exp-driven-dev（经验驱动开发）

Experimentation-Driven Development

When This Skill Activates

Claude uses this skill when:

• Building new features that affect core metrics
• Implementing A/B testing infrastructure
• Making data-driven decisions
• Setting up feature flags for gradual rollouts
• Choosing which metrics to track

Core Frameworks

1. Experiment Design (Source: Ronny Kohavi, Microsoft/Netflix)

The HITS Framework:

H - Hypothesis:

"We believe that [change] will cause [metric] to [increase/decrease] because [reason]"

I - Implementation:

• Feature flag setup
• Treatment vs control
• Sample size calculation

T - Test:

• Run for statistical significance
• Monitor guardrail metrics
• Watch for unexpected effects

S - Ship or Stop:

• Ship if positive
• Stop if negative
• Iterate if inconclusive

Example:

Hypothesis:
"We believe that adding social proof ('X people bought this')
will increase conversion rate by 10%
because it reduces purchase anxiety."

Implementation:
- Control: No social proof
- Treatment: Show "X people bought"
- Sample size: 10,000 users per variant
- Duration: 2 weeks

Test:
- Primary metric: Conversion rate
- Guardrails: Cart abandonment, return rate

Ship or Stop:
- If conversion +5% or more → Ship
- If conversion -2% or less → Stop
- If inconclusive → Iterate and retest

2. Metric Selection

Primary Metric:

• ONE metric you're trying to move
• Directly tied to business value
• Clear success threshold

Guardrail Metrics:

• Metrics that shouldn't degrade
• Prevent gaming the system
• Ensure quality maintained

Example:

Feature: Streamlined checkout

Primary Metric:
✅ Purchase completion rate (+10%)

Guardrail Metrics:
⚠️ Cart abandonment (don't increase)
⚠️ Return rate (don't increase)
⚠️ Support tickets (don't increase)
⚠️ Load time (stay <2s)

3. Statistical Significance

The Math:

Minimum sample size = (Effect size, Confidence, Power)

Typical settings:
- Confidence: 95% (p < 0.05)
- Power: 80% (detect 80% of real effects)
- Effect size: Minimum detectable change

Example:
- Baseline conversion: 10%
- Minimum detectable effect: +1% (to 11%)
- Required: ~15,000 users per variant

Common Mistakes:

• ❌ Stopping test early (peeking bias)
• ❌ Running too short (seasonal effects)
• ❌ Too many variants (dilutes sample)
• ❌ Changing test mid-flight

4. Feature Flag Architecture

Implementation:

// Feature flag pattern
function checkoutFlow(user) {
  if (isFeatureEnabled(user, 'new-checkout')) {
    return newCheckoutExperience();
  } else {
    return oldCheckoutExperience();
  }
}

// Gradual rollout
function isFeatureEnabled(user, feature) {
  const rolloutPercent = getFeatureRollout(feature);
  const userBucket = hashUserId(user.id) % 100;
  return userBucket < rolloutPercent;
}

// Experiment assignment
function assignExperiment(user, experiment) {
  const variant = consistentHash(user.id, experiment);
  track('experiment_assigned', {
    userId: user.id,
    experiment: experiment,
    variant: variant
  });
  return variant;
}

Decision Tree: Should We Experiment?

NEW FEATURE
│
├─ Affects core metrics? ──────YES──→ EXPERIMENT REQUIRED
│  NO ↓
│
├─ Risky change? ──────────────YES──→ EXPERIMENT RECOMMENDED
│  NO ↓
│
├─ Uncertain impact? ──────────YES──→ EXPERIMENT USEFUL
│  NO ↓
│
├─ Easy to A/B test? ─────────YES──→ WHY NOT EXPERIMENT?
│  NO ↓
│
└─ SHIP WITHOUT TEST ←────────────────┘
   (But still feature flag for rollback)

Action Templates

Template 1: Experiment Spec

# Experiment: [Name]

## Hypothesis
**We believe:** [change]
**Will cause:** [metric] to [increase/decrease]
**Because:** [reasoning]

## Variants

### Control (50%)
[Current experience]

### Treatment (50%)
[New experience]

## Metrics

### Primary Metric
- **What:** [metric name]
- **Current:** [baseline]
- **Target:** [goal]
- **Success:** [threshold]

### Guardrail Metrics
- **Metric 1:** [name] - Don't decrease
- **Metric 2:** [name] - Don't increase
- **Metric 3:** [name] - Maintain

## Sample Size
- **Users needed:** [X per variant]
- **Duration:** [Y days]
- **Confidence:** 95%
- **Power:** 80%

## Implementation

if (experiment('feature-name') === 'treatment') { // New experience } else { // Old experience }

