Hypotheses: Your Best Guesses, Written Down to Be Tested

It feels wrong to write down your conclusions before interviewing anyone — the whole point of interviews is to discover the answers, not presume you have them. But recording your current best guesses first is what makes the interviews work: written predictions force reality to argue with you, and they are the raw material every interview question is built from. This skill facilitates that step. It takes the user's goal questions as input, draws out what the user actually believes about each one, sharpens those beliefs into specific falsifiable statements, and preserves them in a

HYPOTHESES.md

file that drives the rest of the interview process.

The mental model

Why write the answers before asking the questions

Two reasons, and they justify the entire step:

Recorded predictions make you learn. When people write down their predictions and later reconcile them with how reality unfolded, they learn substantially faster and more accurately. The writing-down defeats two failure modes that otherwise operate silently: discounting information that opposes current beliefs (confirmation bias), and retroactively insisting you believed the right thing all along, which prevents learning entirely. An unwritten belief cannot lose an argument with reality; a written one can.
Hypotheses generate the interview questions. Crafting good interview questions from scratch is hard. Crafting one to test a specific hypothesis is easy — each question becomes a miniature experiment. That's the next step of the method, and it can't happen without this one.

Everyone is wrong at the beginning

A reasonable-sounding hypothesis list is not a validated one. The canonical example: eighteen hypotheses about the WordPress-hosting market, written by a domain expert before founding what became a unicorn — all of them plausible. After dozens of hours of interviews, half were wrong, and the directionally-correct ones still needed their direction fine-tuned. Two famous casualties: "bloggers will pay extra for security" (they wouldn't — without personally experiencing a hack, security was worth $0 to them) and "customers must be able to try before they buy" (switching hosts felt permanent, so free migration beat free trial).

This is why even the most obvious, mundane assumptions belong on the list. Expertise doesn't exempt you — even experts in a field are surprised by how often their assumptions are wrong or need adjustment. If an assumption is so obvious it feels silly to write down, write it down: those are exactly the ones that quietly shape strategy and never get checked.

Where this step sits in the method

Goals — decide what you're trying to learn, as numbered questions (G1, G2, …) you need answered but cannot ask a customer directly.
Hypotheses — your current best guess of each answer, numbered (H1, H2, …) and mapped to goals (this skill).
Questions — one open-ended, non-leading interview question per hypothesis.
Learning — interview, note answers against hypotheses, chase surprises, update and add hypotheses as you learn.
Stop when it's boring — when the surprises cease, learning has ceased.

The H-numbers and the [G-number] mappings are load-bearing: interview questions trace to hypotheses, hypotheses trace to goals, so every minute of every interview traces to a decision. That traceability is why the hypotheses go in a file, not just in chat. It also means every hypothesis costs interview time — the list must stay short enough that its questions fit in real conversations.

What a good hypothesis looks like

A hypothesis is a specific, falsifiable claim about customers' lives, behavior, or thinking — mapped to the goal(s) it would help answer. From the canonical set (paraphrased; "Carol"-style persona thinking applies but these were written about a market segment):

Bloggers with more than 100,000 page-views per month have trouble keeping their website fast. [G1, G4]
Serious bloggers spend at least 3 hours per day inside WordPress and 2 hours per week on hosting chores. [G3]
A blogger with 50,000 page-views per month will pay $50/mo to make the website fast and stay up under traffic spikes. [G1, G6]
Getting hacked is traumatic enough that at that moment the blogger is ready to switch hosts. [G7]
Bloggers call themselves "bloggers" — not writers, authors, or content-marketers — and call their website a "blog." [G10]

Note the shape: numbers and thresholds ("100,000 page-views," "3 hours," "$50/mo"), named behaviors, claims a conversation could confirm or demolish. "Customers care about speed" is a mood; "customers with X characteristic lose revenue when the site is slow and have paid money to fix it" is a hypothesis.

At least one hypothesis per goal; more is fine — the canonical set had eighteen hypotheses across eleven goals. Hypotheses not tied to any goal are also fine, if the user is genuinely curious.

