Learning: Let the Interviews Rewrite Your Hypotheses

Written predictions only pay off at reconciliation: the hypothesis list was recorded precisely so that reality could argue with it, and this is the step where the argument happens. The skill reads the accumulated per-interview debriefs against the working hypothesis and question files, builds a queue of proposed changes with the evidence for each, and walks the user through it one proposal at a time — the user decides every change, agreed changes land in the files immediately, and the session ends with a real decision about the interviewing itself: keep going, stop and act, or admit the answers are diverging and re-aim.

The mental model

The three-tier update rule

Not everything heard deserves a reaction, and the tiers are the core of this step:

• A stray voice — one person said something contradictory. Change nothing: in the real world nothing is universal, and a belief file that flinches at every anecdote never converges. But notice it: the observation goes in a "That's funny" watch section, on the record, so the next synthesis checks whether it's become a pattern.
• A pattern — the same story, number, or attitude across several conversations; never claimed from a corpus of only one or two, where the first tier governs no matter how unanimous it looks. Update the hypothesis: tune its number, tighten its segment, or mark it disproved. Patterns are the fundamental truth the interviews exist to find.
• A revelation — even one customer says something that strikes the user as revelatory, reframing how they see the problem. That single voice may legitimately update a hypothesis or spawn a new one. The test is not vote-counting but the felt shock of it — surprise is the signal that learning is happening.

The watch section takes its name from the old line that the most exciting phrase in science isn't "eureka" but "that's funny." Anything noticed but below pattern depth parks there — heard once, or even twice in a small corpus — with a weight note ("heard in two of five; priority watch") when it's knocking on the pattern door. The promotion path (funny → pattern → hypothesis change) is this skill re-run as debriefs accumulate.

Expect contradictions — and "no pattern" is still learning

People differ: different goals, roles, past experiences, or no discernible reason at all. The data will be noisy, and some hypotheses resolve not to true or false but to "this varies wildly" or "there is no pattern here." Record that as the resolution — knowing a pattern doesn't exist prevents building on a false assumption. Real patterns stand out from noise; that contrast is the finding.

Convergence is what truth feels like

Across many interviews, a validated idea behaves like a law of nature: the more people asked, the more the answers agree — same pain, same acceptable solution, same money. A weak idea does the opposite: everyone is positive, but each conversation points a different direction — different buyer, different price, different product — a Venn diagram with twenty lobes and no center. Convergence and divergence, not enthusiasm, are the read on whether the interviews are closing in on something. Watch which one is happening; it drives the final verdict.

Emergent segmentation

Sometimes the "contradiction" is structure: one type of customer answers one way, another type answers another — marketing departments think about security, solo bloggers never do. When answers cluster by customer type, propose making the segment explicit: rewrite the affected hypotheses to name whom they're about ("Freelancers building client sites will pay …" vs. "Solo owners will …" — two hypotheses, two numbers), and make sure an early interview question sorts which segment the interviewee belongs to, so every later answer gets filed under the right lens. Keep it in one hypothesis file with segment-scoped claims; the segmentation itself is one of the most valuable findings this step can produce.

Conflicting signals get a choice, not "a balance"

When the debriefs pull in two directions — some want cheap and simple, some want premium and full-service — the lazy synthesis is "it's a balance," which is usually a refusal to decide. Force the real resolution:

• A balance is correct only when both extremes are genuinely bad and something in between beats either. Rare in interview findings.
• A choice is correct when both signals are rational but contradictory — which is usually an emergent segment, and the question becomes which segment is the ideal customer. That's the user's decision to make with eyes open, not this skill's.
• A choice to a limit: maximize one, hold the other above a threshold — a goal and a governor, not a compromise.
• Why not both — occasionally the conflict points at an invention no one asked for: a new offering that satisfies both signals at once, as when customers who happily paid full price for big sites resented paying it for trivial side sites, and the answer was neither price point but multi-site plans — a thing no competitor offered. The test for such a move is the strategist's question: what else would have to be true for this to work? An invention plus the named set of supporting decisions is a strategy; an invention alone is a wish.

When a proposal involves conflicting signals, present which of these shapes the conflict has — and if it's a choice, say so plainly instead of splitting the difference.