## Success Criteria

- Primary metric improved by [X]%
- No guardrail degradation
- Statistical significance reached
- No unexpected negative effects

## Decision

- **If positive:** Ship to 100%
- **If negative:** Rollback, iterate
- **If inconclusive:** Extend or redesign

Template 2: Feature Flag Implementation

// features.ts
export const FEATURES = {
  'new-checkout': {
    rollout: 10,  // 10% of users
    enabled: true,
    description: 'New streamlined checkout flow'
  },
  'ai-recommendations': {
    rollout: 0,  // Not live yet
    enabled: false,
    description: 'AI-powered product recommendations'
  }
};

// feature-flags.ts
export function isEnabled(userId: string, feature: string): boolean {
  const config = FEATURES[feature];
  if (!config || !config.enabled) return false;

  const bucket = consistentHash(userId) % 100;
  return bucket < config.rollout;
}

// usage in code
if (isEnabled(user.id, 'new-checkout')) {
  return <NewCheckout />;
} else {
  return <OldCheckout />;
}

Template 3: Experiment Dashboard

# Experiment Dashboard

## Active Experiments

### Experiment 1: [Name]
- **Status:** Running
- **Started:** [date]
- **Progress:** [X]% sample size reached
- **Primary metric:** [current result]
- **Guardrails:** ✅ All healthy

### Experiment 2: [Name]
- **Status:** Complete
- **Result:** Treatment won (+15% conversion)
- **Decision:** Ship to 100%
- **Shipped:** [date]

## Key Metrics

### Experiment Velocity
- **Experiments launched:** [X per month]
- **Win rate:** [Y]%
- **Average duration:** [Z] days

### Impact
- **Revenue impact:** +$[X]
- **Conversion improvement:** +[Y]%
- **User satisfaction:** +[Z] NPS

## Learnings
- [Key insight 1]
- [Key insight 2]
- [Key insight 3]

Quick Reference

🧪 Experiment Checklist

Before Starting:

• Hypothesis written (believe → cause → because)
• Primary metric defined
• Guardrails identified
• Sample size calculated
• Feature flag implemented
• Tracking instrumented

During Experiment:

• Don't peek early (wait for significance)
• Monitor guardrails daily
• Watch for unexpected effects
• Log any external factors (holidays, outages)

After Experiment:

• Statistical significance reached
• Guardrails not degraded
• Decision made (ship/stop/iterate)
• Learning documented

Real-World Examples

Example 1: Netflix Experimentation

Volume: 250+ experiments running at once Approach: Everything is an experiment Culture: "Strong opinions, weakly held - let data decide"

Example Test:

• Hypothesis: Bigger thumbnails increase engagement
• Result: No improvement, actually hurt browse time
• Decision: Rollback
• Learning: Saved $$ by not shipping

Example 2: Airbnb's Experiments

Test: New search ranking algorithm Primary: Bookings per search Guardrails:

• Search quality (ratings of bookings)
• Host earnings (don't concentrate bookings)
• Guest satisfaction

Result: +3% bookings, all guardrails healthy → Ship

Example 3: Stripe's Feature Flags

Approach: Every feature behind flag Benefits:

• Instant rollback (flip flag)
• Gradual rollout (1% → 5% → 25% → 100%)
• Test in production safely

Example:

if (experiments.isEnabled('instant-payouts')) {
  return <InstantPayouts />;
}

Common Pitfalls

❌ Mistake 1: Peeking Too Early

Problem: Stopping test before statistical significance Fix: Calculate sample size upfront, wait for it

❌ Mistake 2: No Guardrails

Problem: Gaming the metric (increase clicks but hurt quality) Fix: Always define guardrails

❌ Mistake 3: Too Many Variants

Problem: Not enough users per variant Fix: Limit to 2-3 variants max

❌ Mistake 4: Ignoring External Factors

Problem: Holiday spike looks like treatment effect Fix: Note external events, extend duration

Related Skills

• metrics-frameworks - For choosing right metrics
• growth-embedded - For growth experiments
• ship-decisions - For when to ship vs test more
• strategic-build - For deciding what to test

Key Quotes

Ronny Kohavi:

"The best way to predict the future is to run an experiment."

Netflix Culture:

"Strong opinions, weakly held. Let data be the tie-breaker."

Airbnb:

"We trust our intuition to generate hypotheses, and we trust data to make decisions."

Further Learning

• references/experiment-design-guide.md - Complete methodology
• references/statistical-significance.md - Sample size calculations
• references/feature-flags-implementation.md - Code examples
• references/guardrail-metrics.md - Choosing guardrails