I recently shipped Management Craft, a collection of management frameworks distilled into concise “cards,” each drawn from my work coaching startup founders. The response was generous. The cards were almost secondary. What I’d spent six months building was the process I used to make them.

I want to show you that process. All of it. The tools, the editorial system, the weird iterative path that got me here. Not because my version is the one you should copy, but because I think building your own version might change how you write, how you think, and how much fun you have doing both.

Some of you are going to think this is Very Cool™. Some of you will go looking for pitchforks and torches.

Here’s the short version: I dictate. A lot. 320,000 words in the last ten weeks alone a lot. Claude drafts. Then I run that draft through the distilled principles of my four favorite writers on writing, one at a time. And Claude shows me what each of them might change and why. I read every recommendation and I feel like I’m sitting at a table with them. The output of “What might Bill Zinsser say about this?” makes me feel like a little kid witnessing a magic trick for the first time, every time. After I decide what to do with each proposed change, Claude composes a second draft. And from there, I rewrite the piece heavily until I feel like the piece conveys what I hoped to convey from the start.

Here’s the long version.

Talking Is Thinking

I need to back up. Before Management Craft, before the editorial system, before any of this, I, like many of you, spent about a year beginning to learn with LLMs. Not building anything. Learning. I uploaded a dozen psychology textbooks to Claude projects and asked question after question. I worked my way through Bertrand Russell’s A History of Western Philosophy, peppering ChatGPT with dozens of “dumb question” after each chapter. Each time I returned, it felt more and more like the era of everyone having their own digital Aristotle had arrived.

For a long time, I click-clacked away at my keyboard, waited for Ser Chat to respond, and I’d click-clack some more writing down notes in Notion as I learned and learned.

Then a friend suggested I try dictating. Like everybody else, I’d used ChatGPT’s chat function, the one where you talked and it would respond with Not-Scarlett-Johansson’s voice. But he was suggesting something else. He said he’d dictated hundreds of thousands of words in the last month. So I tried dictating journal entries instead of typing them. And now, after making coffee with my wife in the morning, I tell her, “Okay, I’m off to go talk to Claude all day.” And I do.

And while I’ve broken up with Mr. Altman’s Chat, I now talk to Claude about literally everything. Before, my thinking toolkit as a business-of-one was limited to journaling, writing many essays that never saw the light of day, and talking with my friends and family. I didn’t expect dictating to be much more than another neat working habit that might make me a bit more productive. Instead, it has completely changed the way I work, think, and create.

Here’s why. When you write, you’re doing two things at once: creating and judging. That’s Betty Flowers’ insight.¹ You put a sentence on the screen and immediately start evaluating it. Is it good? Is it clear? Should I restructure? Delete. Delete. The internal editor is always on. Delete. Dictation, on the other hand, could not be more different. I actually put headphones on, connected to my phone, volume on loud, and then think out to the microphone on my computer. A person would call it “word vomit.” But thankfully, Claude isn’t a person. Yet. Yet?

Unencumbered by seeing my own thoughts, I’m freed from judging my thoughts. At least for now. And in this flow state, creation now happens sans brakes. And thinking without brakes is just pure goddamned fun.

C.S. Lewis told a young girl named Thomasine in 1956: “Always write (and read) with the ear, not the eye.”² The best prose sounds like speech. Writers have been trying to achieve that for centuries. Dictation just lets you start there.

A Wonderful Chain

Okay so there I was with these dictations. Raw, messy, full of my actual thinking. Big walls of text, line breaks be damned. And I started wondering. What if I put my large amounts of language in these large language models? It felt like such a silly thought, and it is, but really, like, what might happen?

My first instinct: what might Claude do with my dictations after I uploaded a set of psychology textbooks to a project? Roughly what you’d expect, but I was surprised at how thoughtful it was. Interesting.

I’d also been working on a few essays. Essays I never published. I wondered, one day, “What might Bill Zinsser or Mr. McPhee say about this part?” I started asking Claude just that, and it was like they’d teleported in or something, suddenly available and ready to work through the night with me. This wasn’t “have Expert Writing Man rewrite my essay for me and publish it,” instead I was coming from a place of curiosity and frustration as a writer. The output was, insanely helpful.

The next day I loaded up a combination of eight books and essays to a project in Claude only to be met with the disappointment of maxing out the file limit. Shit.

