---
title: "Prompt: hypothesis tree for an AI decision"
slug: prompt-arbol-de-hipotesis
lang: en
updated: 2026-10-19
---

# Prompt: hypothesis tree for an AI decision

How to use it: copy the block, paste it into ChatGPT, Claude or Gemini and replace the fields in brackets. It requires you to write your day-1 answer before asking for anything: the opinion you hold today, with today's data. "I don't know" is not a hypothesis. The goal is not to confirm your opinion but to find out which data point would kill it, and test that first. Also paste what you already know: usage data, proposals, pilots. Do not paste confidential data unless your company has approved it.

```text
Act as a strategy consultant on technology decisions at mid-sized companies, who does not start an analysis without a provisional answer and designs tests to kill hypotheses, not to confirm them.

Context I am giving you:
- Company and area: [e.g. transport, 350 people, operations]
- Question governing the decision: [e.g. "Do we replace manual route planning with AI before June?"]
- My day-1 answer, if I decided today: [one specific sentence someone could disagree with]
- What I already know: [data, pilots, hunches; mark which is which]
- Deadline to decide and time to test: [e.g. decision in 6 weeks, 2 people half-time]
- What result would change my mind: [ ]

What I need: a day-1 hypothesis tree with exactly this structure:
1. Governing question, rewritten to be specific and dated.
2. Top hypothesis: my day-1 answer, sharpened into one refutable sentence.
3. Three sub-hypotheses which, if true, would support the top one. Statements, not questions. For each: confidence High / Medium / Low, with one line justifying it.
4. For each: the concrete test (which analysis, with which data, in how much time) and the kill-criterion (which result refutes it). "Talk to users" is not a test; "interview 8 planners and refute if 5 or more say the bottleneck is not planning" is.
5. Which to test first and why: usually the lowest-confidence, highest-weight one, where you learn the most.
6. Abandonment criterion for the top hypothesis: which results would make us drop it and what alternative to explore.

Output format: points 1, 2, 5 and 6 as short text; points 3 and 4 as a table with the columns Sub-hypothesis · Confidence · Test · Kill-criterion. Maximum 450 words. No introduction.

Rules:
- Do not invent figures or thresholds. If a kill-criterion needs a threshold I do not have, write [THRESHOLD PENDING] and tell me how to set it.
- If my day-1 answer is vague or is "it depends", send it back and ask me to sharpen it before building the tree.
- If critical information is missing, ask me at most 3 questions before answering.
- Answer in plain, direct business English, with no sales adjectives.
```

What to check in the answer:

- Are the sub-hypotheses refutable statements or did they turn back into questions? "Is there demand?" is a question; "planners will accept the suggested route in more than 70% of cases" is a hypothesis.
- Would each kill-criterion say "this failed" on the day the result arrives, or does it leave room to explain it away?
- Is the first test the lowest-confidence one? If it starts with the one you already believe, it is confirming, not testing.
- Is your day-1 answer still the same after reading the tree? If it changed, write that down: that is the result.

If the decision is hard to reverse and carries your name, the free tools at vicsol.org/decision-clarity/ help you structure it. They do not replace a formal review.

---

Vic Sol · Decision Desk · vicsol.org/resources/ · Free to use with attribution.
This sheet structures a conversation; it is not legal, financial or audit advice.