2026-05-08 · 12 min read

The First Draft Is the Work

Key takeaways

AI can write a cleaner first draft than you can. That is exactly the risk: the messy first draft is where the thinking happens, and a polished one skips you past it.

Prototypes test does this work? Writing answers what should we build, and why? They are different cognitive tools, not substitutes.

The experts who skip writing already did the writing that built their judgment. Junior PMs need the reps.

Use AI hard after the first draft (stress-test, restructure, find gaps), and let it draft low-stakes, well-trodden docs. Keep the strategic first draft yours.

I. The Advice I Keep Hearing

Andrej Karpathy posted a while back about building what he called an "LLM wiki," a system where his workflow is essentially: collect information, let AI synthesize it, iterate on the output. Writing, in the traditional sense, has basically disappeared from this particular process. He's not alone. I keep hearing the same thing in PM discourse both internally and externally: PRDs are dead. Stop writing docs. Build prototypes. Ship faster.

I get it. I build prototypes constantly. I've built multiple versions of AI assistants that automate my daily context assembly, doc generation, and communication outlines. In my personal life I have agents running autonomously overnight pumping out working web apps. On net, these tools have made me materially faster and roughly doubled the volume I ship, while improving the quality of the output at the same time. I'm not arguing against any of that.

But there's something missing from the "just build, and build faster" narrative that worries me, particularly for PMs, especially new-ish PMs.

The "PRDs are dead" mantra is half right, half dangerous. The dangerous half is the implication that writing doesn't matter anymore. That the document, the brief, the strategy doc are relics of a slower era. That thinking happens in your head and prototypes are how you express it.

I don't think that's how thinking works. Maybe for a few. But definitely not for me, and I'm guessing not for most people.

II. Writing Is How You Think

Leslie Lamport, the computer scientist behind LaTeX, put it best: "If you're thinking without writing, you only think you're thinking." That sounds like a clever line. It's actually a claim about cognition. The act of putting words on a page is where you discover what you actually think, as opposed to what you assume you think.

Richard Feynman knew this intuitively. When the historian Charles Weiner called Feynman's notebooks "a record of his day-to-day work," Feynman corrected him: "They aren't a record of my thinking process. They ARE my thinking process." He built an entire learning method around this. Write down the concept. Explain it in plain language. Find where your explanation breaks. Go back and learn that part. The method works because writing exposes the gaps that staying in your head never will. Your inner monologue is too forgiving. It lets you skip the hard parts. It lets you think you understand something when what you've really done is gesture at it.

I experienced this firsthand with a project strategy document earlier this year. When I sat down to write the strategy, I thought I knew what I believed. I'd been talking about it for months. I had the conviction.

Then I started writing, and the conviction fell apart. Not because the idea was wrong, but because I hadn't actually worked through the hard parts. The assumptions I was making about conversion rates contradicted the data I had on free-to-paid timelines. The sequencing I had in my head didn't survive contact with the dependency map. The "obvious" first step turned out to depend on three things that didn't exist yet.

None of that was visible to me until I wrote it down. The writing didn't document my thinking. The writing was the thinking. The messy, contradictory, frustrating first draft is where the strategy actually formed.

I have an operating principle I've held for a while now: never AI-generate the first draft of something strategic. I use AI constantly after that point. Tightening prose, stress-testing logic, restructuring for clarity, finding what I missed. But the initial wrestling? That has to be mine.

III. Two Kinds of Thinking

Prototypes and writing serve fundamentally different cognitive functions. This is the distinction the "PRDs are dead" crowd collapses, and it matters.

A prototype answers: does this work? Does this interaction feel right? Does this flow make sense? A prototype is a hypothesis test. You build something, play with it yourself and with other people, and learn whether your intuition was correct.

Writing answers a different question: what should we build, and why? What problem are we solving? What are we betting on? What are we deliberately not doing? Writing is hypothesis formation. It's where you decide what's worth testing in the first place.

You need both. They are not substitutes for each other.

