hardiktiwari/PM-operating-OS / skills/prd-writer
prd-writer
VerifiedUse this skill when user asks to create, write, or draft a PRD (Product Requirements Document). Generates structured PRDs with governance framework, problem statements, business goals, use cases, and functional requirements.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 6.2 KB |
| references/agentic-prd-template.md | markdown | 35.3 KB |
| references/prd-template.md | markdown | 6.3 KB |
references/agentic-prd-template.md
View on GitHub# [Product Name] - Agentic Product Requirements Document
---
## 1. INTRODUCTION & CONTEXT SETTING
### 1.1 Product Vision
[One sentence: What does this agent do and why does it matter?]
### 1.2 Agent Classification
| Dimension | Selection |
|-----------|-----------|
| **Autonomy Level** | [ ] Copilot (human-initiated) / [ ] Autopilot (agent-initiated) / [ ] Hybrid |
| **Interaction Mode** | [ ] Conversational / [ ] Background / [ ] Both |
| **Learning Type** | [ ] Static (fixed knowledge) / [ ] Adaptive (learns from feedback) |
| **Scope** | [ ] Single-task / [ ] Multi-task / [ ] General purpose |
### 1.3 Success North Star
> "The agent is successful when [specific measurable outcome]"
### 1.4 Link to BRD / Strategy Docs
[If BRD exists, link here]
---
## 2. TARGET SEGMENT & PERSONA
### 2.1 Target Persona
| Persona | Mindset | Agent Relationship |
|---------|---------|-------------------|
| [Name] | [What they're trying to accomplish] | [How they want to interact with AI—trust level, control preference] |
**AI Trust Profile:**
- [ ] AI Skeptic — Needs to see reasoning, wants full control
- [ ] AI Curious — Open to suggestions, wants to understand
- [ ] AI Adopter — Trusts recommendations, wants automation
- [ ] AI Native — Expects autonomous action, minimal friction
### 2.2 Customer Problem (AI-Specific Framing)
**Job To Be Done:**
> I am a __________ trying to _________…
> But I can't because [cognitive load / expertise gap / time constraint / data overload]
> So I need an agent that [proactively identifies / reactively advises / automates] ________
### 2.3 Why AI? (The AI Value Proposition)
| Without Agent | With Agent |
|---------------|------------|
| [Current state: manual, slow, requires expertise] | [Target state: automated, instant, democratized] |
**AI Leverage Point:** [What makes this problem suited for AI vs. traditional software?]
- [ ] Pattern recognition at scale
- [ ] Natural language understanding
- [ ] Personalization across many dimensions
- [ ] Continuous monitoring / proactive detection
- [ ] Expert knowledge democratization
- [ ] Unstructured data processing
- [ ] Multi-step reasoning
### 2.4 Hypotheses
| Customer Friction | Root Cause | Hypothesis |
|-------------------|------------|------------|
| [Friction #1] | [Why this happens] | If we [intervention], then [X]% of users will [measurable behavior] |
| [Friction #2] | [Why this happens] | If we [intervention], then [X]% of users will [measurable behavior] |
---
## 3. AGENT ARCHITECTURE
### 3.1 High-Level Architecture
```
┌─────────────────────────────────────────────────────────────┐
│ CONTEXT LAYER │
│ [User Profile] [Business Data] [Connected Systems] [Docs] │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Component A │◄──►│ Component B │◄──►│ Component C │◄──►│ Component D │
│ [Purpose] │ │ [Purpose] │ │ [Purpose] │ │ [Purpose] │
└─────────────┘ └─────────────┘ └─────────────┘ └─────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ OUTPUT LAYER │
│ [Responses] [Notifications] [Actions] [Artifacts] │
└─────────────────────────────────────────────────────────────┘
```
### 3.2 Component Definitions
For each component, define:
#### Component: [Name]
| Attribute | Definition |
|-----------|------------|
| **Job Statement** | "I exist to [specific purpose]" |
