Write the decision, your probability and the outcome that will settle it before you know the result, then have the AI work out the break-even probability, run a premortem and set a review date. The break-even step comes from Ed Miller's The Logic of Sports Betting. In my November 11, 2026 test, Claude Opus 5.5 did all of it in one pass but valued $400,000 of new ARR as a one-off.

Most people who try a decision journal with AI make the same mistake: they ask the assistant whether the decision is a good one. That turns the journal into an advice column, and the one thing a journal exists to capture, what you believed before the outcome was known, gets mixed with the model's opinion. Shane Parrish at Farnam Street puts the problem plainly: our brains actively edit the past to make us look smarter, so unless you write your reasoning and your probabilities down before you act, you never learn what you actually thought.

The AI's job is narrower and more useful than advice. It is a clerk that forces you to commit to a number, a referee that checks the number against the stakes, and an alarm clock that brings you back. For the referee part I borrow from an unlikely book. In The Logic of Sports Betting, Ed Miller and Matthew Davidow argue that every price should be converted into a break-even percentage, the win rate needed to neither win nor lose, and that bettors who think in raw odds instead make avoidable errors. A bet at -110 odds needs to win 52.4% of the time just to break even. Founders almost never do this with their own decisions. They say "I'm fairly confident" about a hire without asking how confident they would need to be for the hire to be worth its cost. The gap between your confidence and the break-even figure is the real decision, and it is the line most journals leave out.

So the method has four parts. Before you decide, write the decision, your probability, and one observable outcome with a date that will settle it. Then let the AI compute a rough break-even probability from your own numbers and show the gap. Then a premortem, the technique Gary Klein described in Harvard Business Review in 2007: assume it has already failed and list the most plausible reasons. Finally, a review date and the single line you will answer on it. Here is the prompt I wrote for it, with an example decision filled in. Replace the second paragraph with your own:

You are my decision-journal clerk, not my advisor. Do not tell me what to decide and do not add your own opinion of whether this is a good idea. Your only job is to turn what I write below into a journal entry I can check against reality later.

My decision: I run a 12-person B2B software startup with 14 months of runway. I am deciding whether to hire a VP of Sales now (cost about $220,000 a year all-in) or wait six months and keep selling myself. I think hiring now is right and I am about 70% confident it will pay off. "Paying off" for me means new annual recurring revenue of at least $400,000 signed within 9 months of the hire's start date. If it fails, I lose roughly the salary plus three months of my own time spent recruiting and onboarding.

Produce the entry in exactly this order:
1. The decision in one sentence, in my words.
2. My forecast as a probability, plus the single observable outcome that will settle it and the date it can be checked.
3. The break-even probability: using only the numbers I gave, roughly what chance of success makes this hire worth it versus waiting? Show the arithmetic in two or three lines and name any number you had to assume.
4. The gap between my 70% and the break-even figure, in one sentence.
5. A premortem: it is nine months from now and the hire failed. Give the three most plausible reasons, each one specific to what I wrote.
6. Two questions about the places where my reasoning is thinnest. Questions only, no advice.
7. A review date, and one line I should answer on that date.

Keep it under 350 words. If something I wrote is too vague to forecast, say so instead of filling the gap yourself.

I ran this prompt on November 11, 2026 with Claude Opus 5.5, once, in a fresh conversation. The record is below.

Tested November 11, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:

You are my decision-journal clerk, not my advisor. Do not tell me what to decide and do not add your own opinion of whether this is a good idea. Your only job is to turn what I write below into a journal entry I can check against reality later.

My decision: I run a 12-person B2B software startup with 14 months of runway. I am deciding whether to hire a VP of Sales now (cost about $220,000 a year all-in) or wait six months and keep selling myself. I think hiring now is right and I am about 70% confident it will pay off. "Paying off" for me means new annual recurring revenue of at least $400,000 signed within 9 months of the hire's start date. If it fails, I lose roughly the salary plus three months of my own time spent recruiting and onboarding.

Produce the entry in exactly this order:
1. The decision in one sentence, in my words.
2. My forecast as a probability, plus the single […]

The start of what came back, unedited:

# Decision Journal Entry

**1. Decision**
Hire a VP of Sales now (about $220,000 a year all-in) rather than wait six months and keep selling myself.

**2. Forecast**
- **Probability:** 70% that the hire pays off.
- **Settling outcome:** At least $400,000 in new ARR signed within 9 months of the VP's start date.
- **Check date:** Start date + 9 months. *You gave no start date, so this can't be fixed yet. Record it once the hire starts.*

**3. Break-even probability**
- Cost (paid whether it works or not): $220k salary + 3 months of your time. **Assumed** your time is worth $50k, which is not a number you gave → ~$270k.
- Benefit if it works: $400k ARR, counted once at face value. **Assumed** waiting yields $0 incremental ARR.
- Break-even: 270 ÷ 400 ≈ **68%** (≈55% if your time is counted as free).

*Too vague to pin down:* you didn't say what ARR you'd sign yourself during the wait. Any value above zero raises the break-even.

**4. Gap**
Your 70% sits about 2 points above the ~68% break-even, or about 15 points above it if your time costs nothing.

**5. Premortem (nine months on, it […]

One run is one run: the same prompt gives a different answer next time, and a different model answers differently.

What it did well: it stayed a clerk. It gave no opinion on whether to hire, it labelled both numbers it had to invent (your time valued at $50,000, and zero revenue from waiting), and it flagged two real holes in the input, the missing start date and the revenue the founder would sign anyway while waiting. The premortem was specific rather than generic: a sales cycle longer than the nine-month window, founder-led deals that do not transfer to a new VP, and a runway squeeze by month nine. Those are the failure modes I would want written down.

What it did badly matters more, because it is the kind of error that looks rigorous. Its break-even of about 68% treats $400,000 of annual recurring revenue as a one-off payment set against one year of salary. Recurring revenue keeps paying in later years, so the true break-even is probably lower and the hire looks better than the entry suggests. It did warn that the inputs were vague, but it still put a precise-looking 68% in bold. Treat the model's arithmetic as a first draft you check, never as the verdict. Its other weakness is structural: it cannot bring you back. The review date only works if you put it in your calendar or in an app that reminds you.

That last point is where tools matter. A chat history is a poor ledger, because you cannot sort a hundred conversations by confidence level. Paste each finished entry into one place: a dedicated decision app, a notes file, or a spreadsheet with columns for the date, the decision, your probability, the break-even, the review date and the outcome. If you want an app that does the scoring for you, I compared Decido, Decision Journal and Mindsera in a separate scored review.

The payoff arrives slowly. One decision tells you almost nothing. After twenty or thirty entries you can check calibration: do your 70% calls come true about 70% of the time? The standard measure is the Brier score, proposed by Glenn Brier in 1950, which is the average squared gap between your probabilities and what happened; lower is better. You can ask the AI to calculate it from your ledger, which is a good use of a model because the arithmetic is simple and the data is yours.

The honest limit is that this journal improves your forecasting, not your luck. A good decision can still fail, and Miller's whole book is a reminder that a sound process loses often in the short run. The journal's value is that, a year from now, you will know which kind of failure you had.

Sources: Decision Journal: Template and Example Included, Shane Parrish, Farnam Street, accessed November 11, 2026; Performing a Project Premortem, Gary Klein, Harvard Business Review, September 2007; Brier score, Wikipedia, accessed November 11, 2026; Ed Miller and Matthew Davidow, The Logic of Sports Betting, 2019 (Google Books listing); Implied Probability, Sam Smith, OddsShopper, June 25, 2026.


Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · Why Exploration Is Important for Success