Skip to content
AI Product Thinking
EN DE

Templates

Five templates from the curriculum. Copy them, adapt them, use them in your projects.

1. AI PRD Template

Section titled “1. AI PRD Template”

Based on Chapter 8: Writing AI PRDs.

AI PRD: [Product Name / Feature]
Date: [Date]
Author: [Name]


---


1. PROBLEM STATEMENT & USER CONTEXT
- Problem: [What is the problem? Quantify the manual effort.]
- Target audience: [Who has this problem? How often?]
- Current workaround: [How do users solve this today?]


2. AI APPROACH & RATIONALE
- Why AI? [Why not rules/traditional code?]
- Approach: [ ] LLM API  [ ] RAG  [ ] Fine-Tuning  [ ] Agent Workflow
- Rationale: [Why this approach?]
- AI suitability (5 Check Questions):
  [ ] Does the use case tolerate occasional errors?
  [ ] Is there enough training data / context?
  [ ] Is the value over rules significant?
  [ ] Is cost per query economically viable?
  [ ] Is the risk of errors acceptable?


3. EVALUATION CRITERIA
- Golden Dataset: [Size, source, labeling process]
- Metrics + thresholds:
  | Metric | Minimum | Target | Measurement |
  |--------|---------|--------|-------------|
  | [e.g. Accuracy] | [e.g. 85%] | [e.g. 92%] | [e.g. Golden Dataset] |
  | [e.g. Hallucination Rate] | [e.g. <5%] | [e.g. <2%] | [e.g. LLM-as-Judge] |
  | [e.g. Latency P95] | [e.g. <3s] | [e.g. <1s] | [e.g. APM] |
  | [e.g. Cost/Query] | [e.g. <$0.05] | [e.g. <$0.02] | [e.g. Provider Dashboard] |


4. MODEL & INFRASTRUCTURE
- Model: [e.g. Claude Sonnet 4.6 / GPT-4o-mini]
- Rationale: [Cost/quality/latency tradeoff]
- Expected volume: [Queries/day]
- Cost projection: [$/month]


5. USER EXPERIENCE
- AI output presentation: [e.g. Inline suggestion, separate panel, chat]
- Confidence indicators: [ ] Yes  [ ] No — Rationale: [...]
- Fallback for low confidence: [e.g. manual review, disclaimer]
- Feedback mechanism: [e.g. Thumbs up/down, Regenerate, Edit]


6. RISK & MITIGATION
- Failure modes:
  | Failure Mode | Probability | Impact | Mitigation |
  |-------------|------------|--------|------------|
  | [e.g. Hallucination on domain terms] | [Medium] | [High] | [RAG with verified source] |
  | [e.g. Prompt injection] | [Low] | [High] | [Input validation + guardrails] |
- Bias consideration: [Which groups could be disadvantaged?]
- Privacy: [What data flows into the model? Tier 1/2/3?]


7. SUCCESS METRICS & ITERATION PLAN
- Launch criteria: [Eval metrics above threshold + human review passed]
- Post-launch monitoring: [Which metrics, what cadence?]
- Improvement cadence: [e.g. prompt updates weekly, model upgrade quarterly]
- Rollback trigger: [At what values do you roll back?]

2. RICE-A Scoring

Section titled “2. RICE-A Scoring”

Based on Chapter 2: Opportunity Identification. RICE extended with AI Complexity.

RICE-A SCORING: [Project / Feature List]
Date: [Date]


| Feature | Reach | Impact | Confidence | Effort | AI Complexity | RICE-A Score |
|---------|-------|--------|------------|--------|--------------|-------------|
| [Feature A] | [1-10] | [1-5] | [0.5-1.0] | [1-10] | [1-5] | [calculated] |
| [Feature B] | [1-10] | [1-5] | [0.5-1.0] | [1-10] | [1-5] | [calculated] |
| [Feature C] | [1-10] | [1-5] | [0.5-1.0] | [1-10] | [1-5] | [calculated] |


Formula: (Reach x Impact x Confidence) / (Effort + AI Complexity x 0.5)


SCALES:


Reach (1-10):
  1-3: <1,000 users affected
  4-6: 1,000-10,000 users
  7-10: >10,000 users


Impact (1-5):
  1: Minimal — nice-to-have
  2: Low — saves some time
  3: Medium — noticeable improvement
  4: High — solves a real problem
  5: Massive — game changer


Confidence (0.5-1.0):
  0.5: Gut feeling, no data
  0.7: Some user signals
  0.8: Solid evidence
  1.0: Validated by prototype/pilot


Effort (1-10):
  1-3: <2 weeks, 1-2 people
  4-6: 2-6 weeks, small team
  7-10: >6 weeks, cross-functional


AI Complexity (1-5):
  1: Low — clear use case, good data, standard model
  2: Moderate — data available but needs preparation
  3: Medium — custom eval needed, data quality unclear
  4: High — model risk, extensive evaluation required
  5: Very high — unsolved problem, significant research needed

3. Red Team Plan

Section titled “3. Red Team Plan”

Based on Chapter 5: Red Teaming.