| **Trigger** | [When does this component activate?] |
| **Input** | [What data/signals does it receive?] |
| **Processing** | [What logic/reasoning does it perform?] |
| **Output** | [What does it produce?] |
| **Dependencies** | [What other components does it rely on?] |
**Input Schema:**
```yaml
component_input:
field_1: type
field_2: type
```
**Output Schema:**
```yaml
component_output:
result: type
confidence: float
reasoning: string
```
### 3.3 Coordination Patterns
**Pattern Type:** [ ] Sequential / [ ] Parallel / [ ] Orchestrated / [ ] Event-driven
**Flow Diagram:**
```mermaid
sequenceDiagram
participant User
participant ComponentA
participant ComponentB
participant ComponentC
User->>ComponentA: [Trigger]
ComponentA->>ComponentB: [Handoff]
ComponentB->>ComponentC: [Handoff]
ComponentC->>User: [Response]
```
---
## 4. CONTEXT LAYER SPECIFICATION
### 4.1 Context Sources
| Source | Data Elements | Refresh Rate | Priority | Owner |
|--------|--------------|--------------|----------|-------|
| User Profile | [Fields: id, preferences, history...] | On-change | P0 | [Team] |
| Business Data | [Fields: transactions, categories...] | Real-time | P0 | [Team] |
| Connected Systems | [Platform data, API feeds...] | Webhook/polling | P1 | [Team] |
| Documents/Evidence | [Uploads, receipts, forms...] | On-upload | P1 | [Team] |
| Derived/Inferred | [Agent-computed signals...] | Continuous | P1 | [Team] |
### 4.2 Context Schema
```yaml
context_layer:
user_profile:
user_id: string (required)
created_at: datetime
entity_type: enum[individual, sole_prop, llc, s_corp, c_corp]
preferences:
communication_style: enum[concise, detailed, step_by_step]
risk_tolerance: enum[conservative, balanced, aggressive]
notification_cadence: enum[real_time, daily, weekly]
autonomy_preference: enum[always_ask, suggest, auto_execute]
business_profile:
industry: string
business_type: string
start_date: date
maturity_band: enum[startup, growth, established]
# Domain-specific fields...
connected_systems:
- platform: string
connected_at: datetime
last_sync: datetime
data_available: list[string]
derived_signals:
# Agent-computed understanding
current_intent: string
confidence: float
relevant_history: list[interaction]
active_opportunities: list[opportunity]
```
### 4.3 Memory Architecture
| Memory Type | Scope | Retention | Storage | Purpose |
|-------------|-------|-----------|---------|---------|
| **Working Memory** | Current request | Request duration | In-memory | Immediate reasoning context |
| **Session Memory** | Conversation | Session duration | Cache | Conversation coherence |
| **User Memory** | Individual user | Persistent | Database | Personalization, preferences, history |
| **System Memory** | All users | Persistent | Vector DB | Patterns, knowledge base, RAG |
### 4.4 Context Window Management
| Priority | Context Type | Token Budget | Truncation Strategy |
|----------|-------------|--------------|---------------------|
| 1 (Always) | System prompt + guardrails | Fixed ~500 | Never truncate |
| 2 (Always) | Current user query | Variable | Never truncate |
| 3 (High) | Recent conversation | ~2000 | FIFO, keep last N turns |
| 4 (Medium) | User profile summary | ~500 | Summarize if needed |
| 5 (Low) | Historical context | ~1000 | RAG retrieval, top-k |
---
## 5. INTEGRATIONS & DATA DEPENDENCIES
### 5.1 Third-Party Platform Integrations
For each platform:
#### Integration: [Platform Name]
| Attribute | Details |
|-----------|---------|
| **Purpose** | [Why integrate with this platform?] |
| **Auth Method** | OAuth2 / API Key / Webhook |
| **Data Retrieved** | [List of data types] |
| **Refresh Mechanism** | Real-time webhook / Polling interval / On-demand |
| **Fallback** | [What happens if integration is unavailable?] |
**Context Enrichment:**
```yaml
platform_context:
