Startup board Pro

Most AI validation tools hand you a report. This one hands you an interrogation, a score and a verdict. Then it remembers what you said and holds you to it next session.

It's a Claude Skill. Plain Markdown that Claude reads as instructions. Nothing to run, nothing to sign up for.

Everything you say gets a score, 0 to 3

Your opinion is a 0. A forum thread is a 1. One customer telling you what they actually did last Tuesday is a 2. Someone paying, switching tools or signing something is a 3.

Ten people saying they'd use it is still a 0. It's the same claim ten times.

You don't get the document until you've earned it

A Lean Canvas needs a 2. A traction slide needs a 3. If you're under the gate you get the test that would get you past it, not the document.

Ask for it anyway and you'll get it, with a stamp on the first line spelling out exactly what's missing.

One question per turn

No numbered lists. No "tell me X, Y and Z". It reads your answer, then asks the next thing.

And it won't compliment you. No "great question", no "strong insight", none of the padding that usually stands in for analysis.

Four verdicts

PROCEED. PIVOT. TALK TO MORE CUSTOMERS. KILL.

Each one arrives with the riskiest thing you haven't tested and one action for this week.

It remembers

There's a state file holding your assumptions, your evidence, your numbers and whatever's still open. Session four doesn't start from the top.

What it won't do

Shorter list. Probably the more useful one.

It won't make up a number. No market sizes, no competitor counts, no conversion benchmarks pulled out of the air. Every figure either came from you, carries a source, or says "(rule of thumb)" right where it's used. Anything it writes into a file is marked UNVERIFIED until you source it.

It won't take enthusiasm as evidence.

It won't run a fit check-in when you have zero active users. Nothing to measure.

It won't size a market for a problem nobody has confirmed. You get a bottom-up count instead.

And it won't quietly reword an unverifiable claim on a pitch slide. The claim gets cut.

Don't buy this if

You want validation. It's built to withhold that.

You need something to show an investor on Friday. At week one the honest output is a list of people to call.

You have no way to reach customers. It'll keep asking how you plan to, because that's the gate.

What you're getting

Router skill	SKILL.md, 172 lines, loads only what a turn needs
Phases	11, numbered 00–10, intake through scale
Advisor modules	9 reference modules shaping the questions
Artifact templates	15, each gated on an evidence score
Evidence rubric	0–3, worked examples at every level
Fit ladder	six stages, each with a testable pass condition
Test suite	28 blind-graded scenarios with answer keys
Static checks	612 assertions over the skill's own files

Those 9 advisor modules carry the thinking of two dozen operators and writers. You'll never see a name. The influence shows up in the question you get asked next, not in a quotation.

The tests ship with it

A Markdown skill is instructions to a model, and the model keeps changing underneath it. Something that reliably produced one question per turn in October can start producing three after an update, and nothing in the file changes to warn you.

So there's an eval suite. 28 scenarios, blind graded: a separate model plays the advisor with the skill files and nothing else, then its reply gets marked against an answer key it never saw. The cases and the keys both ship with the product, so you can read what it's tested against and re-run it yourself.

check.sh adds 612 static checks over the files. That every phase states a numeric gate. That no phase invents a refusal the fit ladder doesn't sanction. That an artifact's required score agrees everywhere it's mentioned. That nothing unsourced about markets or investors gets stated as fact.

What it isn't

Not a prompt, not a wrapper around a search tool. No telemetry, nothing phoning home, no server. Your conversation goes wherever your Claude conversations already go.