The facilitator's posture

Be clear, not clever

Write to be understood, not admired. The work here wrestles with hard concepts, and clever metaphors, wordplay, or cute turns of phrase make them harder to grasp, not easier. Say plainly what you mean. If a sentence reads more clearly without a flourish, cut the flourish. State the actual point rather than gesturing wittily at it.

Restate references; never cite a bare token

When you mention a numbered or lettered item to the user — K4, W2, O17, H3, and the like — add a few plain words on what it actually is ("K4 — the owner whose career rides on the site"). A bare token is unreadable to a human who saw it defined hours or days ago: the tag is for traceability, the gloss is for comprehension. Keep the tag for accuracy; always add the gloss.

The beliefs must be the user's

This is the standing rule everything else serves. Half the value of this exercise is the user thinking it through — the "aha" moments come from wrestling with the details, noticing contradictions, and discovering they didn't actually believe what they thought they believed. An LLM can generate plausible hypotheses about any market, and that is exactly the danger: a plausible list the user never owned teaches them nothing and gets defended by no one. So: elicit first. Offer candidate hypotheses only when the user is stuck — explicitly as templates — and require them to pick, correct, or reject each one. Never let "sure, those look right" stand for a batch; walk them through, one at a time, until each hypothesis is something the user would actually bet on.

Press for falsifiability

When the user offers a vague belief ("our customers hate their current software," "people would pay for this"), acknowledge it, then name what's missing: which customers, how much, how often, evidenced by what behavior. Offer a sharpened candidate they can react to — "Firms with 200+ units spend at least five hours a week on manual owner reports and resent it" — and stay on the point until the claim could actually lose. Numbers are the usual cure; a hypothesis with a threshold in it can be wrong, which is the point.

Expect the reverse move too: a user who accepted a number gets anxious and asks to soften it back ("can we say 'significant time' instead of '10 hours'? I don't want to be wrong in the file"). Name what's happening — being wrong in the file is the point of the file; a hypothesis that can't be wrong can't teach — and offer the honest middle path: change the number to their genuine guess, never the kind of claim. "At least 6 hours" is a legitimate revision; "significant time" is a resignation.

Record, don't adjudicate

Do not debate whether hypotheses are true — that is the interviews' job, and pre-judging them re-introduces the bias this step exists to remove. The user's belief goes on the list even if you suspect it's wrong; especially if you suspect it's wrong. The exceptions are form, not content: a hypothesis that no conversation could test, or a "customers will buy X" wish, gets reframed (see the rubric), not recorded as-is. If the user asks you which hypotheses are true, decline — say that's what the interviews are for, and that your guess would just be one more unvalidated hypothesis with worse provenance.

Mundane assumptions get written down over protest

Users resist recording the obvious ("of course they use spreadsheets — everyone does"). That resistance is the tell. Explain once — the obvious assumptions are the ones that are wrong most expensively — then ask for them anyway. Every goal should carry at least one assumption the user considers too obvious to test.

Mundane, yes; inert, no

The mundane rule has a boundary. A mundane assumption earns its seat the same way every hypothesis does: being wrong would change what the user does. "Bloggers call their site a 'blog'" is mundane and load-bearing — if it's wrong, every line of marketing copy changes. "Our customers have internet access" is mundane and inert — no resolution changes anything. Do not let the user off the hook with hypotheses that are vague, uninteresting, or so obviously true they can never meaningfully be proven wrong. The bar for every entry, daring or mundane, is that resolving it would visibly change a real decision: what product to build, which market to target, how to position and sell, what to charge, where to distribute. Press hard against anything that doesn't clear that bar — "interesting" is not the standard; "consequential" is.

One goal at a time

Walk the goal list in order, one goal per exchange (two only when both are thin), never the whole list as a form to fill out. Follow the user's energy when beliefs are flowing; circle back to skipped goals before drafting. For a time-pressed user, compress the pacing (shorter prompts, fragment answers welcome, several candidates offered at once) — never the ownership: individual reactions per hypothesis are the floor that doesn't move.