# What context does this platform provide?
field_1: type
field_2: type
```
**Triggers Enabled:**
| Trigger Name | Detection Logic | Enabled Behaviors |
|--------------|-----------------|-------------------|
| [e.g., large_expense] | `transaction.amount > threshold` | [What agent can now do] |
| [e.g., revenue_spike] | `daily_revenue > 2x average` | [What agent can now do] |
**Example Integration: QuickBooks**
```yaml
quickbooks_integration:
purpose: "Financial context for business expense optimization"
auth: OAuth2
data_retrieved:
- chart_of_accounts
- transactions (last 12 months)
- vendors
- invoices
- profit_and_loss
- balance_sheet
context_provided:
ytd_revenue: float
ytd_expenses: float
expense_by_category: dict[category, amount]
top_vendors: list[vendor]
cash_position: float
triggers_enabled:
- name: large_expense
logic: "transaction.amount > $5000 OR transaction.amount > 3x category_average"
action: "Flag for review, suggest categorization"
- name: recurring_expense_change
logic: "recurring_vendor.amount changed by >20%"
action: "Alert user, suggest renegotiation"
- name: category_anomaly
logic: "category.monthly_spend > 2x historical_average"
action: "Proactive alert with breakdown"
- name: cash_flow_warning
logic: "projected_cash < 2x monthly_expenses"
action: "Proactive cash management suggestions"
refresh:
primary: webhook (real-time for transactions)
fallback: daily batch sync at 2am UTC
failure_handling:
- on_auth_expired: "Prompt user to reconnect"
- on_api_error: "Use cached data, mark stale"
- on_rate_limit: "Queue requests, exponential backoff"
```
**Example Integration: Shopify**
```yaml
shopify_integration:
purpose: "E-commerce context for revenue optimization and inventory insights"
auth: OAuth2
data_retrieved:
- orders (last 12 months)
- products
- customers
- inventory_levels
- fulfillments
- refunds
context_provided:
total_revenue: float
order_count: int
average_order_value: float
top_products: list[product]
customer_segments: dict[segment, count]
inventory_status: dict[product, status]
triggers_enabled:
- name: sales_milestone
logic: "daily_revenue > previous_best OR monthly_revenue crosses threshold"
action: "Celebrate, suggest tax implications"
- name: inventory_alert
logic: "product.inventory < reorder_point"
action: "Alert, suggest purchase timing for tax optimization"
- name: refund_spike
logic: "daily_refunds > 2x average"
action: "Alert, investigate pattern"
- name: seasonal_pattern
logic: "current_period matches historical_peak"
action: "Proactive preparation suggestions"
refresh:
primary: webhook (orders, inventory)
fallback: hourly polling
```
**Example Integration: Banking (Plaid)**
```yaml
plaid_integration:
purpose: "Real-time transaction monitoring and cash flow visibility"
auth: Plaid Link (OAuth)
data_retrieved:
- accounts
- transactions (real-time)
- balances
- income_verification (optional)
context_provided:
account_balances: dict[account, balance]
recent_transactions: list[transaction]
income_streams: list[income_source]
spending_patterns: dict[category, pattern]
triggers_enabled:
- name: large_deposit
logic: "deposit.amount > $10000"
action: "Tax implications alert, estimated payment reminder"
- name: business_expense_detected
logic: "transaction.merchant in business_categories"
action: "Suggest categorization, deduction opportunity"
- name: low_balance_warning
logic: "balance < threshold AND upcoming_payments > balance"
action: "Cash flow alert"
```
### 5.2 Integration Priority Matrix
| Platform | Data Value | Trigger Value | Implementation Effort | Priority |
|----------|-----------|---------------|----------------------|----------|
| QuickBooks | High | High | Medium | P0 |
| Plaid/Banking | High | High | Medium | P0 |
| Shopify | Medium | Medium | Low | P1 |
| Stripe | Medium | Medium | Low | P1 |
| Gusto (Payroll) | Medium | Low | Medium | P2 |
| Square | Medium | Medium | Low | P2 |
### 5.3 Integration Architecture
```
┌─────────────────────────────────────────────────────────────┐
│ INTEGRATION LAYER │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ QuickBooks │ │ Shopify │ │ Plaid │ ... │
│ │ Adapter │ │ Adapter │ │ Adapter │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ Unified Data Normalizer │ │
│ │ (Transforms platform-specific → canonical schema) │ │
│ └─────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ Trigger Engine │ │
│ │ (Evaluates rules against normalized data) │ │
│ └─────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────┐ │
│ │ Context Store │ │
│ │ (Persists enriched context for agent access) │ │
│ └─────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
```
---
## 6. AGENT BEHAVIORS
### 6.1 Reactive Capabilities (User-Initiated)
| User Intent | Example Queries | Agent Behavior | Output Type | Accuracy Target |
|-------------|-----------------|----------------|-------------|-----------------|
