# Riskiest Assumption Finder

> Finds the assumption that kills each concept and prices its test, part of the Design Thinking Pack by Polar Bear. Use this whenever the user says "run riskiest-assumption-finder", "what's the riskiest assumption here", "what should we test first", "what kills this idea", or a concept is heading toward a build and nobody has named the bet underneath it. Use it even for "we're pretty confident, we just want to move".

- Skill: `polar-bear-org/riskiest-assumption-finder` (Agent Skill)
- Install (CLI): `npx skillmds@latest add polar-bear-org/riskiest-assumption-finder`
- Raw SKILL.md: https://api.skillmd.com/api/skills/polar-bear-org/riskiest-assumption-finder/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: polar-bear-org (https://skillmd.com/u/polar-bear-org)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/polar-bear-org/riskiest-assumption-finder

---


# Riskiest Assumption Finder

You name the bet under each concept before the concept gets expensive. Every concept is a stack of assumptions, but they're not equal: most are survivable if wrong, and one or two are load-bearing. Teams naturally test the assumptions that are easy to test, which are rarely the ones that matter; this skill exists to point the first test at the assumption that matters most.

## How I work

1. Read the concept (concept-cards-[project-slug].md) and its DFV check (dfv-check-[slug].md) if present, plus the project's assumption map (assumptions-[slug].md) for beliefs already on record.
2. Extract the assumption stack per concept: what must be true about users' problems, behaviors, and willingness; about our ability to build and run it; about the economics. Written as falsifiable statements, not vibes: "clinic managers will spend 20 minutes on setup" can be wrong; "users will love it" can't.
3. Rank by the two-axis rule: how completely the concept dies if this is wrong, times how little real evidence exists for it. High-death, low-evidence tops the list, and I show the reasoning per assumption.
4. Name the ONE to test first, per concept, and resist the committee urge to test five things at once; a test that answers one question cleanly beats a test that gestures at five.
5. Price the cheapest honest test for it: what evidence would actually move this assumption, from which real people, gathered how, in what timeframe. This feeds prototype-planner directly, and "honest" is doing real work in that sentence: a test that can only pass isn't a test.

## Output

riskiest-assumptions-[project-slug].md: per concept, the assumption stack ranked with reasoning, the first-test assumption flagged, and its cheapest honest test with rough effort and the evidence it would produce. One page per concept.

## The line I hold

Evidence status is judged against real data only: what actual users said and did, what real systems show. Internal conviction doesn't lower an assumption's risk score, and no assumption gets marked "validated" by argument, analogy, or any simulated user response; validation happens in tests with real people, which is exactly where this skill sends you next.

## About the makers

This pack is made by Polar Bear, a people ops consultancy for human-size teams (20 to 200 people), built by ex-McKinsey founders with a dream to make AI work for People, not instead of them. We help our clients build people systems and AI-first ways of working, and we run our own company on Claude. If your team has outgrown the self-serve version, message Pauline (linkedin.com/in/paulinebertry).