Your MSD Is Not a Document. It's Robot Food.

I’ve been noodling on something for a while now, and I think I’m finally ready to put a stake in the ground.

Your Message Source Document (MSD), the positioning doc, MPF, whatever you call it, is the most important artifact in a PMM’s toolkit. Every product marketing practitioner knows it’s the source of truth everything else pulls from: the pitch deck, the battlecard, the sales discovery guide, one-pager, the website content … pretty much every downstream asset.

But until very recently, MSDs have been written by humans and designed for humans to read … and that’s becoming an increasingly obvious problem as we adopt AI. “Human-readable” is now an after thought, an output should that actually be needed.

What we need now, is a standardized, codified, markdown file that can be used and maintained by AI, with human oversight of course.

The blank page tax

Every time you start a new deliverable from scratch, you pay a time penalty.

Not just the time it takes to write from nothing, but the time it takes to re-orient things. To remember what you decided about positioning six months ago. To dig up the competitive notes from that assessment nobody’s touched since last quarter. To reconcile the three slightly different versions of your value proposition that are floating around in email threads that shifted as you learned.

That re-orientation cost is invisible because it doesn’t show up on a project plan. But it’s real, and it compounds. Every deck starts with a detour. Every battlecard starts with a conversation. Every launch starts with “wait, where is the ICP again?”

The fix isn’t more discipline, most PMM’s have that already, what’s needed it a more efficient and effective way to build and maintain.

The Three Layer Model

Here’s the approach I’ve started to feel is durable for an AI-native PMM:

Layer 1: Source documents: your .md files. ICP, positioning, messaging, competitive landscape. These are the codified inputs in robot format, your AI tools use them efficiently, while humans own and maintain them. This is your foundational single source of truth.

Layer 2: Working templates: structured overlays with AI and human aspects included. these are not blank templates, they are opinionated starting points that are tagged as to which sections need human judgment and which ones AI can draft.

Layer 3: Final deliverables: the outputs in human consumable format … pitch deck, datasheet, battlecard, sales one-pager. These are generated from Layers 1 and 2, not built from scratch. The PMM does final-mile review to ensure a human signs off on emotional resonance. The robots carried the execution heavy lifting.

The lightbulb moment for me was when I realized that the solve to Layer 3 IS Layer 1. Everything downstream from the baseline source is just those critical items like the MSD rendered for a different audience. This is what we’ve said in PMM for a long time, but the shift with AI that is enabled is to codify that process for faster and more efficient execution.

If your MSD is weak, everything downstream is weak. If it’s strong and AI-readable… you’ve got a translater and compiler, not just a document.

What “AI-readable” actually means

An AI-native MSD looks different from a traditional one. Not because the content changes, the core elements are exactly the same: ICP, positioning, messaging, competitive landscape, objection handling and so on. But the structure is designed to do double duty.

A few things that make a document genuinely useful as robot food:

A metadata block. A short YAML header the AI reads first … product stage, audience, tone register, and critically: a do_not_claim list. Claims that aren’t validated yet. Things you’ve decided AI should never say on your behalf until you’ve verified and tested them. (Zero-trust fact-checking, baked into the document itself.)

Confidence ratings. Not every section is equally validated. Your ICP might be HIGH confidence after 50 customer calls. Your messaging for segment 2 might be LOW because it’s a hypothesis you’re still testing. Marking this explicitly means AI doesn’t treat a hunch the same as a hard-won insight.

Human and AI annotation zones. Some sections AI can draft from. Some require human judgment before AI touches them. The template makes this explicit:

<!-- AI: Use these proof points for battlecard generation -->
<!-- HUMAN: Validate this claim before using in paid contexts -->

Structured competitive data. Not long form prose paragraphs about competitors, crisp and clear decision trees. What they do well, where they fall short, our wedge into a competitors stronghold. Formatted so AI can reason over it, not just summarize it.

