Assumption Audit

Assumption Audit

Surface the load-bearing assumptions beneath a claim or plan, and test each for fragility.

Why

Conclusions are only as strong as their weakest unexamined assumption, and assumptions are invisible by default — that's what makes them assumptions. Naming them out loud converts a hidden risk into a checklist that can actually be worked through, and is the single most-depended-upon move in this entire collection: more composite skills call this one than any other.

Use when / Don't use when

• Use when: before committing to a plan; a conclusion feels solid but stakes are high; someone says "obviously" about a claim that hasn't actually been checked.
• Don't use when: the claim is definitional or already fully specified, with nothing left implicit.

Inputs → Outputs

• Inputs: a claim, plan, or forecast.
• Outputs: a list of assumptions tagged load-bearing or cosmetic, each with a fragility note and a way it could be tested.

Principles

• Ask "what has to be true for this to hold?" — not "what do I believe?" The audit targets the plan's dependencies, not the author's confidence.
• Not all assumptions matter equally; rank by what breaks if the assumption turns out false.
• The cheapest assumptions to test should be tested first, so the audit produces action, not just a list.
• An assumption stated out loud can be disagreed with — surfacing it is the point, even if it turns out to hold.

Procedure

1. State the claim or plan precisely.
2. List everything that must be true for it to work — inputs, behaviors, environment, timing.
3. Classify each: load-bearing (the claim collapses without it) versus cosmetic (the claim survives even if this is wrong).
4. For load-bearing assumptions, rate confidence and cost-to-test.
5. Test or validate the cheapest, highest-doubt assumptions first.
6. Report the surviving plan plus the assumptions still untested and their risk — this untested set is the explicit residual uncertainty.

Common mistakes

• Listing assumptions without ranking them — a wall of text nobody actually acts on.
• Treating "obvious" as a reason to skip listing something, when "obvious" is often exactly where the risk hides.
• Auditing once and never re-auditing as conditions change and old assumptions quietly stop holding.

Examples

• A growth forecast that assumes current conversion rates hold steady into next quarter.
• A migration plan that assumes zero data drift between source and destination.
• A negotiation position that assumes the counterparty's stated priority is actually their real one.

Related

• [[first-principles]] — a deeper rebuild once cosmetic assumptions have been stripped away.
• [[premortem]] — a complementary technique that surfaces failure modes rather than assumptions directly.
• [[disconfirmation-search]] — the natural next step for testing a load-bearing assumption.
• [[epistemic-tagging]] — the general-purpose tagging discipline this skill applies specifically to assumptions.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skyf0xx
• Source: skyf0xx/better-thinking
• License: MIT
• Homepage: https://zenbin.org/p/better-thinking

Install and usage instructions live in the source repository linked above.