Agent Skills
› mohitagw15856/pm-claude-skills
› expert-interview-prep
expert-interview-prep
GitHub针对专家访谈的深度准备工具。通过梳理专家公开信息生成作业简报，设计从校准到前沿的问题弧，提供深挖追问技巧及会后记录方案。旨在避免浅层提问，确保在有限时间内获取独家洞察并支持决策。
Trigger Scenarios
请求为专家通话做准备
询问应向顾问/分析师提问的内容
即将与特定领域专家进行一小时对话
现有专家访谈流于表面
Install
npx skills add mohitagw15856/pm-claude-skills --skill expert-interview-prep -g -y
SKILL.md
Frontmatter
{
"name": "expert-interview-prep",
"homepage": "https:\/\/mohitagw15856.github.io\/pm-claude-skills\/skill\/expert-interview-prep.html",
"metadata": {
"openclaw": {
"emoji": "🧠"
}
},
"description": "Get the most from an hour with an expert — the do-your-homework floor (never ask what's googleable), the question arc from calibration to the frontier, the follow-up discipline that goes deep instead of wide, and the capture that survives the call. Use when asked prep me for the expert call, what should I ask this advisor\/analyst\/practitioner, we get an hour with X, or our expert calls are pleasant but shallow. Produces the homework brief, the question arc, the follow-up toolkit, and the capture plan."
}
Expert Interview Prep Skill
An hour with a real expert is expensive and perishable — and most of it gets spent badly: fifteen minutes of questions the expert's own writing already answers, a list marched through on schedule while the expert's best tangent dies unfollowed, and notes that resolve to "great conversation." The prep flips the economics: homework raises the starting altitude (the call begins where public knowledge ends), the question arc is designed (calibration → the core unknowns → the frontier questions only this person can answer), and the discipline in the room is follow-up over coverage — the fifth "why" beats the ninth topic.
What This Skill Produces
- The homework brief — what the expert has already said publicly (writing, talks, the searchable record), so no question re-asks it
- The question arc — calibration openers → the core unknowns (ranked, since time will run out) → the only-this-person frontier questions
- The follow-up toolkit — the probes that go deep: "what's an example?", "who disagrees with that and why?", "what would change your mind?"
- The capture plan — recording/notes division, the same-day synthesis, and the repo deposit (research-repo-setup)
Required Inputs
Ask for these if not provided:
- The expert and the occasion — who, their actual expertise edges (experts are spiky — the prep aims at their peaks), and the relationship (a paid analyst call, a courtesy favor, a potential advisor — the register differs)
- The decision the call feeds — what the team will do differently after; questions trace to it or get cut (desk-research-sprint decomposition applies)
- The current belief state — what the team thinks it knows, including the shaky parts; the best expert questions test beliefs, and hidden beliefs can't be tested
- The time box — 30 vs. 60 minutes changes the arc's ambition; both change what gets pre-sent
Framework: The Prep Rules
- Homework to the public frontier: read what they've written, watch the talk, know their stated positions — then open questions with "you've written that X — …" which simultaneously (a) skips the retread, (b) signals seriousness (experts calibrate their effort to the asker's), and (c) earns the right to push on the position. Asking a googleable question tells the expert which hour this will be.
- The arc: calibrate, core, frontier: two calibration questions (their current view, what's changed lately — cheap, warms the channel, catches stale homework) → the ranked core unknowns (the decision-critical questions, best first — time will run out) → the frontier ("what's nobody asking about this?", "where is the conventional wisdom wrong?") — the questions with no other source.
- Follow-up beats coverage: the standing rule — when an answer surprises, stay ("say more," "what's a concrete case?", "who disagrees?") — a list fully covered at survey depth loses to three questions at mechanism depth. The list is a map, not a contract; ranked order means the sacrificed questions are the right ones.
- Test beliefs explicitly: at least one question stakes the team's current belief for demolition — "we're operating on the assumption that X; where is that wrong?" Experts correct stated beliefs far more usefully than they answer open questions; the correction is often the call's whole ROI.
- Capture with epistemics, same day: note claims vs. opinions vs. hearsay as the expert frames them ("everyone's seeing" ≠ "my data shows"), synthesize same-day while tone and emphasis survive, deposit to the repo with the expert-single-source confidence grade — one expert is one (excellent) source, and the source-triangulation rules apply to their numbers too.
Output Format
Expert Call Prep: [expert] — [T] min, feeds [the decision]
Homework Brief
[Their public positions relevant to us · the quotes worth referencing · the stale-check question]
The Arc
Calibration: [2 openers] · Core (ranked): [the unknowns, best-first, each traced to the decision] · Frontier: [the only-this-person questions]
The Toolkit
[The follow-up probes · the belief-test question, staked · the time-checkpoint rule (at T/2, jump to the top unasked core question)]
Capture
[Recording/notes split · claims-vs-opinions marking · same-day synthesis · repo deposit with confidence grade]
Quality Checks
- No question on the list is answerable by their public record
- Core questions are ranked and traced to the decision
- The belief-test question stakes a real current assumption
- The follow-up rule outranks the coverage instinct, explicitly
- Capture distinguishes claims, opinions, and hearsay
Anti-Patterns
- Do not spend expert-minutes on googleable questions — the hour starts at the public frontier or it's a podcast
- Do not march the list — the surprising answer is the call's best moment; follow it
- Do not ask only open questions — staked beliefs get corrected; blank slates get lectures
- Do not treat expert assertion as established fact — one source, graded accordingly, numbers triangulated
- Do not synthesize next week — expert calls decay overnight into "it was really useful"
Version History
- 471c606 Current 2026-07-22 13:24
