Here’s something I’ve learned. In the beginning when I was first using AI tools for content and writing, I’d start every session from scratch. Set the context, explain who I am, what I want the tool to be, who I’m writing for, what tone I want, what I won’t tolerate. Every … Single … Time.

There’s a better way. A much better way.

It’s called a CLAUDE.md file (if you’re using Claude, but the concept applies to any AI tool you use consistently … CoPilot has a file named copilot-instructions.md and ChatGPT uses custom instructions, but essentially the same thing … more on this below). Think of it as a persistent brief, a single Markdown file that lives in your project folder and tells your AI everything it needs to know about you, your audience, your voice, and your work … before you’ve typed a single prompt.

For a PMM, this is almost embarrassingly natural. We spend our careers building baseline documents, defining audiences, and documenting our tone and voice. Building one to use both yourself as a pmm and consistently as a business is the same muscle.

Here’s what I’ve got in mine. And I continue to expand it as I learn.

Side note: this is about what I have for PMM Kickstarter, I also have this in my day job, similar but different, so some of the below has already be road tested on applicability and ease of transferability :)

1. Who I Am

Start with the basics. Your name, your role, your context. Not a bio, a brief. Two or three sentences that capture what you do and why it matters.

I include my contact info too. Simple, but useful when AI is drafting content that might reference how to reach me.

The goal is for any AI reading this to immediately understand the lens through which everything else should be interpreted. Are you a founder doing your own PMM? A senior PMM at a growing company? A consultant advising? Each of those is a different posture, a different authority level, a different set of assumptions.

Be specific. “I’m a PMM” is not useful. “I’m a PMM who has spent 15 years at technical B2B companies and now advises early-stage founders on their first product marketing hire” is.

2. My Audience

This one should feel familiar. You know how to write an ICP document. Do the same thing here, but for your readers.

Who are you writing for? What do they already know? What are they trying to solve? What would make them stop reading versus lean in?

I define four audience segments in mine, from early-stage founders doing their own PMM through to PMM leaders integrating AI into their teams. Each one reads with different context and different needs. The AI can’t calibrate tone and depth without knowing this.

One line that lives in my file: “My readers are practitioners, not academics. They want actionable, real-world guidance, not theory.” That single sentence shapes every draft in a way that pages of instructions can’t. (and yes, while I write a lot of this content myself including proof reading everything and old skool typing like the old man I am, I do also use the AI to draft things to get me going and certainly to pressure test my content against both positive and negative voices).

3. What I Cover (and What I Don’t)

Your topic territory. The themes and subject areas that are yours to write about. This helps the AI stay in your lane rather than drifting toward generic content.

For me this is product marketing fundamentals, team building, AI in PMM, and career development. It’s also implicitly a list of what I don’t cover … I’m not writing about demand gen tactics, PR strategy, or general startup advice, even though those topics overlap with my world.

The boundary matters to me as much as the content. I don’t run this site as a generic top-of-mind brain splurge. It’s for a particular audience, most in a particular space.

4. My Voice and Style

This is the most important section, and the one most people either skip entirely or fill with vague adjectives (“conversational,” “professional,” “friendly”). Vague adjectives produce vague output.

Be specific and be honest. What does your voice actually do?

Mine includes things like:

These are not style preferences, they are my rules. The more specific your rules, the closer the AI output gets to actually sounding like you. And it gets scarily realistic.

Also document what you won’t do. My file explicitly bans words like “leverage,” “synergy,” “unlock,” and “game-changing.” If you’ve ever read a piece of AI content and winced … those winces belong in this section.

5. My POV on AI

If you write about AI at all, document your actual philosophy here. Not the generic “AI is a tool” which is a total non-take, but your real and specific point of view.

Mine is built around the “Humans + AI” framing, the belief that AI handles the 80% execution and humans provide the 20% that actually matters: intuition, empathy, emotional resonance, and context. I document that AI is a “vibe engine,” not a database. I include the Zero-Trust fact-checking rule.

This matters because AI will default to reflecting the dominant narrative in its training data, which right now swings between breathless AI optimism and existential panic. Your file anchors it to your actual POV.

6. My Vocabulary and Frameworks

Every PMM develops a set of recurring terms over time. The phrases that have become shorthand for the things you believe. Document them.

Mine includes “connective tissue” for PMM’s role between functions, “execution tax” for the overhead burden AI can reduce, “multiplier” for what AI becomes in skilled hands. These aren’t jargon to me, they’re my personal vocabulary. When the AI uses them naturally in a draft, it sounds like me. When it doesn’t, it sounds like a competent stranger attempting to mimic me.