So I did what any god-fearing American would in 2025 and I asked Claude if there was a way to fix that. And there was, it turns out. LLMs don’t like PDFs, you see. They like Markdown. So I figured out how to convert all the PDFs to markdown, celebrated with the thunderous applause of a file limit bar only half filled.

And the next time I consulted Claude on writing, it felt like I had drunk a can of radioactive gatorade. I had acquired superpowers. Perhaps I had forsaken my soul. The radioactive sludge too far consumed. There was no going back.

I realized that when I sat down to write, I was trying to think, structure, and edit simultaneously. Dictation let me separate those. Get the thinking out first. Then let the tools help with structure. Then edit.

I wondered if Claude could help me make an outline from my dictation. It could.

I wondered if it could produce a first draft I could work with. It could.

I wondered if I could load my favorite books on writing into a project and have Claude edit through the lens of those writers’ principles. It could.

Each question led to the next. None of this was planned. I didn’t sit down at a chalkboard and architect a system. I just kept asking “I wonder if it could do that” and finding out that, more often than not, it could. Then I’d iterate on the output until it got better.

That’s how the content for Management Craft got made.

My next “I wonder,” moment is right now, wondering what might happen if I shared my process with other people. So, here it is.

The Full Pipeline

Here’s everything, start to finish.

Step 1: Dictate

I talk. Stream-of-consciousness rounds, sometimes ten minutes, sometimes thirty. Or three hours. I dictate using Wispr Flow, which transcribes in real time. I used the native record feature in ChatGPT, but after three times of losing ten minutes of dictation I was done putting my hand on that hot stove. The dictations go into a database in Notion (the Raw Dictation Archive) so nothing gets lost.

Dictation isn’t just talking. Peter Drucker wrote in “Managing Oneself”³ that people learn differently: some by writing (Churchill), some by taking notes (Beethoven), some by doing, and some by hearing themselves talk (Drucker himself). He described a CEO who called his senior staff into his office weekly and talked at them for hours, not for their input, but because talking was how he processed. Drucker’s verdict: “Learning through talking is by no means an unusual method.”

And yet, the tech world went the other direction. Amazon and Stripe built cultures around the written memo, and a generation of startup founders internalized the idea that clear thinking means clear writing. If you couldn’t produce a six-pager, something was wrong with your thinking. But maybe nothing was wrong with your thinking. Maybe you just needed to talk, and nobody was available to listen for three hours. An LLM is.

The most important thing about dictation is that you feel done at some point. Like there’s a gas tank in your head full of content on whatever it is you’re dictating on. You must always empty that gas tank.

Step 2: The Scaffold

Claude reads the dictation and produces a first draft. I call this the scaffold (thank you, Ross, Bruner and Wood)⁴. It’s far from what I’ll eventually publish. It is the starting material I’ll react to.

What Claude does here isn’t generating structure from nothing. It’s translating. My oral structure (the sequencing, emphasis, and connections I made while speaking) gets translated into written structure (paragraph order, section breaks, the architecture of a prose essay). The thinking was in the dictation. The translation is the craft of converting what I said into something that works on the page.

The scaffold is the start, not the end.

Step 3: “The McPhee Loop”

This is the core of the system, the part I’m most proud of, and the part I have the most fun with.

I call it the McPhee Loop because John McPhee, the legendary New Yorker staff writer, is the one who taught me, through his writing, that structure is a creative act, not a mechanical one. Most writing advice treats structure as the outline you make before you write. McPhee treats it as the thing you discover inside your material after you’ve written it. That’s why he goes first (to find the structure hiding in the scaffold) and last (to make sure the structure survived four rounds of editing that might have broken it). He’s the bookends. Everything else happens in between.

Everything after this phase is written linearly, as if each step follows one after another, but in practice I’ll run or re-run steps in isolation. Finishing Step 6 often leads to running Step 3 and 4 again. By the time I’m ready to publish, I’ve been there and back quite a few times.

Five rounds of editing, each through a different craft lens. Each round loads a “condensed reference” from a writer on writing and asks: “If this author were reviewing this draft, what would they recommend we change?” Mentally, I eagerly sit cross legged and smile as I read through recommendations.

Round 1: McPhee

John McPhee, from Draft No. 4. McPhee has been writing for The New Yorker for over sixty years. His obsession is structure: the shape of a piece, the order of its parts, the way a lead promises something the rest must deliver. This round asks: is the structure right for this material? Does each section make the next one feel inevitable? Does the opening lead with the hook, not the thesis?

