About the author
Jason Gong here. Saturday Robot ships each week with what the most serious AI builders are actually doing. Before working on this and doing fractional growth work, I ran growth at Kite, an AI coding assistant from before ChatGPT (7M users, acquired by Affirm), founded Firezone (YC W22), and led GTM at GrowthX, where we built AI growth machines for companies like Lovable, Webflow, and Surge AI ($12M in 9 months).
This February, my token bill passed what a headcount costs for the first time. About $5,000 in a single month.1
It sounds like a lot, but as I went through what I spent the tokens on, I realized my small team had delivered what would have taken a team 3x the size a year or two ago. Most of the month ran through Claude Code and Codex. The people around me were doing work nothing in their job titles would have predicted.
The work is changing faster than job titles. Count the products launched for sales and marketing teams this year alone; every one is a new tool and a new way of working you now have to learn. That idea has been stuck in my head for months, and I'd guess plenty of people know the feeling without knowing what to do about it.
Coding's head start: code was already canonical, public, artifact-leaving
Almost eight years ago I was head of growth at Kite, an AI coding assistant that existed years before ChatGPT.2 The models then could autocomplete and hint, and that was it. Nothing could reason with you yet. We charged for tokens before most people knew what a token was.
The reason software absorbed AI so fast, I only saw it later. The substrate was already right. Engineers had spent decades getting their work into text: the code was the single source of truth, the docs got maintained, the pull requests got argued in public, and mistakes left a record. When the models learned to reason, decades of those habits started paying off at once.
Marketing gave a model nothing to work with. Strategy lived in a deck going stale, positioning was a paragraph somebody half-remembered from an offsite, and the playbooks were PDFs nobody opened. Almost none of the work left an artifact a model could read.
The same shift is now happening in other functions
Last year I built a go-to-market team where nobody had written a SQL query on day one. By the end of the year we were opening PRs in GitHub every week.
There was no elaborate workflow behind it. Context, tooling, and the person running it kept feeding each other, a flywheel, and it got better the longer it ran.
When Block cut roughly 4,000 people this February, Jack Dorsey wrote that intelligence tools paired with smaller, flatter teams were changing what it means to run a company.4 Not every layoff is downstream of AI. But the market has an incentive to tell that story, and stories like his harden into strategy before the technology catches up.
Why write about this?
The best thinking I've read on any of this comes from practitioners writing as they go: Paul Graham on writing as thinking,5 or Simon Willison, who ships a tool and has the blog post up the same week.6 Ben Thompson naming the structural force under an earnings call.7 Almost everything else is a 101 explainer or a 2030 prediction. The middle, where you decide what to try on Monday, is thin.
I also ran monthly workshops on AI in content marketing this past year, with thousands of registrants across them. The people who ramped fastest all had some version of a shared system under them. The people still stuck were running the same prompt-per-task loop they'd been running six months earlier, and nobody had shown them what a working setup looks like.
The promise: have something to try Monday
Every post will be about the work itself: how a thing got done, what made it hard, what broke, what I'm still figuring out. Mostly field reports from my own work, sometimes someone whose setup taught me something. If I don't have anything substantial, the piece will be short. When I do, I'll go deep. Either way you should walk out with one concrete thing to try at work that week.
That includes this post. Shipping with it is a starter kit for a knowledge base you can run agents on:8 templates and examples you copy over and have running in an afternoon. By Monday you can have your first agent reading your own work.
Saturday is the day you build whatever you want. The robot is the thing you end up building. Didn't exist on Friday, works by Sunday.
If the feeling I opened with sounds like your last year, I think you're in the right place.
A note: the starter kit goes live Monday, May 11. Parts of the site are still rough, I'm fixing things as I go.
Get the next post in your inbox
Companies with unfair distribution win. I'm collecting field notes from the world's best operators to help you think through growth.
Join top operators from Zapier, Clay, Lovable.