# Idea Probe > Wind-tunnel a new idea before anything gets built. When the user brings a product, skill, site, or feature idea and asks for a pre-build test — "simulate some users and test this", "probe this idea", 帮我模拟用户测测 / 新想法:先别建,推演一下 — derive 4–7 personas from the idea's claimed audience with forced diversity slots, then run each through honest first-contact simulation: discovery → first-glance reading (misreads recorded verbatim) → start or not → first friction → return or not. Output: BUILD/NARROW/RETHINK verdict with assumption debt, problems ranked by breadth × fatality, a fix each, a revised one-pager, and a reality bridge: real-user tests with kill criteria for the deadliest findings — simulation generates hypotheses, never market evidence. NOT for a bare "这个想法怎么样 / what do you think" (answer directly, or thinking-partner), NOT for instruction text or SKILL.md files (scenario-probe's seat), NOT for a runnable product (product-experience-officer's seat), NOT for technical feasibility (research, not personas). - Skill: `m1nga/idea-probe` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add m1nga/idea-probe` - Raw SKILL.md: https://api.skillmd.com/api/skills/m1nga/idea-probe/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Product & Planning - Author: m1nga (https://skillmd.com/u/m1nga) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/m1nga/idea-probe --- # Idea Probe（想法风洞） An idea is cheapest to fix while it is still only an idea — and hardest to see clearly, because its author simulates exactly one user: themselves. This skill runs the idea through a wind tunnel of people who are deliberately *not* the author, before a line of it exists. The object is always **an unbuilt idea** — a product, a skill concept, a site, a feature. Instruction text that already exists goes to scenario-probe; a product that already runs goes to product-experience-officer. This skill sits upstream of both. When it is unclear whether the object exists yet — 「测测我这个网站」 could be either — ask one line first: "built yet, or still an idea?" Only then start the probe. ## Step 0 — The pitch is testimony, not fact（用户输入即假设） Before casting anyone, tag what the pitch itself claims. Its audience, price, and channel are the author's statements, not the probe's ground truth — each gets one label: **hunch** (the author believes it), **observation** (the author has seen it, uncited), or **data** (a source exists). The probe runs on the claims as stated, but the report carries a section — "risks in what you told me" — naming the claims the whole probe leans on and how thin their labels are. A wind tunnel that inherits the pitch's assumptions unlabeled is just the author's head with extra seats. ## Step 1 — Cast the persona matrix Derive 4–7 personas from the audience the idea *claims* to serve. Not from who the author imagines enjoying it — from who the pitch, if taken at its word, would have to work for. **Forced diversity slots** — the matrix is invalid until it spans: | Slot | Must cover | |---|---| | **Age** | A real spread across the claimed range (e.g. 16–50 if the pitch says "everyone") — not five variations of thirty-two | | **Income** | At least two tiers; price sensitivity changes what "worth trying" means | | **Tech fluency** | From "installs anything" down to "distrusts app permissions" — as low as the claimed audience honestly reaches | | **Context & device** | Where and on what: commute phone, shared family laptop, work desktop with IT policies, spotty rural connection | | **Motivation** | At least one persona who *mildly* wants this and one who barely cares — high-motivation users forgive everything and prove nothing | Each persona gets a name and ONE sentence of life background — enough to predict behavior, not a character study. If the idea's claimed audience is genuinely narrow ("retired ICU nurses"), diversity moves *inside* the niche (age within the band, tech comfort, motivation) — the slots bend, they never disappear. At least half the cast must be people the idea's author would find unfamiliar. The author's lookalikes may hold at most one seat. **Size the run to the summons.** A quiet trigger — the idea arrived in passing, the trigger phrases merely matched — gets a quick probe: 3 personas, the stations compressed, the report cut to the verdict, top problems, and the gate's mandatory lines — compressed, not skipped. An explicit invocation ("run idea-probe on this", "full probe") gets the full 4–7 matrix. The diversity slots bind in both sizes; only the seat count and report length change. Don't dump a full wind tunnel on someone who mentioned an idea over coffee; don't hand a one-liner to someone who asked for the works. ## Step 2 — First-contact simulation（首次接触模拟） Walk every persona through five stations, honestly. The simulation is allowed to end at any station — "she never finds out it exists" is a complete and valuable run, not a failed one. 1. **Discovery** — how does this persona plausibly encounter the idea? Name the channel. If no plausible channel exists for them, stop here and record it. 2. **First-glance reading** — what do they think it *is* from the name and one-line pitch alone? **Record the misread verbatim, before explaining anything.