In early April, I had what I think of as a burst day: twelve PRDs and a set of deployed prototypes for a certain project, all in one session. The PRDs came first. They defined the problem space, the strategic rationale, the success criteria, the assumptions that needed validation. The prototypes came second. They tested whether the solutions were viable, whether the interactions worked, whether the user experience held together.

Without the writing, those prototypes would have been solutions in search of problems. Technically impressive, maybe. Strategically untethered, definitely.

Consider Karpathy's own progression. He coined "vibe coding" in February 2025 (describe what you want, accept all AI output, don't read the diffs) and called it passé twelve months later. Now he talks about "agentic engineering," agents with oversight and scrutiny, no compromise on quality. I think that progression happened because untethered building produces lots of noise. At some point, you have to stop and ask: what am I actually trying to accomplish here? For me, that question gets answered in writing, not in code.

Writing tethers building to purpose. Remove the tether and you're just making stuff.

IV. The Karpathy Problem

Here's what worries me about the current conversation.

Karpathy created a workflow that skips writing. He has decades of deep technical work, years of building intuitions about what matters and what doesn't, a mental model of his field that's been pressure-tested thousands of times. When he says his workflow doesn't involve traditional writing anymore, he's describing the endpoint of a long process of developing judgment. The judgment is already there. It's pretty much subconscious for him at this point.

A 25-year-old PM reading that tweet takes away something very different. The takeaway is: I don't need to write either. I can just collect, synthesize, and let AI do the rest. That's the corrosive version of the message.

It's like telling someone who just started lifting that they don't need free weights because the experienced athlete uses hyper-specific machines. The athlete earned the machines with years of foundational compound work. The junior needs the hard, unassisted reps precisely because they haven't built the foundation yet.

A senior product leader made exactly this point during a product all hands recently, talking about communication as a core PM skill. When you write something yourself, you discover that Argument 2 contradicts Argument 1, that you're missing key data, that your recommendation needs to change. That messy process is the thinking, and AI skips you past it.

Decision filter: before you let AI draft it

Is it strategic, or low-stakes and well-trodden?

Could you defend paragraph 2 if someone senior challenged it?

Did you wrestle the contradictions yourself, or inherit a clean answer?

If it's strategic and you couldn't defend it, the draft isn't done. It hasn't started.

That last question is everything. It's the difference between a PM who has a position and a PM who has a document. If you're a young PM, you need the reps. The documents aren't busywork.

V. Creating vs. Consuming

There's a broader pattern here that extends beyond product management, and Karpathy's LLM wiki actually illustrates it perfectly.

There are two directions for a personal knowledge base. Karpathy's pattern is compiled external knowledge: you curate raw sources (papers, articles, PDFs, transcripts), and AI authors and maintains the synthesis on top of them. The LLM is the writer. You're the librarian pointing it at what to read.

There's an inverse shape. You write your own notes (half-formed thoughts, meeting fragments, architecture sketches, what you actually decided on Tuesday), and AI maintains that. It never authors the first draft. It cross-references, flags contradictions, catches drift, keeps the graph connected. You're the writer. It's the librarian for the thing you wrote.

Take each pattern to its limit. Karpathy's wiki, scaled to infinity, converges toward a personalized copy of the internet: an ever-more-complete synthesis of what you've read. The inverse, scaled to infinity, converges toward a persistent copy of your own mind: what you've actually thought, how you've changed your mind, what you've kept alive. Genuinely different artifacts.

This maps onto what I've built for myself. My AI agent, Pulse, assembles my context every morning before I'm awake: automated daily briefings, weekly intelligence scans, meeting prep documents, experiment health checks. That's the Karpathy direction. Compiled external knowledge, synthesized for retrieval.

But the thinking happens in the other direction. It happens in my daily notes, my strategy docs, my decision logs. The notes I write by hand (literally. Boomer, I know). And critically, it happens in the connections between them.

I've used Obsidian for years, and when I try to explain to people why I love it, I usually lose them. They see a markdown file viewer. Nice, sure. But just a file viewer.

