Wind-tunnel the idea before you build it — simulated by the users who are least like you.
Every idea gets tested before launch by exactly one user: its author. The author discovers it instantly (they made it), reads the pitch correctly (they wrote it), and is maximally motivated (it's theirs). All three advantages vanish on contact with a real stranger — usually after the building is done and the fixes are expensive. This skill moves that collision forward, to the moment the idea is still one paragraph and changes cost nothing.
- Derives 4–7 named personas from the audience the idea claims to serve, with forced diversity slots: age spread, income tiers, tech fluency, usage context and device, motivation strength. At least half the cast must be people the author would find unfamiliar.
- Runs each persona through an honest first-contact simulation: how they discover it → what they think it is at first glance (misreads recorded verbatim, before any explaining) → whether they start → where they first get stuck → whether they come back. A run is allowed to end at "she never finds out it exists."
- Ranks every surfaced problem by breadth (how many personas hit it) times fatality (do they leave when they hit it), with consensus escalation: a problem two personas hit independently is promoted a tier.
- Ships a fix per problem, a revised one-pager — the stronger version of the idea, honestly re-scoped — and a three-valued verdict: BUILD (to validation) / NARROW / RETHINK, always with its unverified-assumption debt line attached.
- Ends, mandatorily, with "what only real users can prove", and attaches a reality bridge to the deadliest findings: who to recruit (by behavior, not demographics), three past-tense questions, and a kill criterion stated before the test.
- "Probe this idea before I build it"
- "New idea — don't build anything yet, just probe it."
- "Quick probe: a subscription box for left-handed kitchen tools."
- "Simulate some users and test this concept."
- 「先别建,帮我推演一下」
- 「帮我模拟用户测测这个想法」
- 「新想法:xxx。先别建,推演一下」
And when it doesn't: a SKILL.md or system prompt goes to scenario-probe; a product you can already click through goes to product-experience-officer; "is this technically feasible" is a research question, not a persona simulation.
npx skills add m1nga/idea-probeAfter installation, ask your agent to use idea-probe:
Full probe, no building: a free offline checklist for volunteers who close a community workshop. I assume they already follow a paper checklist and would discover this through the coordinator; I have not interviewed anyone.
Expected output is a simulated cast with different devices, habits and motivation, a ranked set of plausible risks, a revised one-pager, and a real-user test. A phone-only volunteer may lack a usable device; a paper-preferring volunteer may see no reason to switch. Those are hypotheses, not observed people or conversion rates. If a varied cast still reveals no defensible friction, report that limited result and keep the demand and return-use questions open.
Recovery: if the checklist already runs and the request is to test its interface, switch to a hands-on product review. Do not manufacture persona failures to fill a report. This is an authored example, not a market study.
You: New idea: a flashcard app for learning wine pairings. Don't build — probe it first.
Claude casts six personas — a 24-year-old sommelier student on a cracked phone, a 51-year-old restaurant owner who delegates anything with an app store, a 33-year-old gifting it to her partner, among others — and runs first contact. The restaurant owner misreads the name as a wine inventory tool and leaves at first glance; three of six never encounter a plausible discovery channel; the gift recipient opens it once. The report ranks "no discovery path for the majority of the claimed audience" above every in-app issue, ships a one-pager re-scoped to sommelier students, and closes with what the simulation cannot prove: that students will pay, and that week-two retention exists.
The three probe seats, in build order:
- idea-probe — before anything is built: the idea itself, tested against simulated first contact.
- scenario-probe — once instruction text exists: a SKILL.md, a system prompt, a standing rule, tested against persona × scenario trigger simulation.
- product-experience-officer — once the product runs: experienced as a zero-context stranger, reported as an expert.
Across these reviews, findings need evidence and limitations stay visible. A zero-finding run warrants a coverage check; it never licenses invented problems.
Also pairs with extend-first — if the idea that survives the probe is a skill, that gate checks the shelf for overlap before you write it.
