@shaykec/bridge 0.4.25 → 0.4.27

package/modules/ai-agents/walkthrough.md

@@ -0,0 +1,85 @@
+# AI Agents Walkthrough — Learn by Doing
+
+## Before We Begin
+
+**Diagnostic Question:** When an LLM uses a "tool" (like a search API or calculator), how is that different from the model just generating text? What has to happen for the tool to be useful?
+
+**Checkpoint:** You recognize that tools give models access to external data and actions. The model must decide when to call a tool, what to pass, and how to use the result.
+
+---
+
+## Step 1: Define a Simple Tool
+
+<!-- hint:code language="json" highlight="1,3" -->
+
+**Task:** Design a tool for an agent that looks up the current weather. Specify: name, description, and parameters (e.g., city, unit). Write it as a JSON schema. Why is the description important?
+
+**Question:** What would happen if the description were vague? How would the model choose between this tool and a "get_forecast" tool?
+
+**Checkpoint:** The user defines a tool with clear name, description, and parameters. They understand the model uses the description to decide when to call the tool.
+
+---
+
+## Step 2: Trace the Agent Loop
+
+<!-- hint:diagram mermaid-type="flowchart" topic="observe think act loop" -->
+
+**Task:** A user asks: "What's the population of Tokyo?" The agent has a `search_web` tool. Trace the loop: (1) Observe — what does the agent see? (2) Think — what does it decide? (3) Act — what tool call? (4) Observe — what comes back? (5) Think/Act — what next?
+
+**Question:** When would the loop stop? What if the search returned no results?
+
+**Checkpoint:** The user traces observe → think → act → observe → respond. They can handle the "no results" case (retry, different query, or admit uncertainty).
+
+---
+
+## Step 3: Design a Multi-Step Workflow
+
+**Task:** Design an agent workflow for "Summarize the top 3 results from a web search for X." List the tools needed, the order of operations, and what the agent outputs at each step. Draw a simple flow.
+
+**Question:** Where could the agent get stuck? What guardrails would you add?
+
+**Checkpoint:** The user designs: search → get top 3 → fetch content → summarize. They identify failure modes (empty results, timeout) and suggest retries or fallbacks.
+
+---
+
+## Step 4: Compare ReAct vs Pre-Planning
+
+<!-- hint:buttons type="single" prompt="Which approach adapts better to surprises?" options="ReAct,Pre-planning,Both equally" -->
+
+**Task:** For the task "Find the GDP of Japan and France, then compare them", sketch two approaches: (A) ReAct — interleave reasoning and tool calls. (B) Pre-plan — write a 3-step plan first, then execute. What are the tradeoffs?
+
+**Question:** When might pre-planning fail? When might ReAct be inefficient?
+
+**Checkpoint:** The user contrasts both approaches. They note: ReAct adapts to surprises; pre-planning can be wrong. ReAct may make redundant calls; pre-planning may miss steps.
+
+---
+
+## Step 5: Map MCP to Your Use Case
+
+<!-- hint:card type="concept" title="MCP (Model Context Protocol)" -->
+
+**Task:** Pick a tool you'd want an agent to use (e.g., read files, query a DB, call an API). Describe how it would look as an MCP tool: name, description, inputs, outputs. What does MCP give you that a custom integration doesn't?
+
+**Question:** How would MCP help if you had 5 different tools from 5 vendors?
+
+**Checkpoint:** The user describes one MCP tool with schema. They understand MCP provides a standard interface so any MCP client can use any MCP server.
+
+---
+
+## Step 6: Add Safety Constraints
+
+**Task:** Your agent can run shell commands and edit files. List 3 safety measures you'd implement. For each, say what it prevents and what the tradeoff is (e.g., slower, less flexible).
+
+**Question:** Where would you require human approval? What would you never allow?
+
+**Checkpoint:** The user proposes: sandbox (no network), allowlist (only certain commands), human approval for destructive ops. They can justify each and name a "never allow" (e.g., `rm -rf /`).
+
+---
+
+## Step 7: Design a Multi-Agent System
+
+**Task:** Design a 3-agent system for "Analyze this codebase and suggest improvements." Define: (1) Orchestrator role, (2) Two specialist agents with distinct tools, (3) How they hand off work and merge results.
+
+**Question:** What does the orchestrator need to "know" to delegate well? How do you avoid duplicate work?
+
+**Checkpoint:** The user defines orchestrator + 2 specialists (e.g., Analyzer + Writer). They describe handoff (orchestrator assigns files, specialists return findings) and a simple merge strategy.

