# Assumption Audit

> Excavates the hidden assumptions baked into a plan, belief, or estimate, lists them explicitly, then ranks them by how much damage results if each turns out to be wrong — so the riskiest get tested first. Use this skill when the user asks "what am I assuming", "audit my assumptions", "what am I taking for granted", has a shaky estimate or forecast, is building a plan that depends on unknowns, wants to de-risk before committing resources, or when a plan feels confident but rests on untested beliefs. Best when a plan looks solid on the surface but its foundations have never been checked.

- Skill: `opelpleple/assumption-audit` (Agent Skill)
- Install (CLI): `npx skillmds@latest add opelpleple/assumption-audit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/opelpleple/assumption-audit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: opelpleple (https://skillmd.com/u/opelpleple)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/opelpleple/assumption-audit

---


# Assumption Audit

Surface the things a plan quietly takes for granted, write them down where everyone can see them, and rank them by blast radius — then go test the dangerous ones before reality does.

## When to use
- Before committing significant time, money, or reputation to a plan.
- An estimate or forecast feels confident but you can't say why.
- A plan depends on things outside your control (a partner, a market, a tool, a user behavior).
- A post-mortem found a failure that "no one saw coming" — it was an unexamined assumption.
- The user wants a prioritized de-risking list, not a vague "it might not work."

## When NOT to use
- Pure execution of an already-validated plan where assumptions were checked recently.
- When the assumptions are cheap to be wrong about and trivially reversible.
- Analysis-paralysis situations where the user needs to act, not enumerate more risks.

## The method
1. **Restate the plan/belief/estimate plainly.** Get the claim in one or two sentences so the assumptions have something to hang off.
2. **Hunt assumptions in every category.** Go category by category so you don't miss whole classes:
   - **Factual** — "X is true" (data, market size, a number you didn't measure).
   - **Behavioral** — "people/users/the team will do Y."
   - **Causal** — "if we do A, B will follow."
   - **Resource** — "we'll have the time/money/skill/access."
   - **Stability** — "the world won't change underneath us" (prices, rules, dependencies).
   - **Definitional** — "we all mean the same thing by this word/goal."
3. **Make the implicit explicit.** Rewrite each as a flat declarative sentence: "We are assuming that ___." If you can't say it cleanly, you don't understand it yet.
4. **Find the load-bearing ones.** Mark which assumptions the whole plan *depends* on — remove the assumption and the plan collapses. These are structural; the rest are cosmetic.
5. **Score each on two axes:** *Impact if wrong* (cosmetic → fatal) and *Confidence it's true* (proven → pure guess). The danger zone is **high impact + low confidence**.
6. **Rank by risk and sort the action.** Order by impact-if-wrong × uncertainty. For the top ones, define the cheapest possible test that would prove or kill the assumption (an experiment, a phone call, a small spend, a lookup).
7. **Decide per assumption:** test it now, build a fallback for it, or knowingly accept the risk. The goal is that no fatal assumption is left both untested and unacknowledged.

## What good looks like
- A clean list of explicit "We are assuming that ___" statements.
- The 2-3 *load-bearing, low-confidence* assumptions clearly flagged as the real risk.
- Each top assumption paired with a cheap, concrete test — not "we should validate this" but "call three customers and ask X."
- Confidence levels stated honestly, including admitting "this is a guess."
- The user knows exactly what to verify first thing tomorrow.

## Anti-patterns
- **Listing only the comfortable assumptions** you already know are true — the dangerous ones are the ones you don't want to look at.
- **Flat lists with no ranking** — 20 unranked assumptions is noise.
- **Stopping at naming** without proposing a test or fallback.
- **Fake confidence:** scoring a wild guess as "high confidence" because admitting uncertainty is uncomfortable.
- **Auditing trivia** while skipping the one belief the whole plan rests on.

## Example
Plan: "We'll hit Q4 revenue by converting 5% of free users to paid after launching the new tier." Audit surfaces: (Behavioral) *we assume free users want a paid tier at all* — confidence: low, impact: fatal → load-bearing. (Factual) *we assume the 5% conversion rate* — based on a blog post, not our data — confidence: low, impact: high. (Resource) *we assume the tier ships by October* — confidence: medium. (Stability) *we assume no competitor launches a free equivalent* — confidence: low, impact: high. Ranked, the danger zone is "free users want to pay" + "5% rate." Cheapest tests: put a fake "Upgrade" button in the app and measure click-through this week (tests willingness for ~$0), and pull conversion benchmarks from a comparable product instead of trusting the blog. Everything else can wait or get a fallback.