hypotheses

/hypotheses makes your bets explicit. Every idea silently assumes things — that the pain is real, that people will pay, that it hurts now — and this skill turns those into short testable statements, then keeps score as evidence arrives.

​

When to reach for it

• “What am I actually assuming is true here?”
• You catch yourself saying “I think users will…”
• Prepping for interviews (scripts are built around hypotheses)
• “How are my hypotheses looking? Anything ready to confirm?”
• Retiring bets that no longer apply after a pivot

​

First-time generation

Rather than asking “so, what’s your biggest assumption?” (which makes everyone freeze), the advisor does the pattern-recognition itself: it reads your idea, names an assumption baked into it, and asks you to react — one bet at a time, working toward a starter set of 3–4. It then opens the floor (“anything that keeps you up at night about this idea?”), checks the set for blind spots — most founders under-specify willingness-to-pay and urgency — and offers an optional web-research pass to sharpen specific bets. You confirm the whole set in one go, and the exit handoff names each bet’s type, how it’s best tested, and which single bet is highest-stakes — the one to test first. Each hypothesis gets a type tag that drives how it should be validated — #problem via conversations, #solution via a prototype, #willingness_to_pay via a pricing gate, #urgency via behavioral signals. See The evidence graph for the full table.

​

State assessment: how are my bets doing?

The skill’s core capability. Whenever you ask for a health check (and automatically after each interview is analyzed), a bias-isolated assessment re-reads every statement linked to each hypothesis across all your interviews and weighs the evidence — distinct interviews counted, framing-induced agreement discounted, cross-interview support required before any status flips. What you get back, per hypothesis: what changed → next action — a one-line delta since the last assessment plus the smallest observable next validation move (“show three freelance designers the one-screen mockup and watch whether they try to add a client unprompted”). Plus one cross-hypothesis Top pick: the sharpest move on the board right now. When unlinked statements from different interviews rhyme, the assessment also proposes brand-new hypotheses you didn’t know you were testing. Status changes and new hypotheses need your per-item confirmation. The advisory bookkeeping (last_assessed, ## Next Action) is written automatically so your hypothesis folder stays a live dashboard.

​

Everyday management

Refining a statement, re-tagging, adding a bet mid-conversation, reviewing the set grouped by status, updating status manually, archiving after a pivot (reversible, with a reason noted) — all handled inline, always propose-before-write.

​

What it writes

​

Good to know

• Good hypotheses are testable, specific, and consequential — “people have this problem” gets pushed back on until it names who, what, and what changes if it’s wrong.
• ## Next Action is overwritten on every assessment — it’s the latest read, not a history.
• Works best once core.md has at least an Audience and a Problem; the skill will suggest firming those up first, but won’t refuse.
• Assessment deliberately doesn’t happen in your chat — the whole point is that it can’t be swayed by your enthusiasm.

​

Related