package/modules/ai-assisted-research/content.md

@@ -0,0 +1,136 @@
+# AI-Assisted Product Research — Using Claude for Discovery
+
+<!-- hint:slides topic="AI-assisted research: use cases, prompting for synthesis, validation techniques, and ethical considerations" slides="5" -->
+
+## AI as a Research Amplifier
+
+AI doesn't replace human judgment—it amplifies it. Product research still requires real users, real data, and real empathy. AI helps you work faster: synthesize feedback, structure analysis, generate hypotheses, and draft frameworks.
+
+Think of Claude as a **research assistant**, not a substitute for discovery.
+
+## Use Cases
+
+| Use Case | How AI Helps | Human Still Does |
+|----------|--------------|------------------|
+| **Competitor analysis** | Summarize feature matrices, positioning, reviews | Validate with real usage, interpret nuance |
+| **User feedback synthesis** | Cluster themes, extract quotes, identify patterns | Conduct interviews, sense emotional tone |
+| **Market sizing** | Structure TAM/SAM/SOM, suggest data sources | Validate assumptions, fill gaps |
+| **Persona generation** | Draft personas from aggregated data | Ground in real user research |
+| **Survey analysis** | Code open-ended responses, surface themes | Design survey, interpret outliers |
+
+## Prompting Techniques for Research
+
+### 1. Provide Context
+
+Give Claude domain, audience, and constraints upfront.
+
+```
+I'm a PM at a B2B SaaS company. Our product helps sales teams track deals.
+I'm researching how mid-market sales managers (50–200 reps) prioritize their week.
+Keep responses focused on that segment.
+```
+
+### 2. Ask for Structured Output
+
+Request formats that are easy to act on: tables, bullet lists, JSON.
+
+```
+Synthesize these 10 user feedback quotes into a table: Theme | Representative Quote | Frequency
+```
+
+### 3. Iterate
+
+Use follow-ups to refine: "Narrow to pain points only," "Add a confidence score," "Reorder by severity."
+
+### 4. Request Citations and Assumptions
+
+Ask Claude to distinguish what it inferred vs what you provided.
+
+```
+For each insight, label: [From your data] or [Inferred assumption]. 
+Flag any assumption that needs validation.
+```
+
+## Validating AI-Generated Insights
+
+AI can be confidently wrong. Validation steps:
+
+1. **Trace to source** — Can you point to user data that supports this?
+2. **Check for bias** — Does the model over-weight common patterns and miss edge cases?
+3. **Triangulate** — Does it match qualitative interviews or quantitative data?
+4. **Pressure-test** — What would disprove this? What evidence would you need?
+
+## Combining AI Research With Real Interviews
+
+```mermaid
+flowchart LR
+  A[AI Synthesis] --> B[Themes & Hypotheses]
+  C[User Interviews] --> B
+  B --> D[Validated Insights]
+  D --> E[PRD / Strategy]
+```
+
+**Workflow:**
+1. Use AI to pre-synthesize existing feedback → draft themes.
+2. Use those themes to shape interview guides.
+3. Run interviews; capture raw notes.
+4. Use AI to cluster and extract from notes.
+5. Human validates: Do the clusters match what you heard? Refine.
+
+## Ethical Considerations
+
+- **Privacy** — Don't paste PII (emails, names) into AI. Anonymize first.
+- **Confidentiality** — Don't upload internal roadmaps or unreleased plans to public AI.
+- **Attribution** — AI summaries are aids, not sources. Don't claim AI output as primary research.
+- **Bias** — AI reflects training data. Watch for demographic or segment bias in synthesis.
+
+## Limitations to Watch For
+
+| Limitation | Mitigation |
+|-----------|------------|
+| **Hallucination** | Always verify facts, URLs, statistics |
+| **Recency** | Check model cutoff; supplement with fresh data |
+| **Over-generalization** | Ask for segment-specific views; validate with real users |
+| **Missing nuance** | AI smooths over edge cases; manual review for outliers |
+| **Anchoring** | Don't let AI themes replace your own reading of raw data |
+
+## AI-Augmented Research Workflow
+
+```mermaid
+flowchart TB
+  subgraph Input
+    I1[Raw feedback]
+    I2[Interview notes]
+    I3[Competitor pages]
+  end
+  
+  subgraph AI
+    A1[Synthesis]
+    A2[Theme extraction]
+    A3[Structured output]
+  end
+  
+  subgraph Human
+    H1[Validate]
+    H2[Interview]
+    H3[Decide]
+  end
+  
+  I1 --> A1
+  I2 --> A1
+  I3 --> A2
+  A1 --> A3
+  A3 --> H1
+  H1 --> H2
+  H2 --> H3
+```
+
+1. **Gather** — Raw data (feedback, notes, competitor info)
+2. **AI Synthesis** — Themes, tables, hypotheses
+3. **Validate** — Human checks against sources
+4. **Interview** — Fill gaps, test hypotheses
+5. **Decide** — PRD, prioritization, strategy
+
+---
+
+AI makes you faster; it doesn't make you right. Use it to scale your analysis, then validate with real users and real data.