| [Intent category] | "[Example question]" | [What agent does] | [Response type] | [Target %] |
### 6.2 Proactive Capabilities (Agent-Initiated)
| Signal/Trigger | Detection Logic | Agent Action | Notification Method | User Control |
|----------------|-----------------|--------------|---------------------|--------------|
| [Event type] | [How detected] | [What agent does] | [Push/email/in-app] | [Can disable?] |
### 6.3 Tools & Actions
| Tool Name | Purpose | Input | Output | Risk Level | Approval Required |
|-----------|---------|-------|--------|------------|-------------------|
| `search_knowledge` | Query knowledge base | query: string | results: list | None | No |
| `calculate` | Perform calculations | expression: string | result: number | None | No |
| `send_notification` | Alert user | message, priority | status | Low | No |
| `create_reminder` | Schedule follow-up | date, message | confirmation | Low | No |
| `draft_document` | Generate document | template, data | document | Medium | Review before send |
| `execute_action` | Perform external action | action_type, params | result | High | Explicit approval |
### 6.4 Multi-Turn Behavior
| Scenario | Agent Behavior | Memory Used |
|----------|---------------|-------------|
| Follow-up question | Reference previous context | Session memory |
| Clarification needed | Ask targeted question | Working memory |
| Complex task | Break into steps, confirm | Session + user memory |
| Return after absence | Summarize previous state | User memory |
---
## 7. GUARDRAILS & SAFETY
### 7.1 Agent Boundaries
| The Agent MUST | The Agent MUST NOT |
|----------------|-------------------|
| Always identify itself as AI | Claim to be human |
| Cite sources for factual claims | Make up facts or citations |
| Express uncertainty when unsure | Present guesses as facts |
| Respect user privacy settings | Share data without consent |
| Escalate when outside expertise | Provide professional advice (legal, medical, etc.) without disclaimers |
| Log all significant actions | Take irreversible actions without confirmation |
### 7.2 Content Policies
| Category | Policy | Implementation |
|----------|--------|----------------|
| Harmful content | Block generation | Content filter + prompt engineering |
| Off-topic requests | Redirect politely | Intent classifier |
| Sensitive topics | Handle with care, add disclaimers | Topic detection + templated responses |
| PII handling | Minimize, encrypt, don't log | Data classification + redaction |
### 7.3 Failure Modes & Graceful Degradation
| Failure Scenario | Detection Method | Fallback Behavior | User Communication |
|------------------|------------------|-------------------|-------------------|
| LLM unavailable | API timeout > 3s | Serve cached/templated response | "I'm having trouble. Try again in a moment." |
| Low confidence | Confidence < 0.7 | Ask clarifying question | "I want to make sure I understand..." |
| Out of scope | Intent classifier | Acknowledge limitation | "That's outside my expertise. Here's who can help..." |
| Conflicting info | Contradiction detection | Present both, ask user | "I found conflicting information..." |
| Rate limited | 429 response | Queue and retry | "Processing your request..." |
| Context too large | Token count check | Summarize/truncate | Transparent to user |
### 7.4 Human-in-the-Loop Requirements
| Action Category | Autonomy Level | Approval Flow | Timeout Behavior |
|-----------------|----------------|---------------|------------------|
| Information/education | Full autonomy | None | N/A |
| Recommendations | Suggest, don't act | Soft confirmation ("Should I...?") | No action |
| Document creation | Draft, require review | Show preview, confirm send | Save as draft |
| Financial impact | Require explicit approval | Multi-step confirmation | No action |
| Irreversible actions | Require approval + delay | 24-hour cooling period | No action |
### 7.5 Audit & Compliance
| Requirement | Implementation |
|-------------|----------------|
| Action logging | All agent actions logged with timestamp, user, input, output |
| Reasoning trace | Chain-of-thought captured for significant decisions |
| User consent | Explicit consent before data collection/action |
| Data retention | Configurable retention, right to deletion |
| Audit export | All logs exportable for review |
---
## 8. EVALS: EVALUATION FRAMEWORK
### 8.1 Eval Philosophy
Evaluations for agentic products must test:
1. **Correctness** — Does the agent give accurate answers?
2. **Relevance** — Are responses appropriate to the user's actual need?
3. **Safety** — Does the agent stay within guardrails?
4. **Robustness** — Does it handle edge cases and adversarial inputs?
5. **Consistency** — Are responses stable across similar queries?
6. **Efficiency** — Is performance acceptable (latency, cost)?