RED TEAM PLAN: [Product Name / Feature]
Date: [Date]
Test window: [Start — End]
Owner: [Name/Team]


SCOPE
- System under test: [e.g. Support chatbot, content generator]
- Model: [e.g. Claude Sonnet 4.6]
- Access path: [e.g. Web UI, API, Mobile App]


PRIORITY MATRIX
| Category | Priority | Test Cases | Status |
|----------|----------|------------|--------|
| Prompt Injection (direct) | Critical | [e.g. "Ignore all previous instructions"] | [ ] Open |
| Prompt Injection (indirect) | Critical | [e.g. Malicious content in uploads/URLs] | [ ] Open |
| Data Extraction | High | [e.g. Extract system prompt, read PII] | [ ] Open |
| Jailbreaking | High | [e.g. Roleplay attacks, multi-turn manipulation] | [ ] Open |
| Bias / Fairness | High | [e.g. Compare results by gender, ethnicity, age] | [ ] Open |
| Hallucination | Medium | [e.g. Ask for facts not in context] | [ ] Open |
| Edge Cases | Medium | [e.g. Empty input, very long input, different language] | [ ] Open |
| Misuse | Medium | [e.g. Generate harmful content] | [ ] Open |


TEST PROTOCOL PER FINDING
- Finding ID: [RT-001]
- Category: [e.g. Prompt Injection]
- Severity: [Critical / High / Medium / Low]
- Input: [Exact prompt]
- Expected behavior: [What should happen]
- Actual behavior: [What happened]
- Reproducible: [ ] Yes  [ ] No  [ ] Partially
- Recommendation: [Fix / Accepted risk / Monitoring]


RESULTS SUMMARY
| Severity | Found | Fixed | Accepted | Open |
|----------|-------|-------|----------|------|
| Critical | [n] | [n] | [n] | [n] |
| High | [n] | [n] | [n] | [n] |
| Medium | [n] | [n] | [n] | [n] |
| Low | [n] | [n] | [n] | [n] |


SHIP DECISION: [ ] Ship  [ ] Fix First  [ ] No-Ship
Rationale: [...]

4. Ship/No-Ship Checklist

Section titled “4. Ship/No-Ship Checklist”

Based on Chapter 5: Ship/No-Ship Decisions.

SHIP/NO-SHIP CHECKLIST: [Feature]
Date: [Date]
Decision maker: [Name]


EVALUATION
[ ] Golden Dataset defined and current
[ ] All core metrics above minimum threshold
    Accuracy: [current] vs. [minimum]
    Hallucination Rate: [current] vs. [maximum]
    Latency P95: [current] vs. [maximum]
[ ] Performance checked on relevant subgroups (no aggregation problem)


RED TEAMING
[ ] Red team completed (date: [...])
[ ] No open findings with severity "Critical"
[ ] Open "High" findings documented with mitigation


BIAS & FAIRNESS
[ ] Metrics disaggregated by relevant groups
[ ] No significant performance differences between groups
[ ] Fairness decision documented (which metric prioritized?)


ROLLBACK
[ ] Feature flag / kill switch in place
[ ] Rollback possible in <1h
[ ] Rollback triggers defined (at what values?)


MONITORING
[ ] Real-time monitoring for core metrics set up
[ ] Alerting configured (thresholds + recipients)
[ ] Feedback loop for user reports in place


COMPLIANCE
[ ] Privacy review completed
[ ] EU AI Act risk category determined
[ ] Guardrails implemented and tested
[ ] Documentation requirements fulfilled


DECISION
[ ] SHIP — all checks passed
[ ] SHIP WITH CONSTRAINTS — [which and why]
[ ] FIX FIRST — [what needs to happen]
[ ] NO-SHIP — [rationale]


Signature: __________ Date: __________

5. AI KPI Dashboard

Section titled “5. AI KPI Dashboard”

Based on Chapter 9: KPIs for AI Products.

AI KPI DASHBOARD: [Product Name]
Date: [Date]
Reporting cadence: [weekly / monthly]


--- QUALITY METRICS (Technical Quality) ---


| Metric | Current | Target | Trend | Alert at |
|--------|---------|--------|-------|----------|
| Hallucination Rate | [%] | [<5%] | [up/down/stable] | [>8%] |
| Groundedness Score | [%] | [>90%] | [up/down/stable] | [<85%] |
| Task Completion Rate | [%] | [>80%] | [up/down/stable] | [<70%] |
| Latency P95 | [ms] | [<2000ms] | [up/down/stable] | [>3000ms] |
| Eval Score (Golden Dataset) | [%] | [>85%] | [up/down/stable] | [<80%] |


--- BUSINESS METRICS (Business Impact) ---


| Metric | Current | Target | Trend | Alert at |
|--------|---------|--------|-------|----------|
| AI Feature Adoption Rate | [%] | [>30%] | [up/down/stable] | [<15%] |
| Regeneration Rate | [%] | [<20%] | [up/down/stable] | [>35%] |
| Revenue Impact | [$] | [...] | [up/down/stable] | [...] |
| Cost per Query | [$] | [<$0.05] | [up/down/stable] | [>$0.10] |
| User Retention (AI vs. Non-AI) | [%] | [...] | [up/down/stable] | [...] |


--- OPERATIONAL METRICS (Operations) ---


| Metric | Current | Target | Trend | Alert at |
|--------|---------|--------|-------|----------|
| Error Rate | [%] | [<1%] | [up/down/stable] | [>3%] |
| API Availability | [%] | [>99.5%] | [up/down/stable] | [<99%] |
| Monthly AI Cost | [$] | [...] | [up/down/stable] | [...] |
| Cost Trend (MoM) | [%] | [stable] | [up/down/stable] | [>+20%] |
| Guardrail Trigger Rate | [%] | [<5%] | [up/down/stable] | [>10%] |


--- LEADING INDICATORS ---


Regeneration Rate is the most important leading indicator:
- Rising → Users are unhappy with AI output → Quality problem
- Falling → AI output is being accepted → Quality is improving
- Correlates with churn? → If yes: Quality is directly revenue-relevant


--- ACTIONS ---


| Metric in alarm | Root cause | Action | Owner | Deadline |
|----------------|-----------|--------|-------|----------|
| [...] | [...] | [...] | [...] | [...] |

Part of AI Learning — free courses from prompt to production. Jan on LinkedIn