Product Discovery & Market Research

A product manager skill that discovers market needs through parallel deep research. It runs 3 specialized agents concurrently — competitor analysis, user pain point mining, and trend tracking — then synthesizes findings into 5-10 prioritized feature recommendations. Research results are cached daily to docs/research/ to avoid redundant agent calls.

Process

Phase 1: Scope Clarification

Before launching research, ask the user one question at a time to establish:

1. Product direction — What are you building? (If not provided as argument)
2. Target users — Who is this for?

Keep it to 1-2 questions max. Extract search keywords from the answers for the agents.

Phase 1.5: Load Previously Rejected Ideas

Before launching research, check for previously rejected or deferred ideas:

1. Use the Glob tool to scan docs/rejected/*.md (skip README.md)
2. Read each file's YAML frontmatter to extract: idea, status, reason
3. Build two lists:
   • Rejected list: Ideas with status: rejected — these MUST be excluded from recommendations
   • Deferred list: Ideas with status: deferred — these may only be re-suggested if there is strong new evidence of changed market conditions

If no files exist in docs/rejected/, skip this phase and proceed to Phase 2.

Store these lists for use in Phase 2 agent prompts.

Phase 2: Parallel Deep Research (with caching)

Step 1: Check research cache

Before launching agents, check if today's research already exists:

1. Derive a {topic-slug} from the product description in lowercase kebab-case (e.g. "CLI developer tools" → cli-developer-tools)
2. Use Glob to check for docs/research/{YYYY-MM-DD}-{topic-slug}.md (using today's date)
3. Cache hit: Read the cached file and skip directly to Phase 3. Do NOT launch any agents.
4. Cache miss: Continue to Step 2 below.

Step 2: Launch agents (cache miss only)

Launch 3 agents in parallel using the Task tool with subagent_type: "general-purpose". Each agent uses WebSearch extensively.

IMPORTANT: Launch all 3 agents in a single message to maximize parallelism.

IMPORTANT: If Phase 1.5 produced any rejected or deferred ideas, append the following block to each agent's prompt (before the 800-word constraint line):

IMPORTANT: The following ideas have been previously evaluated and REJECTED by our team. Do NOT recommend features similar to these:
{for each rejected idea: "- {idea_name} (reason: {reason})"}

The following ideas were DEFERRED (not right now). Only mention them if you find strong evidence that market conditions have significantly changed since {date}:
{for each deferred idea: "- {idea_name} (deferred on: {date}, reason: {reason})"}

If there are no rejected or deferred ideas, omit this block entirely.

CRITICAL: Each agent prompt MUST end with this constraint:

Keep your response under 800 words. Return ONLY a structured summary — no preamble, no methodology explanation. Focus on actionable findings with specific data points.

CRITICAL: After ALL agents return, you MUST immediately proceed to Step 3 (cache write) and then Phase 3 (Synthesis) in the SAME response. Do NOT stop or wait for user input between Phase 2 and Phase 4.

Step 3: Write research cache (cache miss only)

After all 3 agents return, write their combined output to docs/research/{YYYY-MM-DD}-{topic-slug}.md:

---
topic: "{product_description}"
target_users: "{target_users}"
date: YYYY-MM-DD
---
## Competitor Analysis
{agent 1 output}

## User Pain Points
{agent 2 output}

## Market Trends
{agent 3 output}

Create the docs/research/ directory if it doesn't exist. Then proceed to Phase 3.

Agent 1: Competitor Analyst

Prompt template:

You are a competitor analyst. Research the competitive landscape for: {product_description}

Target users: {target_users}

Do the following:
1. Use WebSearch to find 5-8 competing products. Search for:
   - "{product_type} alternatives 2026"
   - "{product_type} best tools"
   - "{product_type} comparison"
2. For each competitor, search for their features and pricing:
   - "{competitor_name} features pricing"
   - "{competitor_name} review"
3. Compile a competitor analysis table with columns:
   - Name | Core Features | Pricing | Strengths | Weaknesses | Popularity

Return ONLY the structured analysis. Be specific with facts and data.