7. Content Format Specs

If you produce content in more than one format (long-form posts, LinkedIn, email newsletters, internal briefs), give each format its own section with specific rules.

I have separate specs for LinkedIn posts versus long-form articles. Although TBH I never use this one, I write whatever feels right at the time I post, I’m sharing information to my network, not trying to game clicks and engagement. LinkedIn: 50-150 words, flowing prose only, link in the post body, no sign-off. Long-form: 600-1500 words, start with a relatable observation not background context, no horizontal rule dividers, always sign off with “Adam.”

The AI cannot guess what “write me a LinkedIn post” means to you. You have to tell it.

8. Standing Instructions

Close with the explicit assumptions you want the AI to carry into every writing request. The things you’d otherwise have to type at the start of every prompt.

“Always write for practitioners, not academics.” “Don’t hedge, take a point of view.” “When in doubt, go shorter and sharper.”

These become defaults. You only override them when you have a specific reason to.

9. Not Using Claude? Here’s Your Version

The concept is the same regardless of which tool you use day to day. The implementation is slightly different. Info is correct as far as I can find on The Internet, let me know if you read this and it’s incorrect or changed :)

ChatGPT gives you a few options, and they’re not all equal. Custom Instructions (Settings, Personalization, Custom Instructions) is the basic version, two text boxes where you describe yourself and how you want the tool to respond. It’s limited, but it’s a start. ChatGPT Projects are the closer match, you can attach files as persistent context, which gets you to something that behaves much like a CLAUDE.md. Custom GPTs are the most powerful option, essentially a configured agent built around your brief, but they take more setup and are probably overkill for most PMMs who just want better writing output.

GitHub Copilot has the most direct equivalent. Create a file at .github/copilot-instructions.md in your repo. Same Markdown format, same concept, drops right in. If you’re already using GitHub to manage content or briefs (more common than you’d think), this is the clearest path.

Microsoft 365 Copilot doesn’t have a native file-based equivalent yet. The closest thing is Copilot Studio, where you can build custom agents with defined personas and knowledge bases. It’s more involved than dropping a Markdown file in a folder … which is a polite way of saying it’s a lot more involved. The underlying principle is the same, but the path there isn’t as clean.

This concept has a name in AI commentary: “system prompt” or “persistent context.” All the major tools are moving toward supporting this in some form. Some have arrived with more elegance than others.

In any case, the point is that whichever tool you use, the brief you’d write is the same. The sections above apply whether you’re talking to Claude, ChatGPT, or Copilot. The format and filename change. The thinking doesn’t.

10. It’s A Living Document

One last thing: this file is not a one-time exercise. It’s a living brief that gets better every time you use it.

The workflow that’s worked for me: whenever I review a piece of AI-generated content and make significant edits, I ask both myself and the AI whether those edits reveal something the AI didn’t know about my style. If yes, it goes in the file.

Over time, the gap between first draft and final draft gets smaller. That’s the compounding return on building this properly.

The AI gets smarter about you. You spend less time fixing. The output gets closer to your voice with every iteration.

For a PMM, this is just good leverage of tools to aid in getting your work done. We do it for products, for campaigns, for launches. Do it for yourself too.

Now Actually Build It

Reading about this is the easy part. Here’s a prompt you can use to get your first version built in one sitting. Copy it, drop it into Claude, and let it interview you.

I want to build a CLAUDE.md file to use as a persistent context brief for all my AI writing work. Please interview me to gather everything you need, one section at a time. Cover: who I am and my role, my audience and what they're trying to solve, the topics I write about (and don't), my voice and style including specific rules and things I won't do, my POV on AI if relevant to my work, any recurring terms or frameworks I use, the content formats I write in with specific rules for each, and any standing instructions I want applied to every writing request. Ask follow-up questions if my answers are vague. When we've covered everything, write the complete CLAUDE.md file in Markdown, ready to save.

If you’re using Cowork, I’ve also packaged this as a downloadable skill that runs the interview automatically. Download the CLAUDE.md Builder skill. As a good security professional, please review the text file before installing (a skill file is just a ZIP archive, rename it and you can see).

The interview will take 10-15 minutes, less if you have the answers, more if you’re thinking about it as you go. The file you get at the end will be rough in places, but it’ll be yours. Edit it, live with it for a few iterations, update it when something doesn’t sound right, and prompt Claude to compare what was generated with the edits you made to learn from that, and then ask the AI to update the claude.md file to adjust for what was learned.

The first version is never the best version. It just needs to exist.

Adam