Tools With Opinions

This spring I ran an experiment. I published three essays here, Great PMs Don’t Agonize, Reset to Zero, and Stay Close Enough to Stop, under a thesis I call the temper line: the boundary between soft knowledge and hard knowledge is permanent terrain, and the craft of product work is learning to operate on that boundary.

The experiment was partly about the ideas. Mostly it was about ownership. After fifteen years of doing this thinking inside other people’s companies, I wanted a version of it that was mine.

Then I went quiet. April, May, then nothing, and the reason is worth being honest about. Drafting essay four felt like defending a thesis instead of practicing one. Every paragraph pulled toward the abstract: here is an idea about how knowledge works, here is why you should believe it. The ideas still hold. The essays were starting to read like documentation, and I’ve already written a whole piece about what happens to documentation.

What I did with those two months was build.

Intervals is a delivery forecasting tool. You give it your scope and your throughput, it runs ten thousand Monte Carlo simulations, and it answers “when will this be done?” the only honest way that question can be answered: as a distribution with confidence bands. The homepage says stop lying about deadlines, because a single delivery date is a guess dressed up as a plan, and every PM reading this already knows it.

Somewhere in the middle of shipping it, the thing I’d been circling for months clicked. Intervals is the temper line thesis. Gut-feel estimates are soft knowledge, probabilistic ranges are hard knowledge, and the tool sits on the boundary pushing things across it, which is what every essay I’d drafted was trying to argue in prose.

The difference between the two forms explains the stall. An essay defends an argument. A tool asserts one. When you drag the scope slider in Intervals and watch the confidence bands spread apart, nobody is persuading you that single-date estimates hide risk; you’re watching it happen to a project you recognize.

So this post makes the reframe official. Cold Work is a workshop. I build small tools with strong opinions about how product work should be done, each one asserting a specific claim, each one living at its own corner of this domain. The writing keeps a narrower job: every tool ships with a short written argument for its point of view, and sometimes, like today, the argument gets a whole post.

I should say something about how I build these, because I’m not an engineer. Intervals is a real deployed SaaS, auth and per-user data isolation and CI included, and I didn’t write the code. I run a system of AI agents that builds overnight against tickets I’ve scoped, reviews its own output against a written rulebook, and stops at a gate only I can open. My rule for that gate: green CI is necessary, not sufficient. Nothing merges until an independent review is clean and I approve it. That system gets its own essay, Running the Factory, if you want the machinery.

The next tool came from noticing something uncomfortable. Intervals points its skepticism at team estimates, and meanwhile I make hand-wavy claims all day: “that’s probably two weeks,” “I’m not worried about that customer.” None of them ever get checked. They’re bets I never wrote down, which means I get to remember being right.

So the next tool is a bet ledger. The rule: if you didn’t write down a number, you don’t get credit for being right. A judgment call gets logged as a falsifiable claim with a confidence percentage and a resolve-by date, or it gets consciously dropped. Right now it’s a plain text file and an agent that nags me at the moment I make a call, which is the right size for finding out whether the practice survives contact with real weeks. If it does, it becomes the next tool here, calibration curve and all, and I’ll open up the whole thing.

I notice that after declaring the tools the arguments, I wrote you an essay about it. The thesis predicted that much; the conversion between soft and hard knowledge was never one-directional, and the tools will keep needing a little prose.

In that spirit, the first entry in the ledger is a bet about the ledger: my confidence that I’m still logging bets a month from now is 75 percent. If I’m wrong, I’ll know by exactly how much, and when the repo opens you’ll be able to check.