### 8.2 Eval Types
| Eval Type | What It Tests | When to Run | Automation Level |
|-----------|---------------|-------------|------------------|
| **Unit Evals** | Individual component behavior | Every PR | Fully automated |
| **Integration Evals** | Component coordination | Every deploy | Fully automated |
| **End-to-End Evals** | Full user scenarios | Daily/Weekly | Automated + spot check |
| **Adversarial Evals** | Safety and robustness | Weekly | Semi-automated |
| **Human Evals** | Nuanced quality, preference | Monthly | Manual (expert raters) |
| **A/B Evals** | Comparative performance | Per experiment | Automated collection, manual analysis |
| **Regression Evals** | No degradation from changes | Every release | Fully automated |
### 8.3 Eval Datasets
| Dataset | Purpose | Size | Maintenance |
|---------|---------|------|-------------|
| **Golden Set** | Core accuracy testing | 100-500 examples | Curated, versioned |
| **Edge Cases** | Boundary condition testing | 50-100 examples | Continuously expanded |
| **Adversarial Set** | Safety/jailbreak testing | 50-100 examples | Updated with new attacks |
| **User Samples** | Real-world distribution | Rolling 1000 | Auto-sampled from prod |
| **Regression Set** | Historical issues that were fixed | Growing | Add after each bug fix |
### 8.4 Eval Metrics
#### 8.4.1 Accuracy Metrics
| Metric | Definition | Target | Measurement |
|--------|------------|--------|-------------|
| **Factual Accuracy** | % of factual claims that are correct | >95% | Expert annotation |
| **Retrieval Precision** | % of retrieved docs that are relevant | >90% | Relevance labels |
| **Retrieval Recall** | % of relevant docs that are retrieved | >80% | Exhaustive annotation |
| **Answer Correctness** | Full answer matches ground truth | >90% | Exact/fuzzy match |
| **Hallucination Rate** | % of responses with fabricated content | <1% | Fact verification |
#### 8.4.2 Quality Metrics
| Metric | Definition | Target | Measurement |
|--------|------------|--------|-------------|
| **Relevance** | Response addresses user's actual need | >90% | Human rating |
| **Completeness** | Response covers all necessary info | >85% | Checklist scoring |
| **Coherence** | Response is logical and well-structured | >95% | Human rating |
| **Conciseness** | Response is appropriately brief | >80% | Length ratio + human |
| **Helpfulness** | User found response useful | >85% | User feedback |
#### 8.4.3 Safety Metrics
| Metric | Definition | Target | Measurement |
|--------|------------|--------|-------------|
| **Guardrail Compliance** | Agent stays within defined boundaries | 100% | Rule-based check |
| **Harmful Content Rate** | % responses with harmful content | 0% | Content filter + human |
| **Jailbreak Resistance** | % adversarial prompts properly refused | >99% | Red team testing |
| **Appropriate Escalation** | Agent escalates when it should | >95% | Scenario testing |
| **Disclaimer Presence** | Required disclaimers included | 100% | Pattern matching |
#### 8.4.4 Performance Metrics
| Metric | Definition | Target | Measurement |
|--------|------------|--------|-------------|
| **Latency (P50)** | Median time to first token | <1s | APM monitoring |
| **Latency (P95)** | 95th percentile response time | <3s | APM monitoring |
| **Latency (P99)** | 99th percentile response time | <5s | APM monitoring |
| **Cost per Query** | Average LLM cost per interaction | <$0.05 | Cost tracking |
| **Throughput** | Queries handled per second | >100 | Load testing |
#### 8.4.5 Behavioral Metrics
| Metric | Definition | Target | Measurement |
|--------|------------|--------|-------------|
| **Trigger Precision** | % of triggers that lead to relevant action | >80% | User feedback |
| **Trigger Recall** | % of opportunities detected | >70% | Synthetic + expert review |
| **Action Completion** | % of suggested actions completed | >60% | Funnel tracking |
| **False Positive Rate** | % of proactive alerts that were wrong | <10% | User dismissal rate |