package/modules/ai-assisted-research/exercises.md

@@ -0,0 +1,80 @@
+# AI-Assisted Research — Exercises
+
+## Exercise 1: Theme Extraction Prompt
+
+**Task:** Write a prompt that takes 5 survey responses (open-ended) and outputs: (1) 3–5 themes, (2) one representative quote per theme, (3) a confidence level (High/Medium/Low) for each theme based on how many responses support it.
+
+**Validation:**
+- [ ] Prompt includes context (e.g., product domain)
+- [ ] Output format is specified (bullets, table, etc.)
+- [ ] Confidence is tied to evidence count, not AI certainty
+- [ ] Prompt is under 200 words
+
+**Hints:**
+1. "You are a product researcher synthesizing user feedback."
+2. "For each theme, cite the number of responses that mention it."
+3. "Low confidence = 1 response, Medium = 2, High = 3+."
+
+---
+
+## Exercise 2: Competitor Feature Matrix
+
+**Task:** Choose a product category (e.g., project management, CRM, note-taking). Write a prompt that produces a 5x4 table: 5 features × 4 products (3 competitors + "Our product"). Include an instruction to note "Unknown" where the AI cannot verify, and to add a "Last verified" caveat.
+
+**Validation:**
+- [ ] Table structure is specified
+- [ ] "Unknown" handling is explicit
+- [ ] Recency/verification caveat is included
+- [ ] You could run this and manually verify a few cells
+
+**Hints:**
+1. List the 5 features and 4 products explicitly
+2. "If you cannot confirm a feature, write 'Unknown (verify)'"
+3. "Add a footer: 'Verify pricing and features on official sites'"
+
+---
+
+## Exercise 3: Assumption vs Evidence Labels
+
+**Task:** Take an AI-synthesized output (from a previous exercise or a sample). For each insight, label it: [E] Evidence-based (traceable to input data) or [A] Assumption (AI inferred). Write 1–2 sentences explaining one [A] and how you'd validate it.
+
+**Validation:**
+- [ ] Every insight has a label
+- [ ] At least one [A] is identified
+- [ ] Validation approach is concrete (interview question, data source, etc.)
+
+**Hints:**
+1. Evidence = "User said X in quote 3"
+2. Assumption = "Users likely want Y because..." with no direct quote
+3. Validation: "I would ask: 'When you do Z, what frustrates you?'"
+
+---
+
+## Exercise 4: Persona Draft + Grounding
+
+**Task:** Use AI to draft a persona for "mid-market sales managers." Then write a "Grounding Checklist": 3 things you would need to validate with real users before using this persona in a PRD. Include one thing the AI likely got wrong or oversimplified.
+
+**Validation:**
+- [ ] Persona has at least: role, goals, frustrations, context
+- [ ] Grounding checklist has 3 concrete validation steps
+- [ ] At least one likely AI error or oversimplification is named
+
+**Hints:**
+1. AI personas tend to be "average" — real users have more variance
+2. Grounding: "Interview 5 people who match this role"
+3. Oversimplification: e.g., "All sales managers care about X" — segment differences?
+
+---
+
+## Exercise 5: End-to-End Research Loop
+
+**Task:** Simulate a mini research loop: (1) Paste 5 feedback quotes, (2) Run an AI synthesis prompt, (3) Write 2 interview questions to validate the top theme, (4) Write a one-paragraph "Research Summary" that states what you learned, what's assumed, and what you'd do next.
+
+**Validation:**
+- [ ] All 4 steps completed
+- [ ] Summary distinguishes learned vs assumed
+- [ ] Next steps are specific (not "do more research")
+
+**Hints:**
+1. Summary template: "We learned X (from data). We assume Y (needs validation). Next: Z."
+2. Next steps: "Interview 3 users about theme A" or "Check analytics for metric B"