Round 2: Minto

Barbara Minto, from The Pyramid Principle. Minto spent decades at McKinsey teaching consultants to think clearly on paper. Her lens is logical architecture: does the argument hold together? Does every section answer a question the reader is already asking? Are there assertions that sound good but say nothing specific? She’s the one who catches you being vague and feeling smart about it.

Round 3: Zinsser

William Zinsser, from On Writing Well. Zinsser’s book has been the standard text on nonfiction writing for fifty years. His doctrine is radical simplicity: every word must do work, every sentence must be the simplest version of itself that preserves the meaning. Is there clutter? Are verbs active and specific? Does it sound like a person talking or like writing?

Round 4: Forsyth

Mark Forsyth, from The Elements of Eloquence. Forsyth is the most surgical of the four. He doesn’t touch structure or logic or clutter. He finds two or three moments where a rhetorical figure would make the difference between a sentence someone reads and a sentence someone remembers. A tricolon, an antithesis, a periodic sentence that holds its payoff to the end. His argument: Shakespeare wasn’t a genius receiving lines from angels. He learned techniques and practiced them. Eloquence is craft, not gift.

Round 5: McPhee again

After four rounds of line-level editing, has the structure held? Tightening sentences always risks breaking the arc. A section that bridged two ideas gets compressed into an assertion. A transition that created momentum gets edited into a flat connector. This round catches that.

Take a look for yourself. Hit the button below and watch one round run on a draft of my next post.

How the Condensed References Work

You can’t just upload a full book and get good results. I tried. The context window means not all the content stays active, and a 60,000-word book contains a lot that isn’t relevant to the specific editorial task.

To make each book something Claude could work with, I bought each book once on Amazon, found Amazon’s DRM too hard to break, bought the books again on ebooks.com. I then used ePubor Ultimate, broke the DRM, converted the books to PDF, and then converted them to Markdown using BookConvert. Did I mention I bought them all in print years ago?

From there, Claude created a “condensed reference” for each: a 1,500-to-2,000-word document that captures the principles, rules, and techniques from that writer as they apply to my work. McPhee’s condensed reference covers structure, leads, greening (cutting without a scar), juxtaposition, and selection. Zinsser’s covers clutter, verbs, unity, the lead and the ending, and rhythm.

Each step solved a specific problem I hit at the previous step.

If you want to build your own version: buy the books, convert them to a format Claude can read, work with them for a while, then distill. I wrote a step-by-step guide for this process. The distillation itself teaches you the book better than reading it did. Distilling McPhee, for instance, forced me to decide which of his principles actually mattered for my editorial workflow versus which ones were fascinating but irrelevant. That act of selection was itself a structural exercise.

Step 4: The Whetstone

After the McPhee Loop, three more passes:

Brevity: Cut ruthlessly. Kill throat-clearing, kill sentences that restate the previous sentence, kill setup paragraphs. But protect the conversational moments that make the voice distinctive.

Flow: Repair the connective tissue the Brevity Pass cut. Check paragraph endings as springboards. Stitch disconnected short paragraphs.

Voice: Four tests. Does it sound like something I’d say in a coaching session? Can you trace phrases back to the original dictation? Could a machine have generated this from a generic prompt? Would I say this to a friend over a drink?

Step 5: The Voice Profile

This one is unusual. I built a document that describes how I sound. Not how I want to sound. How I actually sound, based on analysis of dozens of my own dictations.

It codifies: I lead with the problem, not the framework. I use analogies from everyday life (buying a house, cooking chickens). I use profanity sparingly but naturally. I alternate between short punches and long sentences. I use “you” not “founders.” I say “a lot of” not “numerous.”

The voice profile gets loaded into every session. It’s the check against the system smoothing my voice into generic good writing. It also catches something subtler: when the output is technically excellent but sounds like The Economist. Whether the McPhee Loop drifted there or I wrote it myself, the voice profile catches it.

Step 6: Write

This is where I go through and rewrite. The scaffold gives me something to react to. My internal monologue is often, “that’s not right.” “Dumb.” “I would not say that.” “Deleting this.” Here, I rewrite entire sections, add my own sentences. I cut things that are technically correct but don’t feel like me. Add the humor (all the humor), the asides, the specific details from my life that no machine could produce.

This is where the piece stops sounding like “sounds like Andy” and starts being Andy. Radioactive Andy?