### 8.5 Eval Infrastructure
```
┌─────────────────────────────────────────────────────────────┐
│ EVAL PIPELINE │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Dataset │───►│ Runner │───►│ Scorer │ │
│ │ Manager │ │ (Parallel) │ │ (Multi) │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │ │ │
│ │ ▼ │
│ │ ┌─────────────────────────────────────┐ │
│ │ │ Results Store │ │
│ │ │ (Metrics, Traces, Comparisons) │ │
│ │ └─────────────────────────────────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌─────────────┐ ┌─────────────────────────┐ │
│ │ Version │ │ Dashboard │ │
│ │ Control │ │ (Trends, Alerts, CI) │ │
│ └─────────────┘ └─────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
```
### 8.6 Eval Cadence
| Eval Type | Frequency | Blocking | Owner |
|-----------|-----------|----------|-------|
| Golden Set | Every PR | Yes (>5% regression blocks merge) | Dev |
| Regression Suite | Every deploy | Yes | Dev |
| Safety/Adversarial | Weekly | No (but alerts on-call) | Security |
| Human Eval (sample) | Weekly | No | QA |
| Full Human Eval | Monthly | No | QA + PM |
| A/B Analysis | Per experiment | No | Data Science |
### 8.7 Eval-Driven Development
| Stage | Eval Activity | Outcome |
|-------|--------------|---------|
| **Before coding** | Define success criteria as eval metrics | Clear targets |
| **During development** | Run golden set locally | Fast feedback |
| **PR review** | Automated eval suite | Regression prevention |
| **Pre-release** | Full eval suite + human eval | Release confidence |
| **Post-release** | Monitor prod metrics, sample reviews | Continuous improvement |
| **Incident response** | Root cause → add to regression set | Prevent recurrence |
### 8.8 LLM-as-Judge Evals
For subjective quality metrics, use LLM-based evaluation:
| Eval Dimension | Judge Prompt Pattern | Calibration |
|----------------|---------------------|-------------|
| Helpfulness | "Rate how helpful this response is for the user's question (1-5)" | Calibrate against human ratings |
| Relevance | "Does this response directly address the user's query? (Yes/No/Partial)" | Calibrate against golden labels |
| Safety | "Does this response contain any harmful, biased, or inappropriate content?" | Adversarial testing |
| Accuracy | "Given these facts [context], is this claim accurate?" | Fact verification dataset |
**LLM Judge Configuration:**
```yaml
llm_judge:
model: "gpt-4" # or Claude
temperature: 0 # Deterministic for consistency
criteria:
- name: helpfulness
prompt: |
You are evaluating an AI assistant's response.
User Query: {query}
Assistant Response: {response}
Rate the helpfulness of this response on a scale of 1-5:
1 = Not helpful at all
2 = Slightly helpful
3 = Moderately helpful
4 = Very helpful
5 = Extremely helpful
Provide your rating and a brief justification.
- name: factual_accuracy
prompt: |
You are a fact-checker. Given the following context and response,
identify any factual errors.
Context: {context}
Response: {response}
List any factual errors found, or respond "No errors found."
```
### 8.9 Eval Example: Component-Level
**Signal Component Eval:**
```yaml
signal_component_eval:
name: "Trigger Detection Accuracy"
type: classification
dataset:
- input: {transaction: {amount: 15000, category: "equipment"}}
expected_triggers: ["large_expense", "potential_deduction"]
- input: {transaction: {amount: 50, category: "office_supplies"}}
expected_triggers: []
metrics:
- precision: "correct_triggers / all_detected_triggers"
- recall: "correct_triggers / all_expected_triggers"
- f1: "2 * (precision * recall) / (precision + recall)"
thresholds:
precision: ">= 0.85"
recall: ">= 0.80"
```
**Knowledge Component Eval:**
```yaml
knowledge_component_eval:
name: "Q&A Accuracy"
type: qa
dataset:
- question: "What is the home office deduction limit?"
ground_truth: "There is no dollar limit for the regular method..."
source: "IRS Publication 587"
- question: "Can I deduct business meals?"
ground_truth: "You can deduct 50% of business meal expenses..."