package/modules/ai-assisted-research/module.yaml

@@ -0,0 +1,42 @@
+slug: ai-assisted-research
+title: "AI-Assisted Product Research — Using Claude for Discovery"
+version: 1.0.0
+description: "Use AI to amplify product research—synthesis, competitive analysis, and validation."
+category: product-management
+tags: [ai, product-research, user-research, claude, discovery, analysis]
+difficulty: intermediate
+
+xp:
+  read: 15
+  walkthrough: 40
+  exercise: 25
+  quiz: 20
+  quiz-perfect-bonus: 10
+
+time:
+  quick: 5
+  read: 20
+  guided: 50
+
+prerequisites: [writing-prds]
+related: [prompt-engineering, ai-pair-programming]
+
+triggers:
+  - "How can I use AI for product research?"
+  - "How do I analyze user feedback with AI?"
+  - "Can Claude help with competitive analysis?"
+  - "How do I use AI to synthesize research?"
+
+visuals:
+  diagrams: [diagram-mermaid, diagram-flow]
+  quiz-types: [quiz-timed-choice, quiz-matching]
+  playground: bash
+  slides: true
+
+sources:
+  - url: "https://docs.anthropic.com/en/docs/overview"
+    label: "Anthropic Claude API Docs"
+    type: docs
+  - url: "https://svpg.com/books/transformed-moving-to-the-product-operating-model"
+    label: "SVPG Transformed"
+    type: book

package/modules/ai-assisted-research/quick-ref.md

@@ -0,0 +1,67 @@
+# AI-Assisted Research Quick Reference
+
+## AI as Amplifier, Not Replacement
+
+| AI Does Well | Human Still Essential |
+|--------------|----------------------|
+| Synthesize large feedback sets | Conduct interviews |
+| Cluster themes, extract quotes | Interpret emotion, nuance |
+| Draft competitor matrices | Verify facts, pricing |
+| Structure TAM/SAM/SOM | Validate assumptions |
+| Generate persona drafts | Ground in real users |
+
+## Prompting Checklist
+
+- [ ] **Context** — Role, product, audience
+- [ ] **Output format** — Table, bullets, JSON
+- [ ] **Citation** — "Cite source" or "Label [Evidence] vs [Inferred]"
+- [ ] **Constraints** — Segment, time period, focus area
+- [ ] **Recency caveat** — "Note if info may be outdated"
+
+## Validation Steps
+
+1. **Trace to source** — Can you point to user data?
+2. **Check bias** — Over-weighting common patterns?
+3. **Triangulate** — Matches interviews or analytics?
+4. **Pressure-test** — What would disprove this?
+
+## Ethical Checklist
+
+- [ ] PII removed or anonymized before pasting
+- [ ] No confidential roadmaps in public AI
+- [ ] AI output not claimed as primary research
+- [ ] Watch for demographic/segment bias
+
+## AI-Augmented Workflow
+
+```
+Gather → AI Synthesis → Validate → Interview → Decide
+```
+
+## Limitations
+
+| Limitation | Mitigation |
+|------------|------------|
+| Hallucination | Verify facts, request citations |
+| Recency | Check model cutoff, add fresh data |
+| Over-generalization | Segment-specific prompts, validate with users |
+| Missing nuance | Manual review of raw data, outliers |
+| Anchoring | Read raw data yourself first |
+
+## Sample Synthesis Prompt
+
+```
+I'm a PM for [product]. I have [N] user feedback quotes below.
+Synthesize into themes. For each theme:
+- Name
+- 1 representative quote
+- Count of supporting quotes
+- [E] if from data, [A] if inferred (flag [A] for validation)
+```
+
+## Research Readiness Template
+
+- **Validated:** [What we confirmed with data/users]
+- **Assumed:** [What AI inferred, needs check]
+- **Missing:** [Gaps we couldn't fill]
+- **Next step:** [Concrete validation action]