Multiple message registers. Your value prop written three ways: (by way of an example) for an exec (outcome-focused, short attention span), for a practitioner (peer-level, very specific), for an aspiring user (empowerment-focused). AI picks the right register for the output it’s generating, all you did was write it once.

None of this is technically complex, but it is very deliberately intentional. You’re writing for two audiences now, the human who needs to understand it and the AI that needs to use it. Turns out, what’s good for the AI is often clearer for the human too. Maybe this Humans+AI thing has legs.

Meet AgenticPMM OS

This has all been theoretical, so in the spirit of “seeing is believing”, I built a sample to show what this looks like in practice.

Using a fake product I just made up, meet the AgenticPMM OS, an AI-native PMM workflow system for founders and solo PMMs. (Yes, I’m using a fake PMM tool to explain … how to do PMM.)

This link … agenticpmm-os-msd.md is a fully populated AI-native MSD with sample data. All the annotation zones, the metadata block, the confidence ratings, the competitive structure, the objection table. This is an example of what Layer 1 looks like when it’s built to be robot food.

You can use this two ways:

Option A: Prompt-driven (no special tools needed): Drop the MSD into a Claude conversation and use this prompt:

You are a PMM working from this MSD. Please generate two outputs:

**Output 1: Word document MSD**
Format the MSD as a clean, human-readable Word document with the following sections: Executive Summary, Product Overview, Ideal Customer Profile, Positioning Statement, Messaging Framework (with exec, practitioner, and aspirational registers), Competitive Landscape, Objection Handling, GTM Motion, and Open Questions. Remove all HTML annotation comments from the output — this version is for stakeholders, not AI. Flag any section marked LOW confidence with a note that it needs human review before distribution.

**Output 2: Pitch deck outline**
Generate a 10-slide pitch deck using the exec messaging register. Slides: (1) Problem — the blank page tax, (2) Who feels it — ICP in one sentence, (3) The insight — MSD as compiler not document, (4) Solution overview — what AgenticPMM OS is, (5) How it works — the three-layer model, (6) The Human + AI division of labor, (7) What you get — the deliverables it generates, (8) Proof point or sample output, (9) Who it's for — ICP detail, (10) Call to action. For each slide include: headline, 2-3 supporting bullets, and a speaker note. Flag any slide that draws from a LOW confidence section of the MSD.

Note: what you get back is a structured text outline, the raw material for a deck, not a designed one. Of course you’d take this into your brand template and make it look like pretty. That’s the point. AI gives you the thinking; you (or your designer) bring the polish. You can see a web-based demo of what this looks like here →

Option B: Turn it into a skill: If you’re using Claude in Cowork mode and want this to run automatically against any MSD you drop in, you can build it into a reusable skill. Just paste this into Claude:

I want to turn this MSD-to-deliverables workflow into a reusable Claude skill. The skill should: read an MSD .md file from a specified folder, respect the <!-- AI: --> and <!-- HUMAN: --> annotation zones, check the metadata block for do_not_claim items, and offer to generate either a formatted Word document MSD or a 10-slide pitch deck outline. Walk me step by step through creating this as a Cowork skill I can reuse.

That’s it. Claude will walk you through the rest. You’ve just turned a workflow into infrastructure.

Why this matters … today!

PMMs who figure out how to use AI systematically become force multipliers. PMMs who use it ad-hoc, pasting random prompts into ChatGPT, getting plausible-but-inconsistent output, are just adding noise to their workflow.

The difference isn’t the AI (although that helps), it’s the structure YOU bring to it.

An MSD that’s also robot food is that structure. It’s the thing that makes every downstream AI interaction smarter, faster, and more consistent than the last one, and it’s how your strategy compounds instead of staying trapped in a doc nobody reads.

Download the template. Swap in your product. See what falls out.

I’ll be honest with you: it might look a little weird at first. The YAML header, the HTML comments, the confidence ratings … it doesn’t look like a “normal” document. That’s on purpose. This isn’t just an MSD, it’s a new instruction set.

And instruction sets, in my experience, tend to actually get followed.

Adam