The Shape of Work is Changing

If you draw a chart of effort applied in different phases of making things, the shape of the curve on that chart is changing.

For the purposes of this essay, it’s useful to think of work in three stages: planning, producing, and perfecting.Yes this is highly reductive! And while these reductive phases happen roughly sequentially, you can jump back and forth if needed. For example, you could produce a draft of a presentation, share it, then refine it.

Planning is all the stuff you do before you start to produce any particular work artifact. It can include gathering data, putting together a strategy, researching, reviewing, deciding what to work on. It’s doing the thinking.My friend Duncan used to wear a baseball cap that had “thinking” embroidered on it. It always made me laugh, and writing this essay reminded me of it.

Producing is making the artifacts of knowledge work: memos, deals, contracts, presentations, templates, code, designs, agendas, plans, betas. You are making an actual thing a colleague can see, touch, interact with, and react to.

Perfecting is the most obvious. It’s editing, refining, revising, adjusting.

Before AI, I believe most work followed a chart that roughly looks like a shifted sine curve. For a typical work product, the level of effort would be small, large, then small again.This has been how work felt to me for my entire career, across academia, investing, and operating. It’s the stage where I am most likely to get into flow state. I can plan all I want but until I’m creating (producing), I’m not really in it in the same way.

Now, the three phases are inverted. With AI tools, production is asymptotically driving to zero effort. A deck, code, a memo: they can be made nearly for free now.

But the other two phases require more effort than before. Gathering or producing novel data, or slurping it up into a way that an AI tool can use, that takes a lot more time and effort than it used to.Or said another way: you still need to think. Planning out what to produce, and then prompting, is a new part of this phase that didn’t exist before.There’s an argument that PRDs are the planning phase for engineering work. But I think because PMs aren’t engineers, the PRD is the end of the producing phase for that person, and the code is the product for the engineer. So this model still applies. And perfecting! As anyone who has reviewed AI output knows, this phase takes longer than it used to.

The shape of the curve looks like a shifted cosine now.

One strange consequence of this is that what used to look like work doesn’t look like work anymore. Thinking, walking, reading, interviewing, staring off into space. This is the biggest part of “productivity” now.Arguably it always has been, but when producing took more effort, the percentage has shifted. I wonder how long it will take us to catch up to this. It could be a very long time, because what the work curve used to look like has been consistent for so long.I know for myself that writing and rewriting a prompt, or prompting an agent 137 times, doesn’t put me into flow state the same way writing does.

I spoke with a CEO friend of mine last year. She complained about AI memos. Her executive team kept sending her memos for this, memos for that. Now that writing a memo takes near-zero effort, they could be more productive! But they hadn’t caught up. Thinking and perfecting are the work now. But I feel for them; the main thing they used to bring to their boss is now free to make, so they are making more of it.

Last year I went to a vibe coding hackathon and experienced this first hand. I spent most of the day banging my head against the proverbial design wall. I don’t usually write memoir-style essays, but I’m including it below as it describes the feeling of expecting a sine curve and experiencing a cosine one.

One thing I haven’t figured out is if the total amount of work (the integral under the curve) has gone down or just changed shape. AI maximalists argue it’s gone down and that will make us more productive as a society. But has it? I’m not sure yet.

A Vibe Coding Nadir

I went to build a personal app: a recipe keeper with version control.I like to follow the scientific method of cook > taste > tinker > adjust recipe. I want software to re-home my messy recipe notes that currently live in post-its and Google Docs, including one titled “Neverending Experimental Perfection Bagel Recipe.” Although it stopped after Covid, I used to have a habit of collecting cookbooks and reading them cover-to-cover. I imagined an app that evoked those afternoons reading a cookbook purchased that day, marking the recipes I wanted to make first.