What opened up my thinking about Obsidian, and honestly about note-taking in general, was bidirectional links. That sounds incredibly tactical, but stay with me. A bidirectional link means that when I connect a daily note to a person to a project to a decision, each of those nodes knows about the connection. The web of relationships is visible from any entry point.

The power of this is not in reading. It's in writing. Every time I create a link, I'm making a micro-decision, often fuzzy and instinctual at first, about what relates to what. That act of linking is itself thinking. I'm deciding: this conversation is connected to this project, which is connected to this person's concern, which relates to this strategic bet. AI can suggest connections, but when I write the link myself, I'm the one who has to decide what matters in my specific context. When you're the author, AI must not overwrite your voice. It surfaces the contradiction; you decide whether you changed your mind. It finds the older note that connects; you decide whether the link belongs.

The PKM tradition has a word for this distinction. "Second brain" isn't about having a fast retrieval system for stuff you've read. It's about the act of writing your own words, in your own voice, while the ideas are still percolating. Karpathy's pattern builds a brilliant personalized research index. But the AI is doing the distilling and the expressing. By the standard the PKM world actually uses, only the system where you do the writing qualifies. Both are valuable. Only one of them is the work.

VI. The First Draft Is the Work

So here's the principle.

The first draft of anything strategic must be yours. Not because AI can't produce a competent first draft. It absolutely can. I've seen it. The output is often better structured, more comprehensive, and more polished than what most people would produce on their first pass.

That's precisely the problem.

A polished first draft feels like progress. It feels like you've done the work. But you haven't. You've skipped the part where your half-formed ideas collide with the blank page, where you realize you don't actually know what you think about the second point, where your "obvious" recommendation turns out to depend on an assumption you can't defend. That collision is the work. Everything after it is editing.

After the first draft, AI is extraordinary. "What am I missing in this argument?" is one of the best prompts I use. "Stress-test this logic" is another. "Restructure this for a VP audience" is a third. The tightening, the sharpening, the restructuring: AI does all of that better and faster than I can alone. The output is materially better, not just faster.

But the tools amplify thinking that already happened. They don't replace it.

I'll add one more layer to this. Sometimes I write with a pen. Actual pen, actual paper. And I notice something different happens when I do. The pace forces a different quality of attention. When you can only put down 30 words a minute instead of 90 (or 150-200+ with voice typing), each word carries more weight. You're not generating text. You're choosing words, carefully. It's not productivity advice (it's the opposite of productive). It's a thinking technique. When I really need to work through something, sometimes slowing down is the fastest path to clarity.

VII. Build and Write

I believe PMs should learn to build. AI dissolved the barrier between understanding a problem and making something real, and that's incredibly transformative and empowering. I believe we should automate the scaffolding: context assembly, communication templates, data retrieval. Let AI handle the parts that aren't judgment.

And I believe we need to keep writing.

These are not contradictions. They're a complete picture. The PM who builds prototypes, automates context assembly, and writes their own first drafts is operating at a level that wasn't possible two years ago. They're faster at everything except the part that should be slow. And the slow part (the thinking, the wrestling, the messy first draft) is what makes everything else worth doing.

The rule has edges. AI-first drafting is the right call when the stakes are low or the path is well-trodden: release notes, a status update, a runbook for a problem you've solved ten times, a doc where the thinking is already settled and you just need it on the page. Let AI draft those and edit lightly. The rule is for the strategic first draft, where you don't yet know what you think.

To new and early PMs: The people telling you that writing doesn't matter anymore have already done the writing that made them who they are. Karpathy wrote papers, built curricula, explained concepts in public for decades before he arrived at a workflow without writing. The sharpest communicators I've worked with didn't get that way by letting AI draft their messages.

Don't skip the step they've already completed. The reps are the point. The struggle is a feature, not a bug.

AI gave us the most powerful set of thinking tools in history. Use them. I do, every single day. But the first draft is still yours. That's not a boomer take. That's how thinking works.