How I Make Decisions

Copy the decision system I run on my own business: filter what gets in, argue both sides before you commit, check the objective is still yours, and keep score on your own rulings.

By James Schramko · Updated August 2026

Decisions are the real product of your role. The team executes, the market responds, but the calls you make are the only thing that was ever exclusively yours. Most founders run those calls on mood, memory, and whoever they spoke to last. I run mine through a system, and the system is simple enough to copy.

Only Diamonds Get In

For years I filled journals at events. Pages of capture, copious notes, everything recorded. Later I noticed I never referenced any of it. The notes had no return. The capture felt like work and produced nothing.

So I inverted it. Now nothing gets into my system unless it survives compression. After a key moment I write one or two lines. If an idea cannot survive being compressed to two lines, it was not a diamond, and I let it go past.

The current trend runs the opposite way: hardware that records everything, transcribes everything, stores everything. That is coal-catching dressed up as productivity. It gives you more capture, not better judgment. Information is now infinite and free. The scarce thing is meaning, and a system built around capture instead of judgment is broken at the foundation.

Run it yourself: before anything enters your notes, your task list, or your plans, compress it to two lines. What survives is worth keeping. What does not survive was not worth carrying.

The Three-Question Filter

Every commitment already in the business gets tested against three questions. I built this filter years ago and it has outlived every tool I have run it in.

If it does not support revenue, delivery, or proof, remove it.
If two options exist, one is wrong. Pick.
If I hesitate, it does not belong.

These are blunt on purpose. They are the fast pass for the standing clutter: the subscriptions, campaigns, and commitments competing for the same time, money, and attention. The big, hard-to-reverse calls never get the fast pass. They go through the full system below. Inside its lane, the third question does the most work, because repeated hesitation about whether something belongs is a reason to re-test it, never a reason to keep it.

There is a fourth question that sits underneath the other three, and I ask it before any new piece of work starts: does this work need to exist at all? Most projects fail that question before they fail in the market. The cheapest decision you will ever make is deleting something before it starts.

Argue Both Sides Before You Decide

For any decision that matters, I do not ask my AI what to do. I make it argue. Both sides, fully, before I form a position. I assign roles and make them fight it out: an accountant, a lawyer, a sceptic, a buyer, whatever chairs the decision deserves. Marc Andreessen described a similar approach publicly in May 2026: steelman both sides, build a panel of expert personas, then make the judgment yourself. I had been running my version for years before that, and it has saved me from expensive commitments that sounded excellent in my own head.

Three rules make it work in practice, and each one exists because I learned it the hard way.

Discount the marginal findings. A reviewer never comes back empty, because finding nothing feels like failing the job. So an absence of objections is never proof you are right, and a pile of small objections is never proof you are wrong. Weigh the material findings, discard the rest.

One round per decision. Argue it once, adjudicate the findings, decide. If you re-run the challenge until it agrees with you, you are not validating, you are shopping. A second round needs new evidence, not a new mood.

Check the base rate. Ask what usually happens to people who make this exact move. Your inside view rarely matches the record. When your models disagree about that, the disagreement itself is signal, and I keep a separate playbook on exactly that: AI Outside View.

Is the Objective Still Yours

The most expensive failure class in decision-making is correct reasoning aimed at a stale objective. You can argue a decision perfectly and still get it wrong, because the goal it optimises stopped being your goal a year ago and nobody told the plan.

So the last check before committing is one question: what objective does this decision serve, and do I still want that objective? Businesses outgrow their goals quietly. The founder chasing scale who actually wants freedom, the operator defending a product line that belongs to an earlier identity. Ask the question out loud. If the objective is stale, no amount of good reasoning downstream can save the decision.

Keep Score On Your Own Rulings

I log my strategy decisions with a status: active, resolved, or superseded. Superseded is a healthy word. It means new evidence arrived and the ruling got updated, which is the system working.

Count your own overturns. If nothing in your decision log ever gets revised, check whether new evidence is genuinely capable of changing your mind, because a validation process that never changes your mind is a yes-man with extra steps.

The Human Outside the Storm

Most of this system should run without anyone else, and you should run it that way. The exception is a decision where your own exposure to the outcome makes calibration difficult. When you are inside that decision, you are also inside the weather that made it urgent: the cash pressure, the ego, the sunk years, the identity attached to the old plan. The models will argue whatever you point them at, but you chose where to point them.

The calibrator has to stand outside your storm. Someone who has seen the pattern enough times to say: I know it feels unique to you, and here is what usually happens next. That role does not compress into a prompt, because its whole value is that it does not share your weather.

Quick Reference

Compress everything to two lines before it gets in. What cannot survive compression was not a diamond.
Run the fast-pass filter on standing commitments: revenue, delivery, or proof; two options means one is wrong; repeated hesitation means re-test it.
Ask whether the work needs to exist before asking how to do it.
Make your AI argue both sides in assigned roles before you form a position. One round. Discount the marginal findings. Check the base rate.
Before committing, confirm the objective the decision serves is still your objective.
Log rulings with a status and count your overturns. If nothing ever gets revised, check the process can actually change your mind.
Keep one calibrator outside your own weather for the decisions you are too close to.

Being the calibrator outside the storm, with the pattern library to back it, is the seat I hold for every founder inside Mentor.