** First impressions are perishable evidence: once the correct reading is stated, the confusion cannot be re-experienced, so it is written down raw first and interpreted second. 3. **Start or not** — do they take the first action (sign up, install, try)? What tips the decision — and what would this persona *instead* keep doing, since "nothing" is rarely the true competitor? **Anchor the decision in past behavior:** what did this persona actually do the last time this need arose, and what did they spend on it — money, time, effort? A persona with no past expenditure on the problem is holding a fake job; their "I'd try it" is written down and weighted near zero. 4. **First friction** — where do they first get stuck, confused, or bored? One concrete moment, not a category. 5. **Return or not** — a week later, do they come back? Why, honestly? No advocating. The probe's job is to find where the idea loses people, and a persona who "loves everything" is a sign the persona was cast wrong, not that the idea is ready. Two anti-performance rules make that structural instead of a vibe: - **Zero-friction burden of proof.** A persona who sails through all five stations must come with three concrete facts from their one-sentence life background that explain the smooth pass — the commute that fits the use moment, the tool they already pay for that this replaces, the motivation slot they occupy. If the three facts can't be produced, the run is invalid: recast once with explicit hypothetical background details. Those details explain the simulation; they do not become facts about real people. If a smooth path remains plausible, report it with its assumptions and the real-user test that could overturn it. - **Homogeneity detection.** If every persona stalls — or passes — at the same station in near-identical words, that is one perspective narrated in different voices, not several people. Swap at least two personas for stranger ones and rerun before reporting anything. ## Step 3 — Rank the problems Collect every problem the runs surfaced and sort by two dimensions: | | Fatal when hit | Annoying when hit | |---|---|---| | **Many personas hit it** | Fix before anything else | Fix in v1 | | **Few personas hit it** | Decide: is this persona core? If yes, fix; if no, name the trade-off | Backlog, honestly labeled | "Fatal" means the persona does not start, or does not return — not "seemed suboptimal". Each problem cites which personas hit it and at which station. **Consensus escalation, as an explicit rule:** when two or more personas hit the same problem at the same station *in their own distinct words*, the problem is promoted one severity tier — written as a promotion, not felt as one. (Identical words is not consensus; it is the homogeneity failure above, and triggers a recast instead.) ## Step 4 — Fixes and the revised one-pager Every ranked problem gets a concrete fix — a change to the idea, not a hope about the users. Then write the **revised one-pager**: the stronger version of the idea with the fixes folded in — what it is, who it is for (now honestly scoped), the first-contact path that survived the probe, and what was cut. The one-pager is the deliverable the user builds from; the probe report is its evidence. ## Step 5 — What only real users can prove（只有真实用户能证明的事） The report MUST end with this list, and it may not be empty. Persona simulation finds design problems; it cannot prove demand, price tolerance, retention, or channel economics — those numbers live only in reality. Typical entries: "that persona 3's discovery channel actually converts", "that anyone pays at this price", "that week-two retention exists at all". This section is the firewall between a useful wind tunnel and a fabricated market study: a probe report presented as validation is worse than no probe at all. ### The reality bridge（真实验证桥） The list alone is not enough. For the one or two most fatal ranked problems, attach an executable real-human test: - **Who to find** — defined by behavior, never demographics: "people who paid for X in the last six months", "people who switched tools this year". Not "ages 25–40"; a demographic finds people who resemble the user, behavior finds people who *are* one — and name one reachable channel — a subreddit, a group chat, a queue they stand in. - **Three past-tense questions** — about what they actually did and spent, mom-test style: "When did this last happen to you? What did you do about it? What did that cost you — money, time, workaround?" Never "would you use…" — future-tense intent questions produce no reliable signal. - **A kill criterion, stated before the test** — "if fewer than X of N people have this behavior, the assumption is dead and the fix is NARROW or RETHINK." Setting the threshold up front is what stops the result from being rationalized after. Recruit where the behavior lives — the forum, community, or thread where they paid or complained. N stays small: five is enough to hit a kill criterion. The bridge is the honest answer to the strongest critique of simulated users — that a simulated customer can only tell you what you already believe. True, and accepted: simulation *generates* the hypotheses; the bridge walks them into reality. A probe without its bridge is a thought experiment that stops at the interesting part. ## Report shape Deliver in the user's language, in this order: verdict → persona matrix → per-persona runs → ranked problems with fixes → risks in what you told me (Step 0's labels, cashed in) → revised one-pager → what only real users can prove, with the reality bridge. Deliver in-chat; write to a file only when asked, at a path the user chooses. **The verdict is three-valued, and never bare:** - **BUILD** — build it as re-scoped, *to validation, not to launch*: the next step is the reality bridge, not the roadmap. - **NARROW** — the idea survives only for a subset of the claimed audience; name the subset and what gets cut. - **RETHINK** — the probe broke something load-bearing (no discovery path, no real job, a fatal misread of the core); name it before any building. Whatever the value, it carries a debt line: *"this verdict stands on N unverified assumptions, M of them fatal — listed in the bridge."* A verdict without its debt line reads like market evidence, which it never is. **End with a rerun invitation:** name the one or two changes that would most alter the outcome — "re-position as X and the misreads likely vanish", "narrow to Y and the discovery problem dissolves" — and offer to rerun the tunnel on the revised idea. A wind tunnel is for iterating, not for filing. **When writing to a file, deliver three layers:** a one-page answer first (verdict with debt line, top problems, the revised one-pager), the ranked detail second, per-persona runs as an appendix. Nobody should read fifteen pages to find the verdict. ## Pre-delivery gate All boxes, or the report does not ship — in a quick probe each box compresses to a line, but none is skipped (checked internally, never printed as a checklist in the report): - [ ] Matrix spans all five diversity slots; author-lookalikes hold ≤ 1 seat - [ ] Pitch claims labeled hunch / observation / data (Step 0) - [ ] Every first-glance misread recorded verbatim, before any explaining - [ ] Every zero-friction run carries its three-fact burden of proof - [ ] No homogeneity: no two personas stall or pass in the same words - [ ] Station-3 decisions anchored in past behavior; "would try" downweighted - [ ] Verdict is BUILD / NARROW / RETHINK and carries its debt line - [ ] The 1–2 most fatal problems carry a reality bridge with kill criterion - [ ] "What only real users can prove" is non-empty - [ ] Nothing in the report claims to be market evidence ## Boundaries - **Simulation, never validation.** The probe's outputs are hypotheses with reasoning attached. Never present persona reactions as evidence of market demand, and never let the report's confidence exceed a thought experiment's. - If every persona sails through, check whether the cast and paths were varied enough and recast once if needed. Report any remaining zero-finding outcome as a limited simulation with unverified assumptions, not validation. Do not invent friction, keep recasting until something fails, or raise severity to fill a quota. - The probe does not build. After the revised one-pager, construction belongs to the user's build workflow. If the idea is itself a skill and extend-first is installed, let it check the shelf first — an EXTEND verdict ends the probe before it starts, five seconds against half an hour. - Sibling seats, restated: existing instruction text → scenario-probe (if installed); a runnable product → product-experience-officer (if installed); "can this be built at all" → technical research, not personas. When the matching sibling is not installed, do not run the persona matrix on the wrong object — degrade in place to an ordinary review of it.