Using ChatGPT and Claude, I wrote a PRD, then uploaded that PRD into the vibe coding tool. I tinkered with the functionality on the edges, nipping and tucking in a few places. Watching the agent work was pleasant enough, evoking the lumbering but forward-moving feeling of a large rock rolling along the bottom of a stream. Then vYes this is highly reductive! And while these reductive phases happen roughly sequentially, you can jump back and forth if needed. For example, you could produce a draft of a presentation, share it, then refine it. was done, and it felt good. This is it: from idea to clickable thing in less than an hour!

Predictably, pizza arrived. As my fellow participants vacated our communal table, I looked down at the app I had created. It was hideous. It was white with accents of construction-sign orange: neither vibrant nor neon, squint-inducing yet with putrid undertones. It didn’t invite an appetite. Buttons had the rounded corners of 2019, the font was indiscernible muck, and the recipe creation form was less appealing than a form on the California DMV website. Why was it a form anyway?

I hated it.

In my mind, I wanted a bookish, spacious place to tinker with the most sensual thing: food. I wanted to combine the expansive white space of writing, the minutiae of scribbling recipe changes in book margins, and the hearth and beauty of home cooked food. My mother-in-law used to say, “food works all the senses. First you hear it, then you see it, then you smell it, then you taste it.” This app repulsed all my senses.

For the next two hours, I fought to get the agent to design something I could stand.

I started by going back to planning, and writing down in more detail what I thought I wanted. But I hit the boundaries of my vocabulary almost immediately. Although I have been managing creatives and creative processes for over ten years, until this moment I hadn’t realized how much my design counterparts had been doing to translate what I thought I was saying into what I was really saying. If I said “make it more bookish” while waving my hands in the air, then my Creative Director might say to himself “let’s switch to a softer serif, like Iowan Old Style, reduce that line height, use spacing to create hierarchy rather than big swings in font weight and size, and maybe have asymmetrical margins to make that bookish feel.”I actually had to email Frank to get this quote here. If I said “I want this to feel more vibrant and warm,” he might mentally translate that into contrast, spacing, or color combinations. I railed against the limits of my language for about thirty minutes, getting nowhere with the single exception of removing the putrid orange. Now I had a functional app in black and white.

I decided to use an example to explain what I meant. Dear agent, here is what I’m looking for: here is a sample site designed by a friend. Isn’t it great? See the fonts and lines? It feels literary. Make mine like that.

But the agent couldn’t parse what characteristics of that model website I was attempting to replicate, even when I told it in my halting designglish, instead it implemented the shape of the grid system and left fonts, spacing, and lines out of it.

Okay! I will give it more examples that are similarly bookish. But there’s one problem. I can’t think of any. I am not a designer in lexicon nor in visual memory. I don’t have a mental library of brands, websites, images, concepts, or artists.

I imagine a designer’s mind as a neural net of visual nodes. With a stimulus prompt, they can reach into their neural net and find associated other visual nodes that branch off along varying patterns or characteristics. Bookish like this news site? Or this site with images of libraries? Do you mean bookish like the books that this publisher creates? Or by bookish do you mean large serif font like this brand uses, which is not actually how books are printed at all? Or do you mean the types of fonts this foundry makes? Et cetera.

I went to chat with my friend who works at the vibe coding company. He’s a designer, so I asked him, “how do you give the agent you design design input?”

“Oh it’s really hard,” he said. “The best thing to do is come up with example websites, around four or five. And if that doesn’t work sometimes I just ask it to build the app three different times, then adjust from whichever one is the starting point that most looks like I wanted it to. It never builds the same thing twice.”Funnily enough, this is exactly what designers do with clients. They present three concepts, then play “Marco Polo” as my friend Frank calls it. Even though no one wants to admit it, the first round is just for calibration, then it goes from there.

It had been nearly three hours of vibe coding. The prizes were going to be awarded in an hour, and my black and white, DMV-inspired recipe keeper showed no signs of change. My best hope was to start over, three more times, in parallel.

I packed up my laptop and left 30 minutes before anyone showed their work.