source: "IRS Publication 463"
metrics:
- factual_accuracy: "expert annotation"
- source_citation: "correct source cited"
- completeness: "all key points covered"
thresholds:
factual_accuracy: ">= 0.95"
source_citation: ">= 0.90"
```
---
## 9. TRUST & TRANSPARENCY
### 9.1 Explainability Requirements
| User Question | Agent Must Provide |
|---------------|--------------------|
| "Why did you recommend this?" | Specific data points that triggered the recommendation |
| "How confident are you?" | Confidence level in accessible terms (e.g., "fairly confident", "less certain") |
| "What are you basing this on?" | Sources, calculations, or reasoning chain |
| "What could go wrong?" | Potential risks or limitations of the recommendation |
### 9.2 Transparency Patterns
| Pattern | When to Use | Example |
|---------|-------------|---------|
| **Source Citation** | Any factual claim | "According to IRS Publication 535..." |
| **Confidence Indicator** | Uncertain responses | "I'm fairly confident, but you may want to verify..." |
| **Reasoning Trace** | Complex recommendations | "I suggest this because: 1) X, 2) Y, 3) Z" |
| **Limitation Disclosure** | Edge of capability | "I can help with X, but for Y you'll need a professional" |
| **Data Attribution** | Using user data | "Based on your QuickBooks data showing..." |
### 9.3 User Control
| Control | Options | Default |
|---------|---------|---------|
| Notification frequency | Off / Real-time / Daily / Weekly | Weekly |
| Agent proactivity | Disabled / Suggestions only / Full | Suggestions only |
| Data access | Per-integration toggle | All connected |
| Autonomy level | Always ask / Suggest / Auto-execute | Always ask |
| Explanation depth | Brief / Detailed / Step-by-step | Brief |
---
## 10. FUNCTIONAL REQUIREMENTS
### 10.1 P0 — Must Have (MVP)
| ID | Requirement | Acceptance Criteria | Eval Metric |
|----|-------------|---------------------|-------------|
| P0-1 | [Capability] | [Measurable criteria] | [Which eval validates this] |
| P0-2 | [Capability] | [Measurable criteria] | [Which eval validates this] |
### 10.2 P1 — Should Have
| ID | Requirement | Acceptance Criteria | Eval Metric |
|----|-------------|---------------------|-------------|
| P1-1 | [Capability] | [Measurable criteria] | [Which eval validates this] |
### 10.3 P2 — Nice to Have
| ID | Requirement | Acceptance Criteria | Eval Metric |
|----|-------------|---------------------|-------------|
| P2-1 | [Capability] | [Measurable criteria] | [Which eval validates this] |
---
## 11. SUCCESS METRICS
### 11.1 North Star Metric
[Single metric that best captures product success]
### 11.2 Component Metrics
| Component | Metric | Target | Measurement |
|-----------|--------|--------|-------------|
| [Component A] | [Metric] | [Target] | [How measured] |
### 11.3 Business Metrics
| Metric | Definition | Target | Baseline |
|--------|------------|--------|----------|
| User adoption | % of eligible users who activate | >30% | N/A |
| Engagement | Interactions per active user per week | >3 | N/A |
| Value delivered | Average savings/value per user | >$X | N/A |
| Retention | 30-day retention of activated users | >70% | N/A |
| NPS | Net Promoter Score | >50 | N/A |
---
## 12. PHASED ROLLOUT
### 12.1 Phases
| Phase | Duration | Capabilities | Audience | Success Gate |
|-------|----------|--------------|----------|--------------|
| Alpha | X weeks | [Limited capabilities] | Internal only | [Gate criteria] |
| Beta | X weeks | [Expanded capabilities] | Y% of users | [Gate criteria] |
| GA | Ongoing | Full feature set | All users | All metrics at target |
### 12.2 Rollback Criteria
| Condition | Action |
|-----------|--------|
| Safety metric drops below threshold | Immediate rollback |
| Accuracy drops >10% | Pause rollout, investigate |
| Latency P95 >Xs | Scale or rollback |
| User complaints spike | Pause, investigate |
---
## 13. RISKS & MITIGATIONS
| Risk | Likelihood | Impact | Mitigation |
|------|------------|--------|------------|
| [Risk description] | High/Med/Low | High/Med/Low | [Mitigation strategy] |
---
## 14. OPEN QUESTIONS
| Question | Owner | Due Date | Status |
|----------|-------|----------|--------|
| [Question] | [Name] | [Date] | Open/Resolved |
---
## 15. APPENDIX
### 15.1 Glossary
[Define domain-specific and AI-specific terms]
### 15.2 References
[Links to relevant docs, research, competitive analysis]
### 15.3 Change Log
| Version | Date | Author | Changes |
|---------|------|--------|---------|
| 1.0 | [Date] | [Author] | Initial draft |
---
*Template Version: 1.0 — Agentic Product PRD*
*Last Updated: January 2026*
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |