The 2x2 I use to fork past decisions
Counterfactuals aren't thoughts. They're places. Build a worktree for the alternate timeline.
The reframe
Most "what if" tooling treats counterfactuals as thinking. Paragraphs of speculation that vanish into the void. "What if we'd priced lower?" "What if we'd shipped Procore first?"
The unlock: treat counterfactuals as PLACES. A folder. A git worktree. A wiki entry with resolvable conditionals and a date.
The Shell 2x2
When I fork a past decision, I use a 2x2 grid borrowed from scenario planning at Shell:
| | Best-case world | Worst-case world | |---|---|---| | Best-case decision | Ideal outcome | Decision saved you | | Worst-case decision | World saved you | Lesson learned |
Four quadrants. Each one is a separate worktree, each one with a wiki entry, each one with a resolution date.
Example: pricing for Struvo Pro. Best-case decision was $349. Worst-case world is "JDA pulls out, $349 is too high for the next 5 GCs we pitch." Worst-case decision was $199. Best-case world is "$199 anchors low and we never recover the unit economics."
Each scenario lives in business-ops/llm-wiki/counterfactuals/2026-05-08-struvo-pro-pricing/<quadrant>.md. Each one has a resolution criterion: "By 2026-08-01, what's the actual outcome?"
Why this is different from journaling
The crucial bit: each entry has a resolution date. On 2026-08-01, I open the four entries and grade them. The wiki becomes a self-grading forecaster — over time, I learn whether my decision-making is calibrated.
Without resolution dates, the wiki accumulates speculation. With them, the wiki becomes data.
What you can steal
If you've made a decision in the last week that wasn't obvious:
- Pick the next significant decision you're about to make
- Draw the 2x2 — your decision options vs. world states
- For each quadrant, write 3 sentences: what happens, why, what the resolvable condition is
- Pick a resolution date. Put it on your calendar.
- Open the file on the resolution date. Grade your forecasting.
This is what I'm building into the /counterfactual skill. V1 is fork generation. V4 is a self-grading forecaster after 6 months of data.
The compounding move isn't the forecasting. It's the calibration — knowing whether your gut is right at 60% or 80%.