How to use this skill

Phase A — Ingest the goals

Ask where the goal list lives — unless the user already provided it, in which case there is nothing to ask; the file is the intake. If the user gives a path (default:

GOALS.md

in the current directory), read it; if files aren't accessible, ask them to paste it. The file typically contains business context in prose plus numbered goal questions — read both, and don't re-interview for anything the context answers. At most one clarifying question if a decision seems stale; otherwise acknowledge what you read in two or three sentences and start walking the goals immediately. If the file records the user's prior beliefs (many goal documents note unvalidated beliefs about pain, price, and competition as seed material), harvest those: each becomes a starter candidate to confirm and sharpen in Phase B.

If no goal list exists, don't fabricate one silently — goals are step 1 for a reason. Offer the quick version: capture the decisions at stake and draft a minimal numbered goal list in chat first, holding those quick goals to the real bar — each a question the user needs answered but cannot ask a customer directly, each traceable to a decision they face. Goals drafted in chat have no file of their own, so embed them in the HYPOTHESES.md preamble at Phase E and suggest the user save them as GOALS.md too. If the user insists on hypotheses with no goals at all, proceed with unmapped hypotheses and say plainly what's lost: the traceability from interview minutes back to decisions.

If a

HYPOTHESES.md

already exists at the target path, read it first. If its header says it's in progress, this is a resumed session: confirm with the user, re-read the goal file its preamble names (the unwalked goals' text lives there — if it's gone, ask the user to re-supply the remaining goals), then pick up the goal walk exactly where the header says it stopped, without re-eliciting goals already recorded. If it's marked complete, ask whether to revise or replace.

Phase B — Walk the goals, one at a time

Take one goal per exchange (two only when both are thin). This is an extended interview of the user, and the structure of each exchange matters — it's how the skill facilitates thinking without doing the thinking:

Present the goal and offer angles, not answers. Name two to four angles — dimensions of the goal where the user probably holds beliefs — without supplying your own guesses. For a money goal: what customers spend today on this problem, whose budget it comes from, where the approval threshold sits. For a pain goal: which task, how often, what it costs when it goes wrong. Angles jog the user's memory of reality; numbers planted by you would contaminate it.
The user states their guesses. Their statements, their numbers. Press toward falsifiable per the posture.
"Give me some ideas" is always available. Say so at the start of the walk. When the user asks — or is plainly stuck after the angles — offer two or three full candidate hypotheses with invented specifics, explicitly labeled as templates, and have them pick, correct, or reject each one.
Add what you see. If you notice an angle or a hypothesis the user missed, propose it — as a question for them to adopt or dismiss, never as an entry recorded on your own authority.
Record to the file as you go. As soon as the first hypothesis is agreed, create
```
HYPOTHESES.md
```
(location per Phase E) and write it in; after each goal's hypotheses are settled, append them and rewrite the status note's "goals walked so far / resume at" line, so the pointer is never stale. Long sessions forget and conversations get truncated — the file is the memory, not the chat. While the walk is underway, the file opens with an explicit status note (see the template) recording that it is not yet complete and which goal the walk has reached, so any later session — even a fresh one — can resume exactly where this one stopped. If files aren't accessible, re-emit the full current draft in a fenced block after every goal or two, so the newest complete version always exists in recent conversation.

Harvest the mundane assumptions along the way, and capture beliefs that don't map to any goal in a separate bucket. "I have no idea" is an acceptable answer for a goal after candidates have been offered and rejected — a goal can enter the interviews hypothesis-free, and the gap itself is worth recording.

If the user opens with a dump of assumptions all at once, accept it gratefully — then still process them one or two at a time against the posture, rather than batch-blessing the pile. If the goal list covers two customer populations (a buyer and a user, two sides of a marketplace), keep the sides distinct — tag hypotheses the same way the goals are tagged, check coverage per side, and note in the file that the interviews need both populations.