Notice how the draft’s academic framing (Bruner, Vygotsky, misattribution corrections) becomes a single clean sentence with a cooking analogy that makes the concept stick. Warm highlights were cut. Blue highlights are new.

The system exists to get me to this point as efficiently as possible. Everything before Step 6 is scaffolding. Step 6 is where I smith the words.

The Apprenticeship

But now for my favorite part. And the reason I think this is worth sharing beyond “look at my workflow.”

I designed the McPhee Loop to print out each step’s analysis for me to read. I didn’t want a black box editorial process. I wanted to see the reasoning. This way, I get to see Claude channeling Zinsser’s principles and explaining why a sentence should be shorter. I get McPhee’s structural logic applied to my material with reasoning I can follow. I get Forsyth pointing out where a tricolon would land harder than the four-item list I wrote.

It’s like sitting at a table with four writers, absorbing craft by osmosis. An apprenticeship to multiple people’s ideas all at once.

An editor tells you “don’t bury the lede.” Then tells you again. Then tells you a third time. What I’ve always wanted from that experience is: show me what it looks like when you actually fix it. Writing isn’t like cooking, where you can stand in the kitchen next to the chef and watch them fold the dough. You can’t watch a writer write.

The McPhee Loop shows you the move happening. It diagnoses problems in your text, explains why they’re problems, and shows the fix. That’s the equivalent of watching the chef fold the dough. You’re seeing the mechanism, not just the rule.

Jerome Bruner coined the term scaffolding in 1976:⁴ the idea that there’s a zone between what you can do alone and what you can’t do even with help, and the right assistance lets you operate in that zone. A recipe from a wonderful chef lets you cook a dish you couldn’t have invented. You follow it, you learn from it, then you start adding your own touches.

I don’t have a writing coach sitting with me every day. But I have a partner that can channel McPhee, Minto, Zinsser, and Forsyth. Not replacing the books. Making it possible to interact with their ideas at a volume and frequency no traditional path allows.

And it costs $20 to $200 a month. A skilled developmental editor costs $2,000 to $5,000 per piece. That’s a gate. And behind that gate, most writing dies.

Here’s the part I didn’t expect: the product that this process makes is itself scaffolding. Every card in Management Craft is a condensed framework designed to help a founder think through a problem they haven’t mastered yet. The same way McPhee’s condensed reference scaffolds my writing, a card on hiring scaffolds a founder’s first attempt at building a structured interview process. The same way a recipe lets you cook a dish you couldn’t have invented, a card lets you run a process you couldn’t have designed from scratch.

Scaffolding all the way down. The books scaffold the editorial system. The editorial system scaffolds the writing. The writing produces cards that scaffold the reader. I didn’t plan that. It emerged from the process, and I didn’t see it until I was deep inside it.

What I Want You to Take From This

Build your own version. Not mine. Yours.

Pick the writers whose principles matter to your work. Distill them. Play with them. Build the loop. Start with dictation if you’re a talker, or start with rough drafts if you’re a writer. The point isn’t the specific tools. The point is that the process of building the process teaches you more about your craft than any single piece of writing will.

Every step of what I showed you happened through iteration. There’s a meta-process underneath the process. The McPhee Loop didn’t exist six months ago. The condensed references didn’t exist four months ago. The voice profile didn’t exist two months ago. Each one solved a problem I encountered in the previous version of the system.

Your version will look different from mine. It should. That’s the whole point.

To that end, I open-sourced the routing layer that ties all of this together as a skill called /word-smithy. It ships with guides on how to condense your own books, write your own protocols, and build your own voice profile. The skill is the engine. What you put into it is yours.

npx skills add andysparks/word-smithy

Supports Claude Code, Codex, Cursor, and more.

One More Thing

I think part of what we love about handcrafted things, whether they’re tables or essays, is the thoughtfulness, the creativity, and the effort. I suspect part of what’s going to piss people off about this is that it might feel like a shortcut or a way to pass the machine’s work off as my own. But I think, and I’m biased, there’s quite a bit of creativity in building a system like this for myself. And that creativity certainly has led me to make things I wouldn’t have otherwise.

There’s more to say about what this process taught me. Not about the tools, but about what it means to think alongside a machine. About cognitive surrender and why the fear of it is both valid and incomplete. About what it means to be one thinking species that might not be the only one for much longer.

That’s the next piece. This one was the turpentine. The next one is about the painting.