---
title: "Snacks are cheaper and tastier than ever"
date: 2026-07-31
tags: post
---

Perhaps the single most consequential essay I've read is Will Larson's [Work on what matters](https://lethain.com/work-on-what-matters/). This is ostensibly advice targeted towards staff engineers, but the content generalizes across disciplines. One of the failure modes he talks about is snacking:

> If you're in a well-run organization, at some point you're going to run out of things that are both high-impact and easy. This leaves you with a choice between shifting right to hard and high-impact or shifting down to easy and low-impact. The latter choice–easy and low-impact–is what Walk refers to as _snacking_.
>
> When you're busy, these snacks give a sense of accomplishment that makes them psychologically rewarding but you're unlikely to learn much from doing them, others are likely equally capable of completing them (_and_ for some of them it might be a good development opportunity), and there's a tremendous opportunity cost versus doing something higher impact.

As my role as a founder has shifted to be more explicitly managerial over the past few years, I find myself returning to it whenever I am feeling particularly guilty of the behavior it espouses.

One of the things that has made this much more difficult over the past year has, of course, been the rise of LLMs. One of the things they have done is made it incredibly easy to avoid what [Horst Rittel and Melvin M. Webber](https://en.wikipedia.org/wiki/Wicked_problem) refer to as wicked problems: problems with no clear success criteria, no right to be wrong, and no prior art (amongst other things.)

A few years ago, I could procrastinate on thinking deeply about go-to-market for a new segment by maybe hammering out a single bug fix or spending an hour in the support queue. Now, suddenly, my alternative to doing deep work contains many new Cyrenaic properties:

1. I can pretend that it is in fact managerial work. Asking Cursor to build a visualization of metrics feels like it is something that is akin to strategy work, even though no decisions were made differently as a result of it.

2. I can build and ship features trivially. And these are features that I earnestly believe are things that, given an infinite time scale, I would want the product to have — even if the reason we haven't built them yet is because it's not obvious how it actually generates leveraged value for the business, nor is it obvious how to prove it.

3. I can spec out and create high-fidelity prototypes for interesting new product directions. Even if those prototypes end up gathering dust, and by the time I'm ready to seriously consider them, the context which necessitated or informed their construction is obsolete.

So what to do?

---

One answer is simply that I don't know, and it feels very important -- like, "I have written and discarded six other versions of this essay over the past eight weeks" important -- to try and figure it out, because if you have good eyesight you can probably see a Red Queen problem lurking on the horizon where all of the software factories in the world nullify any marginal advantage that the others might have. [^1]

Another answer is "exercise more personal and organizational discipline". This is true but facile.

Here's the answer that feels best right now: the definition of something being done must shift from the code being shipped to customer realizing that. That sounds both abstract and obvious, I know, and I'm sure you can compellingly argue that every good company should already be operating from that position. But this is a personal blog, and in my personal experience, I can say that I've gotten surprisingly far with Budden and other projects, pretending I am Kevin Costner and knowing that if I build it, they will come. 

LLMs optimize the transformation of input into output; in my experience, all the context in the world does not aid them in creating customer value from first principles. Now more than ever, the job is to build both wisdom of what users find valuable and knowledge of how to decompose that into legible work.  

[^1]: We are far away from this. The median software company, despite what you read online, is still figuring out how to get its CRM hooked up to Claude, but we're not going to be far away forever.