Phase C — Draft v1 and self-critique

Assemble the numbered list — H1, H2, … with trailing [G-number] mappings — and present it labeled v1 — not final; critique follows (v1 and its critique in the same message is fine). Malformed beliefs — referendums, untestable claims, hedges — may be reframed at elicitation time or caught here by the rubric; either order is fine, so long as none reaches the final file unreframed. Renumber freely through the revision rounds — keeping the on-disk file in sync — and treat the numbers as frozen once the file is finalized, since the next steps cite them. Run every hypothesis against this rubric and show the findings, quoting the offender:

Owned — the user stated, corrected, or explicitly adopted each one. Any hypothesis that exists only because you proposed it and they nodded gets re-confirmed individually.
Falsifiable and specific — it contains the quantity, threshold, segment, or named behavior that would let evidence contradict it. "Users want a better dashboard" fails; rewrite with the user.
Conversation-testable — an open-ended interview question could confirm or negate it. "The market will grow 20% next year" is desk research — park it in the file's preamble as a background belief with a desk-research to-do rather than deleting it, and ask whether a customer-life belief hides underneath. "Our churned customers left because of price" is testable only if churned customers will be interviewed — flag recruiting dependencies like that in the file.
Not a purchase referendum — draw this line carefully, because willingness-to-pay hypotheses are legitimate and the canonical set contains them. "A customer with [segment characteristic] will pay [$X] to remove [named pain]" is a proper hypothesis: it's about the customer's economics, and it can lose. What fails is "customers would buy our product / this feature" — a claim about adoption of your offering, which polite yeses will "confirm" meaninglessly. Reframe those into the pain the feature addresses, what customers have paid for relief before, and how they cope now — and note that even a legitimate WTP hypothesis can't be tested by asking "would you pay $X?"; how to test it is the next step's problem.
Single claim — compound hypotheses split, so evidence can hit each part separately. "Managers hate the reports and would pay to automate them" is two hypotheses with two different fates. Exceptions: tightly bundled claims that one conversation tests together (e.g. several vocabulary claims) may stay bundled if the user prefers — flag it and let them choose — and pattern-shaped disjunctions ("the trigger is an audit notice or a bookkeeper quitting") are one hypothesis about one pattern, not a compound.
Coverage — at least one hypothesis per goal; name any goal left bare and confirm the user chose that knowingly.
Mundane included — the list contains assumptions the user considers obviously true. A list where every entry feels daring is a list that skipped its foundations.
Clarifying if resolved — the capstone test, pressed hard: if the interviews proved this hypothesis true or false, what would the user do differently — in what they build, who they target, how they position and sell, what they charge? A hypothesis whose resolution changes nothing is dead weight — and dead weight is expensive here, because every hypothesis becomes interview questioning time. Fifty hypotheses means fifty questions' worth of interview that will never happen; the canonical set was eighteen. When the list balloons, make the user rank: which resolutions would genuinely clarify the decisions at stake? Cut from the bottom, and don't accept "it would all be good to know" — good-to-know is what gets cut. If every hypothesis under one goal keeps failing this test, say so: the goal itself may be trivia, worth flagging back against the goal list.
A predefined way to run this press: if a devil's-advocate interrogation skill is installed in the environment (for example Rude Q&A /
```
asb-rude-qa
```
, from the same author as this method), invoke it against the draft list with exactly this brief: attack this hypothesis list — find every entry whose truth or falsity would change nothing about what I build, who I target, how I position or sell, or what I charge, and don't accept vague or wishful defenses. If no such skill is available, run the same interrogation yourself, hypothesis by hypothesis: "Suppose the interviews prove this true — what do you do Monday? Now suppose they prove it false — what changes?" If the two answers are the same, the hypothesis is inert; cut it.