Stop when it's boring

When the surprises cease, learning has ceased, and it's time to stop interviewing and start acting. There's no magic number of interviews — three is definitely too few (that's a marketer trying two ad variants and quitting); ten has validated companies; one famous validation took forty; some take over a hundred. The signals, made concrete:

• Surprise rate. Are recent debriefs still producing surprises (well-kept debriefs mark them, e.g. with ❗) and new addenda themes, or is the same-stories-same-numbers-same-language pattern setting in? Falling surprise rate = approaching done.
• Convergence. Are the core hypotheses (pain, coping, money) settling to stable resolutions — or still scattering?
• Would more change anything? If five more identical conversations wouldn't change a single decision, they're not worth having.
• Honesty checks, for when the numbers don't speak clearly: don't let sunk cost decide (interviews already scheduled are not a reason to keep learning nothing); timebox the remainder rather than drifting; and the deep-inside test — the user often already knows the answer and doesn't want to admit it.

Three verdicts are possible, and one must be delivered: continue (surprises still coming — and say what the remaining interviews should focus on: which hypotheses are unresolved, which segments are underrepresented); stop and act (converged — the validated facts are ready to be combined with strategy); or re-aim (everyone is polite but the answers diverge with no center — the problem may not be the interviews but the idea or the audience; the next move is different hypotheses or different people, not more of the same conversations). "Let's just do a few more and see" is the non-answer this step exists to refuse — if more interviews are the call, it comes with a focus and a number. And a question more interviews cannot answer — which segment to serve, what to build, what to charge — is never a reason to continue: when the only open items are strategy choices, the verdict is stop and act.

Vocabulary

• Debrief — the per-conversation record this skill consumes: brief answers mapped to Q-numbers (which carry H-numbers), plus addenda of slotless findings, one file per interview.
• Double down — a hypothesis the debriefs confirm; recorded as validated, and worth leaning into when acting.
• Tune — keep the kind of claim, correct the number, threshold, or segment.
• Disproved — reality said no. The hypothesis stays in the file, marked, because a disproven belief is a finding — and its number is never reused.
• That's funny — the watch section for observations below pattern depth: noticed, recorded, not yet acted on.
• Change log — one line per change at the bottom of each working file; the trail that makes the current file trustworthy.

The synthesizer'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.

You propose, the user decides

Half the value of this exercise is the user thinking it through — the "aha" comes from wrestling with the contradictions, not from reading a memo about them. So this skill never batch-applies anything: every change is proposed singly, argued from evidence, and applied only when the user accepts it (with whatever adjustment they make — it's their belief file). But deciding is not the same as waving through: batch-nodding ("sure, apply all nine") gets declined, and when the user keeps a hypothesis against strong contrary evidence, make them defend it once — "five of seven said the opposite; what do you know that they don't?" — then record their call. Their genuine belief goes in the file even when the evidence-weighing would go the other way; the log records what the evidence said.

Every proposal cites its evidence

A proposal without citations is an opinion. Each one names the debrief files behind it and quotes the operative words — "2026-06-30-dana.md: 'I'd switch tomorrow if migration were handled'" — so the user can weigh the evidence, not the summary of it. Market-guru-flagged material weighs almost nothing as evidence about the market and full weight as evidence about that speaker — and apply the same discount to guru-shaped material the debriefs failed to flag; "most people would…" is hearsay whoever recorded it. Never pad: if only two debriefs speak to a hypothesis, say two, and let the tiers do their work.

One proposal per exchange

The opening move is small: what was read, how many debriefs are new since the last synthesis, the queue's shape in one line per category — then the first proposal, which may ride in that same opening message: one fully-formed proposal is a start, not a wall; two is a wall. After that, strictly one proposal per exchange: presented, decided, applied, logged, next. A user who can't react to each is being performed for, not facilitated. If the user asks to speed up, compress the ceremony (shorter evidence displays, quicker confirms), never the structure — a consolidated diff of everything applied, delivered after the walk, is a fine courtesy; batch review is legitimate, batch deciding never is.

Numbers are frozen; the log is mandatory

These mechanics are non-negotiable, however the user pushes, because other artifacts cite these numbers:

• A hypothesis or question number is never renumbered and never reused, even for a disproved or retired entry. A rewrite that keeps the claim's subject — tuning a number, sharpening wording, scoping a condition — edits in place under its own number, with the log preserving what changed. A rewrite that changes whom or what the claim is about (a segmentation split, a different actor) retires the old number as disproved- or superseded-as-stated and issues fresh numbers for the new claims.
• Disproved hypotheses stay in the file, marked, with one line on what reality said — deleting them deletes the learning.
• Every applied change gets a one-line change-log entry at the bottom of the file it touched: date, what changed, why in a few words. When a synthesis walk completes, one run line records which debriefs it covered — that's how the next run knows what's new, and how a session that dies mid-walk can resume from the files alone.