package/modules/ai-assisted-research/quiz.md

@@ -0,0 +1,73 @@
+# AI-Assisted Research — Quiz
+
+## Question 1
+
+What is the best way to think about AI's role in product research?
+
+A) AI replaces the need for user interviews
+B) AI amplifies your analysis but doesn't replace human judgment
+C) AI is only useful for quantitative data, not qualitative
+D) AI should be used only for final report writing
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: AI works best as a research assistant—speeding up synthesis, structuring output, generating hypotheses. Human researchers still need to conduct interviews, validate insights, and make judgment calls. AI amplifies, it doesn't replace. -->
+
+## Question 2
+
+When prompting AI for research synthesis, which technique helps reduce hallucination?
+
+A) Asking for longer responses
+B) Requesting structured output (tables, bullets) and citation of sources
+C) Using more creative, open-ended prompts
+D) Running the same prompt multiple times and averaging
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: Structured output (tables, bullets) and explicit requests to cite sources or label [From data] vs [Inferred] make it easier to verify outputs and catch unsupported claims. Creative open-ended prompts can increase hallucination. -->
+
+## Question 3
+
+Before using AI-synthesized insights in a PRD, you should:
+
+A) Accept them as-is since AI is objective
+B) Trace insights to source data and validate with real users
+C) Use them only for internal drafts, not stakeholder-facing docs
+D) Run them through another AI for a second opinion
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: AI can be confidently wrong. Validating means tracing insights to source data and testing them with real users. Internal vs stakeholder-facing doesn't change the need for validation. A second AI can repeat the same biases. -->
+
+## Question 4
+
+Which of these is an ethical concern when using AI for research?
+
+A) AI responses are too slow
+B) Pasting user PII (names, emails) into AI tools
+C) AI sometimes gives short answers
+D) AI requires an internet connection
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: Pasting PII into AI tools risks privacy violations and may violate data policies. Always anonymize before sharing. Speed, answer length, and connectivity are operational concerns, not ethical ones. -->
+
+## Question 5
+
+In the AI-augmented research workflow, where do real user interviews fit?
+
+A) They're optional once AI has synthesized existing data
+B) They validate and fill gaps in AI-generated themes
+C) They should be replaced by AI analysis of support tickets
+D) They happen only before any AI use
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: The workflow is iterative: AI can pre-synthesize existing data → human designs interview guide from themes → interviews happen → AI can help cluster interview notes → human validates. Interviews validate and deepen AI synthesis; they're central, not optional. -->
+
+## Question 6
+
+What is a key limitation of AI for research?
+
+A) AI cannot write in bullet points
+B) AI tends to over-generalize and smooth over edge cases
+C) AI only works with English
+D) AI cannot process more than 1000 words
+
+<!-- ANSWER: B -->
+<!-- EXPLANATION: AI often produces "average" or consensus views and can miss outliers, edge cases, and segment-specific nuances. Manual review of raw data remains important to catch what AI smooths over. Other options are not typical limitations. -->

package/modules/ai-assisted-research/resources.md

@@ -0,0 +1,33 @@
+# AI-Assisted Product Research — Resources
+
+## Videos
+
+- [How to Use ChatGPT for User Research](https://www.youtube.com/results?search_query=chatgpt+user+research) — Various PM channels. Prompting for synthesis and analysis (adapt for Claude).
+- [AI for Product Managers](https://www.youtube.com/results?search_query=AI+product+management) — Product School, Reforge. Broader AI-in-PM context.
+
+## Articles and Readings
+
+- [Building with Claude](https://docs.anthropic.com/en/docs/overview) — Anthropic. Official Claude API docs; prompting and best practices.
+- [How to build your PM second brain with ChatGPT](https://www.lennysnewsletter.com/p/how-to-build-your-pm-second-brain) — Lenny Rachitsky. AI-augmented PM workflows.
+- [Using AI for Product Discovery](https://www.productplan.com/blog/ai-product-discovery/) — ProductPlan. AI in discovery and research.
+- [AI in User Research](https://www.nngroup.com/articles/ai-user-research/) — Nielsen Norman Group. UX research perspective on AI tools.
+- [The Ethics of AI in Product](https://www.svpg.com/) — SVPG. Product ethics and responsibility (search for AI-related posts).
+
+## Books
+
+- **Transformed** by Marty Cagan et al. — SVPG; product operating model and how discovery fits in.
+- **Inspired** by Marty Cagan — Discovery, opportunity assessment, and working with users.
+- **Continuous Discovery Habits** by Teresa Torres — Research cadence and synthesis practices that pair well with AI assistance.
+
+## Podcasts
+
+- [Lenny's Podcast: AI for Product](https://www.lennyspodcast.com/) — Lenny Rachitsky. Episodes on AI tools for PMs.
+- [Product Talk](https://www.producttalk.org/) — Teresa Torres. Discovery and research practices.
+
+## Tools
+
+- [Claude](https://claude.ai/) — Anthropic's AI assistant for synthesis and analysis.
+- [ChatGPT](https://chat.openai.com/) — Alternative for feedback synthesis; compare outputs.
+- [Dovetail](https://dovetail.com/) — User research repository; some AI synthesis features.
+- [Maze](https://maze.co/) — User testing; AI-assisted analysis in some plans.
+- [Notion AI](https://www.notion.so/product/ai) — Summarize research notes and feedback in Notion.

