AI Interview Kit

Professional-grade user interviews at scale —powered by prompt engineering, not bigger budgets.

34 → 94 % follow-up hit rate  ·  50+ iteration rounds  ·  3 500+ production calls

     

中文版 →

50 s demo  ·  research question → auto-classify → pick methodology → generate full prompt

---

⚡ Quick Start

git clone https://github.com/CyannSHI/ai-interview-kit.git
cd ai-interview-kit
# Open with any supported AI tool and say "generate prompt"

Skill	Trigger	What It Does
generate-prompt	"generate prompt" / "new project"	Guided info collection → methodology pick → auto-assembled prompt
generate-input	"prepare input variables"	Natural language → structured input variables
evaluate	"evaluate calls"	Batch review call transcripts; detect bad cases & output Excel report

---

Who It's For

UX Researchers

Your ceiling: 5–8 deep interviews a week. Outsourcing just trades one problem for another. When the budget runs out, sample size gets cut — quality never goes up.

PMs & Founders

You know you should talk to users — but what do you ask, and how deep do you go? ChatGPT gives surface-level answers that dead-end after two turns. You finally invest the time, only to realize you missed every key question.

The idea: engineer research methodologies into AI prompts. Non-experts get professional interview quality. Experts get 10× scale.

---

🎯 Core Innovation — AI Control Precision

Most AI interview prompts are either too rigid or too loose. Different goals need different AI latitude — and seasoned researchers dial this by instinct. We parameterized that instinct.

AI freedom:   Low ◄━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━► High

              Precise Control       Balanced Mode        Exploratory Mode
              confirmatory          most projects        discovery

	Precise Control	Balanced	Exploratory
Use case	Targeted validation	Clear direction, some flex	Open-ended discovery
Key-info markers	Per question	Per question	None — AI decides
Probing limit	4 rounds / Q	6 rounds / Q	AI discretion
Min coverage	80 % required Qs	75 %	60 %
Example	NPS callback	JTBD migration	New-product exploration

We distilled interviewers' tacit knowledge — when to drill down, when to skip, when to follow a thread — into three tunable parameters: info-point density, probing-round cap, and minimum coverage rate.

---

🧩 How It Works — Prompt Engineering Architecture

Universal framework (how to ask)  +  Pluggable methodology (what to ask)  +  Project variables (whom to ask)→ auto-assembled into a production-ready prompt

The SLOT Mechanism

Add a new methodology by writing just 3 SLOTs — no framework changes needed. Framework upgrades automatically benefit every methodology.

---

📚 Methodology Library

Methodology	Best For	Core Dimensions
JTBD Migration	User decisions · churn · competitor switching	Push · Pull · Anxiety · Habit · Destination
Journey Mapping	Experience flows · friction points · action chains	Stage · Touchpoint · Behavior · Emotion · Breakpoint
NPS / Satisfaction	Satisfaction callback · service improvement	Positive driver · Negative driver · Expectation gap
Laddering	Deep motivation · value discovery	Attribute · Functional benefit · Emotional benefit · Core value
User Lifecycle	Conversion · retention · churn	Acquisition · Conversion · Usage · Retention · Churn
Brand Diagnostics	Brand perception · competitive positioning	Awareness · Association · Preference · Comparison · Loyalty

Custom methodology? Copy methodologies/_template.md, fill in 3 SLOTs, save — done.

---

📊 Validation — Iterated Through 50+ Experiments

v0.1  ███████░░░░░░░░░░░░░  34%   Flat question list — no probing
v0.2  ████████████░░░░░░░░  61%   + Key-info markers
v0.3  ██████████████████░░  89%   + Probing cap & 3-strike rule
v0.4  ███████████████████░  94%   + Methodology SLOT mechanism

Problems Discovered

Low-quality small-talk — users roll their eyes

User: "Train-ticket scalping doesn't even matter… what are you asking?"

Broken pacing — too many questions at once

AI asked 3 questions in one turn; the user only answered the last one.

After Optimization

Precise summarization — real-time synthesis, zero leakage

AI consolidates scattered key info — booking channels, decision factors, membership perception, desired benefits — user responds "yes, exactly!"

Layer-by-layer drill-down — from vague complaints to concrete events

From one vague complaint, AI drills down in 3 steps: city → hotel → specific incident.

Natural expansion — context-aware topic transitions

AI naturally extends from hotel booking experience to cross-channel price comparisons — smooth and unforced.

Targeted probing — stays on point, layer after layer

AI probes the reason behind "two channels," drilling into price differences step by step.

 Stress-test plan: 6 extreme scenarios

#	Scenario	What It Tests
1	Memory activation	Can AI gently help users recall when they say "I don't remember"?
2	Deep drill-down without leading	Can AI ask purely open-ended questions — no options, no nudging?
3	Factual contradiction detection	Can AI catch and probe when users contradict themselves?
4	High-pressure emotion handling	Can AI de-escalate anger and steer back on track?
5	Signal extraction from noise	Can AI identify key info when users ramble?
6	*