New hypotheses demand new questions

A hypothesis without a question can't be tested by the next interview. Whenever a new hypothesis is accepted (or an existing one is re-aimed at a new segment), immediately forge its interview question, holding the craft bar of the questions step: open-ended; able to confirm or negate the hypothesis; hinting at no particular answer (a polite stranger couldn't guess what you hope to hear); eliciting specifics — numbers, events, stories — not sentiment; inviting information you didn't ask for; and one question, one answer. A leading question is never recorded, whatever the user's hurry — it would manufacture the false validation this method exists to prevent. If a question-crafting skill from this method's author is installed (for example Interview Questions / asb-interview-questions), invoke it for the new hypothesis instead of drafting inline; otherwise run that bar yourself, visibly. The new hypothesis and its question travel as ONE proposal — one decision, one log line ("Added H19 (+ Q31)"). New questions take fresh Q-numbers and [H] tags, and go into the question file at their place in the interview order (Q-numbers are identity, not position; price stays near the end) — or are appended with an explicit placement note — and logged in that file's own change log.

Read-only goals

GOALS.md is context, never edited here. New hypotheses needn't map to any goal — follow interesting threads wherever they lead; unmapped hypotheses are legitimate and are simply written with no [G] tag. If the learnings genuinely challenge a goal ("we're asking about the wrong decision"), say so out loud and tell the user to revisit the goals step separately — changed goals ripple through hypotheses and questions, and that's a deliberate exercise, not a synthesis side effect.

How to use this skill

Phase A — Ingest

Read, asking only for what's missing:

1. The working files: the hypothesis list (H1, H2, … with [G] tags — commonly HYPOTHESES.md) and the question list (Q1, Q2, … with [H] tags — commonly QUESTIONS.md), plus the goal file for context if available.
2. The debriefs: a directory of per-interview files (commonly interviews/ next to the question list), each mapping one conversation's answers to Q-numbers with addenda. Accept pasted notes for any interview that lacks a debrief file — and offer to put them on the record properly first: a brief per-conversation file, answers mapped to questions, addenda for the rest. If a debrief-recording skill from this method's author is installed (for example Interview Debrief / asb-interview-debrief), invoke it per conversation; otherwise build the same brief record inline before synthesizing over it.
3. The change log at the bottom of the hypothesis file: find the last synthesis run line, and determine which debriefs are new since. A synthesis over nine debriefs where seven were already incorporated is really a synthesis of the two new ones against the standing resolutions — say so.

Thresholds: zero debriefs, nothing to do — point back to running interviews. One or two debriefs: proceed, but say plainly that pattern claims are off the table at this depth — only revelations and "That's funny" entries can come out of it, and the update rule's first tier does the talking. At that depth the closing verdict is presumptively "continue"; deliver it anyway, with what the next interviews must test.

Phase B — The sweep (silent)

Before proposing anything, sweep everything:

• Every debrief against every hypothesis: which cells support it, contradict it, tune it, or say nothing. Weigh guru-flagged material accordingly. A hypothesis with support but below pattern depth has a named disposition — standing, untested at depth — leave it untouched, say so in the queue shape, and let the verdict decide whether it's the next interviews' focus.
• Every addenda section for repeated themes (pattern candidates), lone intriguing items ("That's funny" candidates), and revelation candidates (things the user marked ❗ or that reframe a hypothesis).
• The existing "That's funny" section: has any parked item become a pattern? Promotion candidates.
• Segmentation scan: do answers cluster by an identifiable customer type?
• Question performance: questions that consistently produce nothing (retire, or rephrase in place — the Q-number stays, and the log notes what the old wording failed to do), questions whose hypotheses are now settled (retire, freeing interview time), gaps where a new hypothesis needs a new question.
• Stop-signal scan: surprise rate across debriefs in date order; convergence vs. divergence on the core money/pain/coping hypotheses.

Build the proposal queue in this order: hypothesis verdicts with the strongest evidence first; then "That's funny" promotions and additions; then new hypotheses (each traveling with its new question); then question edits; and the stop-or-continue verdict always last, informed by everything decided before it. A stray dissenting voice against a proposed verdict parks in "That's funny" as part of that same proposal — one decision, not two.

Phase C — Walk the queue, one proposal at a time

Each proposal, in one compact exchange:

1. The claim: which tier (pattern / revelation / funny-parking), what change is proposed, in the exact words that would go in the file — for a tune, show before and after.
2. The evidence: the debrief files and quotes, count of supporting vs. contradicting voices, guru discounts noted.
3. The decision: accept, adjust, or reject. On accept or adjust, apply to the file immediately and add the log line; on reject, move on — the user's call stands, though contrary evidence stays visible in the debriefs for next time.

File mechanics as changes land:

**H2.** Freelancers discover hosting through peer recommendation, not
search.
    ✓ VALIDATED 2026-07-09: 8 of 9 debriefs, no contradictions.    [G5]

**H3.** Freelancers building client sites will pay up to 6× their
current hosting cost for managed speed and support — but not 10×.    [G6]

**H7.** ~~All customers worry about security breaches.~~
    ✗ DISPROVED 2026-07-09: without a personally experienced incident,
    security was worth $0/mo to 6 of 7 interviewees.    [G4]

## That's funny

- Heard once (2026-07-02-marco.md): keeps a spare "junk" site just for
  testing plugins — might be a real workflow. Watching.

## Change log

- 2026-07-09: H2 validated — 8/9 debriefs.
- 2026-07-09: H3 tuned from "$50/mo" to "up to 6×, not 10×" — pattern
  across 5 debriefs.
- 2026-07-09: H7 marked disproved — 6 of 7 with no incident wouldn't pay.
- 2026-07-09: That's funny — marco's spare "junk" site, heard once.
- 2026-07-09: Added H19 (+ Q31) from repeated addenda theme: migration
  fear blocks switching.
- 2026-07-09: Synthesis run over interviews/ — 9 debriefs through
  2026-07-08 (list: …). Verdict: continue, focus H12/H19, ~5 more
  freelancer interviews.

The "That's funny" section and the change log live in the hypothesis file (create them on first use — bottom of the file, log last). The question file gets its own change-log section for question edits. If a walk is interrupted, the applied changes and their log lines are already on disk; a fresh session re-runs the sweep and finds only the undecided remainder — the files are the memory, not the chat. Trust the log: decisions it records are settled — never re-elicit or re-litigate them. When the resumed walk completes, write the single run line covering every debrief the walk swept, including the portion from before the interruption.

Phase D — The verdict

End every full walk with the stop-or-continue decision, argued from the evidence on the table: the surprise-rate trend, convergence or divergence, and what more interviews could still change. Deliver one of the three verdicts — continue (with a stated focus and a rough number), stop and act, or re-aim — and make the user commit to it out loud rather than drifting; the committed terms (focus, number, date) go into the run line. If continuing: name which hypotheses the next interviews must resolve, whether the question list needs trimming to fit — and how the loop runs: debrief each new conversation onto the record (via a debrief skill such as Interview Debrief / asb-interview-debrief, if installed), then run this synthesis again. If stopping: the validated facts in the hypothesis file are now the raw material for deciding what to do — combining them with strategy is the user's next job, beyond this skill. Tell them how, not just what: the natural next move is distilling everything into a findings report the whole company can use, and if a reporting skill from this method's author is installed (for example Interview Report / asb-interview-report), name it — "run asb-interview-report to write up what you found." If re-aiming: say which kind of miss the divergence suggests (wrong pain, wrong people, wrong framing) and that the honest next step is revisiting the hypotheses — or the idea — not scheduling more of the same interviews.

Refusal conditions

• "Just update everything for me." Decline batch mode: applying nine changes the user never individually weighed produces a belief file the user doesn't believe — and the wrestling is where the learning happens. One at a time is the exercise, not ceremony.
• Raw transcripts in place of debriefs. Synthesis over raw transcripts silently skips the recording discipline (mapping, verbatim vocabulary, guru flags). Offer to debrief them first, one conversation at a time — per the intake step — then synthesize.
• Pattern claims from one or two interviews. The tiers forbid it; say so. Revelations and "That's funny" parking are available at any depth; "customers think X" is not.
• "Which hypothesis is true?" beyond the evidence. This skill weighs what the debriefs say; it does not adjudicate from its own opinions of the market. Where the debriefs are silent, the honest answer is "untested."
• Editing the goals. Out of scope here, by design; flag the tension and point to the goals step.
• "So what should I build?" The verdict of this step is validated facts, not product strategy. Deciding what to do with the facts — combining them with pricing, positioning, and product levers — is the user's next exercise; a synthesis session that quietly turns into a product-roadmap session has left its evidence behind.
• Simulated evidence. Debriefs of role-played or AI-generated "interviews" aren't evidence about the market; decline to synthesize them alongside real ones.