package/modules/ai-assisted-research/walkthrough.md

@@ -0,0 +1,85 @@
+# AI-Assisted Research Walkthrough — Learn by Doing
+
+## Before We Begin
+
+**Diagnostic:** What can AI *not* do well in research? When you feed 50 user feedback snippets to an LLM for synthesis, what might it get wrong—or miss entirely? Why does that matter?
+
+**Checkpoint:** You can name at least two limitations (e.g., confabulation, bias toward common themes, missing nuance) and why human validation is non-negotiable.
+
+---
+
+## Step 1: Prepare Raw Feedback
+
+**Task:** Gather 10–15 real or simulated user feedback snippets (support tickets, survey responses, or review quotes). Paste them into a single document. Remove or anonymize any PII.
+
+**Question:** Before using AI—if you had to synthesize these manually, what would you look for? Themes? Sentiment? Feature requests?
+
+**Checkpoint:** The user has a clean, anonymized set of feedback ready for synthesis.
+
+---
+
+## Step 2: Write a Synthesis Prompt
+
+<!-- hint:code language="markdown" highlight="1,3" -->
+
+**Task:** Write a prompt for Claude that will synthesize your feedback into themes. Include: (1) your role/context, (2) the format you want (e.g., table: Theme | Quote | Count), (3) a request to flag assumptions.
+
+**Question:** How might the prompt change if you wanted to focus only on *pain points* vs *feature requests*? What wording would you add?
+
+**Checkpoint:** The user has a prompt that specifies context, output format, and validation instructions.
+
+---
+
+## Step 3: Run Synthesis and Inspect
+
+<!-- hint:card type="warning" title="AI can infer patterns not grounded in your data—always trace themes back to source quotes" -->
+
+**Task:** Run your prompt with Claude (or another AI). Inspect the output. For each theme: Can you trace it back to at least one source quote? Label any theme that feels like an AI inference vs grounded in your data.
+
+**Question:** Did the AI surface anything you hadn't noticed? Did it miss or oversimplify anything?
+
+**Checkpoint:** The user can distinguish AI-inferred themes from data-backed ones and identify gaps.
+
+---
+
+## Step 4: Validate With a Quick Check
+
+<!-- hint:buttons type="single" prompt="How do you validate an AI-synthesized theme?" options="Trust it; AI is accurate,Interview users; ask directly,Run another AI prompt,Ignore; themes are subjective" -->
+
+**Task:** Pick the top 2 themes from the synthesis. For each, write: (1) one piece of evidence that supports it, (2) one thing that would disprove it, (3) one question you'd ask in an interview to validate it.
+
+**Question:** If you couldn't use AI, would you have arrived at the same themes? What does that tell you about AI's role?
+
+**Checkpoint:** The user has a validation plan for at least 2 themes.
+
+---
+
+## Step 5: Draft a Competitor Analysis Prompt
+
+**Task:** You're analyzing 3 competitors. Write a prompt that asks Claude to produce a feature comparison table. Specify: columns (Feature | Competitor A | B | C | Us), rows (which features to compare), and a request to note where info might be outdated.
+
+**Question:** What would you need to verify manually after getting the table? (Pricing? Feature availability? Positioning?)
+
+**Checkpoint:** The user has a structured competitor-analysis prompt and knows what to validate.
+
+---
+
+## Step 6: Combine AI Output With Human Judgment
+
+**Task:** Take your synthesized themes and draft 3 interview questions that would validate or challenge them. Use the themes to focus the questions—don't let AI replace the need to talk to users.
+
+**Question:** How would you use the interview answers? Would you feed them back into AI for another round of synthesis?
+
+**Checkpoint:** The user connects AI synthesis to a human validation loop.
+
+---
+
+## Step 7: Document Assumptions and Gaps
+
+<!-- hint:celebrate -->
+
+**Task:** Create a short "Research Readiness" checklist for your synthesis: What's validated? What's inferred? What's missing? What's the next step to reduce uncertainty?
+
+**Question:** If a stakeholder asked "How confident are you in these findings?", how would you answer using this checklist?
+
+**Checkpoint:** The user can articulate confidence levels and next validation steps.