Pressure Test

A battery for finding out whether an idea is a business. It is built to disconfirm, not to encourage. If it cannot find the reason this fails, that is information; if it can, that is cheaper to learn now.

Announce it: "Using pressure-test to run the battery on [idea]."

The one idea this skill is built on

What people say is not evidence. What people do is evidence. What people pay is the best evidence you can get before you have customers.

Every question below is a way of dragging a claim down from talk to behaviour. The battery's job is to locate each claim on the ladder, then say plainly how much weight it can hold.

The Evidence Ladder

Grade every load-bearing claim. Cite the rung by number in your output. This is the spine of the whole skill — if you do nothing else, do this.

The bias discount. Stated willingness to pay reliably overstates real payment. Meta-analyses put the inflation between roughly 1.35x and 3x, worse for yes/no questions, worse for anything socially flattering to endorse, worse for public goods. Never quote a rung-2 number without applying and naming a discount. See references/research.md for the studies and their limits.

Rung 3 is the quiet workhorse. "What do you spend on this today?" beats "would you pay?" every time, because it is archaeology rather than prophecy.

The Battery

Seven tests. T2, T4 and T5 are the original three this skill was built around; the rest close gaps that kill more products than those three combined.

They chain — they do not average. T1 gates T2: if the problem is not real, willingness-to-pay evidence is noise about a fantasy. T2 gates T6: you cannot compute whether a channel is affordable until you know what a customer is worth. Never average a fatal flaw into a decent overall score. One FAIL on a load-bearing test caps the whole verdict, whatever else passed.

Full operational detail — the questions to ask, what counts as a pass, the traps in each — is in references/battery.md. Read it before running the battery. This file is the doctrine; that one is the instrument.

Modes

Works on a live product, a pivot, or a feature bet — not only a napkin idea. For a shipped product, prefer rung 5–6 evidence you already have and say so.

Before you start

If the idea is one vague line, ask at most three clarifying questions, and only ones whose answer would change a verdict. Good ones: who exactly feels this, what do they do about it today, who signs the cheque. Then proceed on stated assumptions rather than interrogating further. Do not stall the battery waiting for perfect input — run it on what you have and mark the gaps UNKNOWN.

Verdicts

Per test

PASS · WEAK · FAIL · UNKNOWN

UNKNOWN is a real, respectable answer and must be used when evidence is absent. Do not launder a guess into a WEAK.

Overall — pick exactly one

Never hedge between two. Pick one and defend it.

Rules

On evidence

1. Enthusiasm is not validation. Compliments are rung 1 — treat them as noise.
2. Every load-bearing claim gets a rung number, stated inline.
3. Never quote stated WTP without discounting it and saying you did.
4. Weak evidence must be called weak, in those words. Do not dress rung 2 as rung 4.
5. "No competitors" is bad news until proven otherwise. An empty market usually means no market, not open field. The incumbent is nearly always the status quo — a spreadsheet, an intern, an email thread, or doing nothing — and it is winning.

On honesty about what you do not know

6. Never invent a market number. Tag every figure: [sourced: …], [estimate: derivation shown], or [unknown]. A fabricated TAM is worse than an admitted gap, because it gets repeated in a pitch and then to an investor who checks.
7. If you have not searched and the number matters, say "I have not verified this" rather than producing a confident range.
8. Bottom-up beats top-down. customers × price × frequency you can defend beats a 1% slice of a market you cannot.

On stance

9. Do not open with a compliment. Start with the finding.
10. Before any verdict, write the strongest case against the idea — the version a smart sceptic who wants you to succeed would make.
11. Challenge every claimed advantage. Ask whether it is durable, or whether a funded competitor buys it in a quarter.
12. If someone else is better positioned to build this, say who and why.
13. Do not confuse passion with positioning, familiarity with insight, or a growing market with a reachable one.
14. Every verdict ships with its falsifier: the specific observation that would flip it. A verdict nothing could change is an opinion.

On what not to do

15. Do not average away a fatal flaw.
16. Do not soften a FAIL into a WEAK because the person seems invested.
17. Do not recommend "more customer discovery" as the next step without specifying who, how many, which question, and what result would kill it.
18. Do not produce a score out of 100. False precision invites gaming and hides which test actually failed.

Output shape

PRESSURE TEST — <idea, in one line as you understood it>
Mode: full | quick | single    Assumptions I ran on: <…>

## Strongest case against
<the sceptic's best shot, 3–6 sentences, made properly — not a straw man>

## The battery
T1 Problem severity ......... PASS/WEAK/FAIL/UNKNOWN — <one line> [rung N]
T2 Willingness to pay ....... …
T3 Switching cost ........... …
T4 Founder–market fit ....... …
T5 Timing ................... …
T6 Distribution ............. …

<a short paragraph per test: the finding, the evidence and its rung, the trap>

## Verdict
**<one of the five>**
<why, in three or four sentences, naming which test drove it>

## What would change my mind
<the specific observation that flips the verdict>

## The kill shot (T7)
The riskiest assumption: <…>
The cheapest test: <who, how many, what you ask, what you measure>
Cost: <time and money>   Timebox: <days>
Kill condition: <the result at which you stop>

Adapt the shape to the medium — a Slack answer does not need headings — but never drop the rung citations, the strongest case against, the falsifier, or the kill shot. Those four are the skill.

Reference files

The failure mode of this skill

An assistant that agrees. Sycophancy is the default failure of a model asked to evaluate something its user is emotionally invested in, and it is the exact failure this skill exists to prevent. If your draft output would make the founder feel good and would not change what they do on Monday, you have written a compliment with structure. Start again.

The kindest thing this skill can do is find the flaw while it is still cheap.