Claude produces a first draft you spend 10 minutes editing. Same model. Completely different output. The gap is entirely in the setup.
What does a Claude Projects brand voice system actually look like?
A Claude Projects brand voice system has three layers. Each one does a different job.
The first layer is the system prompt. This is the set of standing instructions Claude reads before it does anything. It tells Claude who it is writing for, what the brand sounds like, what it never says, and what the content needs to accomplish. A good system prompt is not three sentences. It is a detailed document that covers tone, vocabulary, sentence structure, what to avoid, and what every piece of content must do.
The second layer is the uploaded documents. These are the files you store inside the Project Knowledge section of Claude Projects. For HowToMarketer, this includes the brand voice guide, the avatar profiles, the keyword list, and the content calendar. Claude reads these before every response. They give Claude the specific context that no prompt instruction can fully capture.
The third layer is skill files. These are individual instruction documents for specific tasks. A skill file for article writing tells Claude exactly how to structure a HowToMarketer article from the opening line to the FAQ section. A skill file for social posts tells Claude the exact Marky AI output format. Claude does not guess. It follows the skill file.
How do you build the system prompt that trains Claude on your voice?
The system prompt is the most important part of the whole setup. Get this right and everything downstream improves. Get it wrong and the uploaded documents cannot save you.
A brand voice system prompt needs to answer five questions clearly.
Who is Claude writing for? Not a broad demographic. A specific person with a specific bad day and a specific thing they need. Marketing Maya, for example: 28 to 38, time-starved, frustrated that her content sounds like every other AI article, needs something she can actually use today.
What does the writing sound like? Give Claude examples of sentences you would write and sentences you would never write. The contrast matters more than the description. Saying "write conversationally" means nothing. Showing Claude a bad opener and a good opener for the same topic is something it can actually use.
What does it never do? List the specific phrases, openers, and patterns Claude must avoid. For HowToMarketer: never open with "In today's digital landscape." Never use the word "leverage." Never write a heading and then take two paragraphs to answer it. These are specific, enforceable rules.
What structure does every piece follow? Claude needs to know the article structure before he starts writing. If you have a 9-step article format, put it in the system prompt. If every article ends with a FAQ section using People Also Ask phrasing, say that explicitly.
What is the single goal of each piece of content? Every article serves a CVJ stage. Claude needs to know which one and what the CTA is before it writes the first sentence.
Which documents should you upload to Claude Projects for brand voice training?
Upload documents that give Claude context it cannot infer from the system prompt alone. The system prompt tells Claude the rules. The uploaded documents give Claude the evidence.
The brand voice guide is the most important upload. It should include real writing samples: a paragraph Claude should try to sound like and a paragraph it should never produce. Examples do more work than instructions.
The avatar profile comes second. Claude writes better when he knows exactly who it is writing for. A detailed avatar with a specific bad day, specific fears, and specific vocabulary does more for tone and relevance than any prompt instruction about "knowing your audience."
Past articles that represent your best work belong in Project Knowledge, too. Not every article. Two or three that you consider the clearest expression of your voice. Claude will use them as reference points without you having to explain what you mean by "peer-to-peer tone."
What you do not need to upload: everything. More documents are not always better. Claude has a context window limit, and if Project Knowledge is full of loosely related files, the most important ones get diluted. Keep it tight. Five well-chosen documents beat fifteen average ones.
What are skill files, and why do they change how Claude writes?
Skill files are the part most marketers skip because they take an extra hour to build. They are also the part that makes the biggest difference.
A skill file is a dedicated instruction document for one specific content task. The article writer skill tells Claude the exact 9-step structure for a HowToMarketer article, the keyword filter it must run before writing, the Before and After Grid it must map before writing the hook, and the Medium companion brief it must produce at the end. Claude does not have to remember any of this. It reads the skill file and follows it.
The result is consistency. Every article Claude produces from that skill file has the same structure, the same GEO-optimised formatting, the same FAQ section, and the same CTA placement. Not because Claude is guessing at what you want, but because the skill file removes the guesswork entirely.
How long does it take before Claude actually sounds like you?
The honest answer: the system prompt and the brand voice guide do most of the work on day one. If those two are detailed and specific, the first article Claude produces inside the Project will be closer to your voice than anything you got from three months of prompting.
What improves over time is the skill files. The first version of a skill file is never perfect. You use it, you notice what Claude still gets wrong, and you update the skill file. After three or four iterations, the output is tight enough that editing a first draft takes minutes rather than hours.
The marker that the system is working: you stop noticing that Claude wrote it. The content reads like something you would have written on a good day, structured better than you would have organised it yourself, and formatted for Google and AI Overviews without you having to think about it.
What did not work the way I expected
Uploading too many documents made things worse before they got better. The first version of this Project had eleven files in Project Knowledge. Claude was reading so much context that the most important instructions got buried. Cutting it down to five core files and putting the rest in skill files made an immediate difference.
The bottom line
If your Claude first drafts still sound like AI, the problem is not Claude. It is the absence of a training system. The model is capable of writing in your voice. It needs the system prompt, the brand documents, and the skill files to do it. Build those three things once inside Claude Projects and the rewriting stops.
Want the full Claude Projects setup we use at HowToMarketer?
Join the email list and get the system prompt template, the brand voice document structure, and the skill file format we use to produce every article on this site.
→ howtomarketer.com/subscribe
