菜单
Use when designing AI products for long-term retention — stickiness patterns, daily engagement hooks, workflow integration depth, habit loops specific to AI, and measuring whether users actually keep using your AI feature.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | ai-product-retention |
| description | Use when designing AI products for long-term retention — stickiness patterns, daily engagement hooks, workflow integration depth, habit loops specific to AI, and measuring whether users actually keep using your AI feature. |
Only 5% of AI tools achieve meaningful retention (>50% month-12). Compare:
| Product | Month-12 Retention | Why |
|---|---|---|
| GitHub Copilot | 80% | Daily coding habit, workflow integration |
| ChatGPT Plus | 71% | Daily use, replaces multiple tools |
| Google Gemini | 25% | Redundant with free alternatives |
| Most new AI tools | <30% | One-time use, low frequency problem |
The retention formula: Frequency × Depth × Irreplaceability
Four dimensions that predict retention:
1. FREQUENCY: How often does the problem occur?
Daily → excellent retention ceiling
Weekly → good
Monthly → hard to retain
Quarterly → don't build a standalone product
2. DEPTH: How embedded in user's workflow?
Surface (one-off queries) → easily replaced
Embedded (part of daily process) → medium sticky
Critical path (breaks workflow if unavailable) → very sticky
3. PERSONALIZATION: Does it know the user's context?
Generic (same for all) → commodity
Adapted (learns preferences) → medium sticky
Deeply personal (knows my specific situation) → very sticky
4. SWITCHING COST: What does user lose by leaving?
Nothing (stateless) → zero retention lock
Some history (conversation, settings) → low lock
Irreplaceable data (years of context, fine-tuning) → high lock
Scoring: Rate 1-3 on each dimension. Total <6 = retention risk. Target 8-12.
B.J. Fogg's model applied to AI tools:
Trigger: What brings user back?
→ Daily notification ("Your AI morning brief is ready")
→ Workflow trigger ("You opened Figma — AI design assistant activated")
→ External event ("New data in your CRM — AI analysis ready")
→ Internal cue ("It's Monday morning → I always check AI dashboard")
Action: What user does (must be frictionless)
→ One click → AI generates something useful
→ Open existing tool → AI is already there (embedded)
→ Not: navigate to separate app + log in + configure
Reward: The dopamine moment
→ "AI just saved me 2 hours" (visible time saved)
→ "That insight surprised me" (serendipitous discovery)
→ "I couldn't have found that myself" (genuine augmentation)
Investment: What makes next session more valuable
→ AI learns preferences over time
→ User adds context (notes, corrections) that improves AI
→ Integration creates network effects (more data = better AI)
Users who don't activate in session 1 almost never return:
Activation target: User experiences core value within 10 minutes of signup
10-minute activation checklist:
✅ No complex onboarding wizard (max 3 steps)
✅ Pre-populated with sample data so AI can show its value
✅ First AI output is impressively relevant (not generic)
✅ Clear "aha moment" — AI does something user couldn't do alone
✅ One clear next action after aha moment
Activation measurement:
Activated user = completed [key action] within first session
Activation rate target: >40% (great), 20-40% (needs work), <20% (critical)
Find your activation event by comparing activated vs not-activated users:
"What did users who paid do in their first session that churned users didn't?"
Beyond standard cohort retention:
# AI-specific retention metrics
metrics = {
# Standard
"day_1_retention": users_return_day1 / new_users,
"day_7_retention": users_return_day7 / new_users,
"month_1_retention": users_return_day30 / new_users,
# AI-specific
"ai_acceptance_rate": outputs_accepted / outputs_generated, # did user keep AI output?
"ai_edit_rate": outputs_edited / outputs_used, # how much did user change AI output?
"ai_dependency_score": tasks_with_ai / total_tasks, # % of work involving AI
"regeneration_rate": regenerations / total_generations, # proxy for quality
"daily_active_ai_users": users_using_ai_today / total_users, # not just logged in
}
# Alert thresholds
alerts = {
"acceptance_rate_drop": acceptance_rate < 0.6, # quality problem
"edit_rate_spike": edit_rate > 0.8, # AI output not useful enough
"dependency_decline": dependency_score declining week_over_week, # abandonment signal
}
ai-product-design for streaming/UX patterns that support habit loopsllm-observability to track AI acceptance rate and quality signalsai-product-validation (validate problem frequency before building)@solo-ai-builder reviews retention design before launchUse when conducting quality audits — reviewing process compliance, identifying gaps between defined process and actual practice, conducting structured inspections (code review audits, test quality reviews), and producing audit reports with remediation plans.
Use when managing quality risk — identifying quality risks in a product or release, applying risk-based testing prioritization, creating risk mitigation plans, and communicating quality risk to stakeholders for go/no-go decisions.
Use when designing or implementing test automation — choosing the right automation framework (Playwright, pytest, JUnit), Page Object Model, selector strategies, test isolation, managing flaky tests, and CI integration.
Use when managing defects — writing effective bug reports, applying severity/priority matrix, tracking defect lifecycle, conducting root cause analysis, and measuring defect metrics for process improvement.
Use when measuring and reporting QA quality — defect escape rate, test coverage analysis, flaky test rate, mean time to detect, shift-left metrics, and building quality dashboards for stakeholders.
Use when managing test data — designing test data strategies, using factories and builders, creating fixtures, generating synthetic data, masking PII for testing, and managing test database state.