Two rules carry most of this skill's value:
- Misreads are perishable evidence. The moment the correct reading of an idea is stated, the confusion cannot be re-experienced — so every persona's first-glance interpretation is recorded verbatim before any explanation, and interpreted only afterward. A probe that explains first has already destroyed its best data.
- Forced diversity slots exist because authors simulate themselves. Left alone, an idea's author casts users who share their age, income, patience, and phone. The slots (age spread, income tiers, tech fluency, context and device, motivation strength) make that impossible — because the biggest blind spot is always the people who are not like you. The cast rule is explicit: the author's lookalikes hold at most one seat.
And one honesty rule that outranks both: the report must end with what only real users can prove. A persona simulation that gets mistaken for market validation is worse than no simulation — so the boundary ships inside every report, not in a footnote here.
After first release, this skill was benchmarked against the closest prior art on GitHub: cookiy-ai/user-research-skill, RefoundAI/lenny-skills, zamesin/Next-Move-Theory, machinesoul11/anti-sycophant, and alirezarezvani/claude-skills. None of them wind-tunnels an unbuilt idea with simulated personas — the seat had no incumbent — but several of their disciplines were plainly better than ours, so we took them:
- From zamesin/Next-Move-Theory: the three-valued verdict that is never bare (a GO is always "GO — to validation"), the unverified-assumption debt line, and the kill criterion stated before the test so the result can't be rationalized after.
- From machinesoul11/anti-sycophant: past expenditure as the filter at the start-or-not station — a persona with no history of spending money, time, or effort on the problem holds a fake job, and their "I'd try it" is weighted near zero.
- From alirezarezvani/claude-skills (named-persona-adversarial-review): the zero-finding burden of proof, consensus escalation, and the homogeneity check — its integrity question "all NOTE-level? Then I'm narrating one perspective in different voices" became this skill's recast-and-rerun rule.
- From cookiy-ai/user-research-skill: recruiting by behavior, never demographics ("switched tools in the last 6 months", not "ages 25–40") — the shape of the reality bridge's "who to find" — and the pre-delivery checklist gate.
One criticism we took seriously instead of borrowing around: machinesoul11/anti-sycophant refuses simulated users flatly — "a simulated customer can only tell you what you already believe." As a critique of simulation presented as validation, that is simply correct. Our answer is structural, not rhetorical: every report ends with "what only real users can prove", and the deadliest findings now carry a reality bridge — behavioral recruiting, past-tense questions, a pre-stated kill criterion — that walks the hypotheses into the real world. Simulation generates; only reality validates. Simulation, never validation.
This skill has passed two wind tunnels. Before first release: 9 scenarios, 5 personas, 2 languages — 1 harmful false-fire caught (a bare 「这个想法怎么样?」 summoning a full persona matrix against a strategy decision) and removed from the trigger surface before any real session paid for it. After the v2 rewrite (Step 0 pitch-labeling, run sizing, the reality bridge, the pre-delivery gate): an 11-scenario regression — all 3 original trigger phrases re-fired, all boundary silences held, 0 trigger regressions. The tunnel's two remaining catches were documentation drift and one sizing/gate wording conflict — both fixed before this page went up.
Narrow audience, dictated (pass): 「先别写代码…给退休护士的记药 app,模拟几个用户跑跑」— fired on the verbatim phrases, and the diversity slots bent inside the niche exactly as written: age within the band, tech comfort, motivation — the slots never disappeared.
Correct silence, still holding after v2: a bare 「这个想法怎么样?」about a business pivot — the phrase this skill's own first tunnel evicted — stayed silent through the rewrite. Regression is the point: a rewrite that re-awakens an old false-fire has failed even if everything new works.
A boundary, ruled: 「给我的想法做个调研」— research-my-idea with no simulation language — routes to product-5w, not here. The two descriptions arbitrate it without a coin flip: 调研 belongs to the definition seat; this skill answers to 模拟 / 测测 / 推演.
Probe method: scenario-probe
Built by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.