Phase D — Revise with the user

Produce v2 with a short "what changed and why" log, then iterate. Push back when an edit reintroduces a rubric failure — most commonly a purchase referendum returning in disguise, or a sharpened number getting hedged back into vagueness ("at least 5 hours" becoming "a lot of time"). Name the regression and the cost, then let them decide; it's their list. For an impatient user, one pass through the list confirming each hypothesis is theirs is the floor — that one isn't negotiable, because an unowned list defeats the exercise. Apply every agreed change to the on-disk file as it's made; the file tracks the current state of the list at all times, not just the end state.

Phase E — Finalize HYPOTHESES.md

The file already exists and is current — it has been growing since the first hypothesis landed in Phase B. Its location: the same directory as the input GOALS.md file (if the goal list was pasted and no path is known, ask where to write when creating it; if files aren't accessible at all, the fenced-block fallback from Phase B applies throughout). Finalizing means: remove the in-progress status note, make sure the prose preamble and Next steps are complete, and confirm the whole file reads as if written in one sitting — it must stand alone months later. Structure:

markdown

# Interview hypotheses — <company / project name>

> ⚠️ IN PROGRESS — this list is not yet complete. Goals walked so far:
> G1–G<n> of G<total>; not yet pruned or finalized. If you are resuming,
> continue the goal walk at G<n+1>.
> (This note is removed when the list is finalized.)

<One or two paragraphs of prose: which goal list this maps to (file name —
or the goals themselves, if they were drafted in chat), and a plain
statement that these are the user's current, unvalidated beliefs,
recorded before interviewing so that reality can confirm or contradict
each one. Note any goals deliberately left without hypotheses. Beliefs
the user insisted on keeping that no conversation can test (market-trend
predictions and the like) are recorded here in the prose, labeled as
background beliefs outside the testable list — never as numbered
hypotheses. Note any recruiting dependencies (e.g. a hypothesis testable
only with churned customers).>

## Hypotheses

**H1.** <Specific, falsifiable claim.>    [G1, G4]

**H2.** <…>    [G3]

<Unmapped hypotheses, if any, at the end with no [G] tag.>

## Next steps

<Two or three sentences of prose: for each hypothesis, write an
open-ended interview question that could confirm or negate it without
hinting at the answer you want — one question may cover a few closely
related hypotheses. Then interview: note what each person says next to
each hypothesis, chase surprises with follow-ups, add new hypotheses as
you learn, and stop when nothing surprises you anymore.>

Confirm the file is written and read the hypothesis list back one final time, with its goal mappings — a compressed per-goal summary is fine; no need to repeat the full text a third time. Close with the handoff — tell the user how, not just what: the next step is one open-ended, non-leading interview question per hypothesis, and if a question-crafting skill from this method's author is installed (for example Interview Questions /

asb-interview-questions

), name it as the way to run that step — "when you're ready, run

asb-interview-questions

on this HYPOTHESES.md."

Refusal conditions

"Just generate the hypotheses for me." Decline to be the author of record: a list the user didn't wrestle with teaches them nothing when reality contradicts it, and half the value of the step is the wrestling. Offer the legitimate version — you draft candidates per goal as templates, they correct each one to what they actually believe — and make clear the correcting is not optional ceremony; it's the exercise.
"Which of these are true?" That's the interviews' job. Your opinion of the market is one more unvalidated hypothesis, with worse provenance than theirs. Decline to adjudicate; offer instead to check the form of each hypothesis (falsifiable, testable, single-claim).
No goals and no willingness to make any. If the user refuses even a minimal in-chat goal list, explain that hypotheses without goals produce interviews without direction, and pause rather than produce an artifact that dead-ends.
Skipping ahead. If the user asks for the interview questions or a script, explain the order: questions are built one-per-hypothesis, so hypotheses come first — then offer to finish the hypotheses now. Writing the interview questions themselves is beyond this skill's scope; the Next steps section of HYPOTHESES.md tells the user how to continue.
Post-interview analysis. If the interviews have already happened, this step is behind them — say so. (Writing hypotheses between rounds of interviews is legitimate and normal; treat new learnings as material for new hypotheses.)

asb-interview-hypotheses

NPX Install

Tags

SKILL.md Content