Keep your response under 800 words. Return ONLY a structured summary — no preamble, no methodology explanation. Focus on actionable findings with specific data points.

Agent 2: User Pain Analyst

Prompt template:

You are a user pain point researcher. Find real user complaints and unmet needs for: {product_description}

Target users: {target_users}

Do the following:
1. Use WebSearch to find user discussions. Search for:
   - "{product_type} pain points reddit"
   - "{product_type} frustrating"
   - "{product_type} missing features"
   - "{product_type} wish list forum"
   - "{product_type} complaints"
2. For each pain point found, note:
   - What the pain point is
   - How frequently it's mentioned (high/medium/low)
   - Source (Reddit, forum, review site, etc.)
   - Whether any existing product has solved it

Return a structured list of pain points ranked by frequency. Be specific — quote real user feedback when possible.

Keep your response under 800 words. Return ONLY a structured summary — no preamble, no methodology explanation. Focus on actionable findings with specific data points.

Agent 3: Trend Analyst

Prompt template:

You are a technology and market trend analyst. Research industry trends for: {product_description}

Target users: {target_users}

Do the following:
1. Use WebSearch to find trends. Search for:
   - "{product_type} trends 2026"
   - "{product_type} emerging technology"
   - "{product_type} market growth"
   - "{product_type} future"
   - "{related_technology} adoption rate"
2. For each trend, note:
   - Trend direction
   - Maturity stage (early/growing/mature)
   - Related technologies
   - Potential opportunity for a new product

Return a structured trend analysis. Focus on actionable insights, not hype.

Keep your response under 800 words. Return ONLY a structured summary — no preamble, no methodology explanation. Focus on actionable findings with specific data points.

Phase 3: Synthesis (MUST happen in the same response as Phase 2 results)

CRITICAL: Do NOT stop after agents return. Immediately synthesize and present findings in a single response. Never wait for user input between Phase 2 and Phase 4.

After all 3 agents return, synthesize their findings:

1. Cross-validate — Needs mentioned across multiple agents get higher weight:

   • Appears in competitor gaps AND user pain points → strong signal
   • Appears in user pain points AND aligns with trends → strong signal
   • Appears in all three → highest priority

2. Score each potential feature on 4 dimensions:

   • Market demand (0-3): How many users want this?
   • Competitive gap (0-3): How poorly do competitors serve this?
   • Trend alignment (0-3): Does this match where the market is going?
   • Feasibility (0-3): How realistic is this to build?

3. Select top 5-10 features by total score.

Phase 4: Collaborative Decision

Present findings to the user in this format:

## Market Research Results: {product_description}

### Key Findings

**Competitive Landscape**: {1-2 sentence summary}
**User Pain Points**: {1-2 sentence summary}
**Market Trends**: {1-2 sentence summary}

### Recommended Features

#### 1. {Feature Name} — Priority: HIGH
- **What**: {feature description}
- **Why**: {market evidence from agents}
- **Competitors**: {who does/doesn't have this}
- **Score**: Demand {x}/3 | Gap {x}/3 | Trend {x}/3 | Feasibility {x}/3 = {total}/12

#### 2. {Feature Name} — Priority: {HIGH/MEDIUM}
...

... (5-10 features total, ranked by score)

### What should we build first?

Then engage in discussion:

• Answer follow-up questions about any finding
• Adjust recommendations if user provides additional context
• Help narrow down to a final decision

Phase 4.5: Record Review Decisions

After the user has made their decisions on which features to pursue, record rejected and deferred ideas:

1. Ask for each recommended feature — Use AskUserQuestion to ask the user's decision for each feature that they did NOT choose to pursue:

   • Options: "Reject permanently" / "Defer for later" / "Already accepted"
   • For rejected/deferred: ask reason with options: "scope" / "direction" / "timing" / "duplicate" / "other"
   • If "other", ask for a brief explanation

2. Write rejection files — For each rejected or deferred idea, use the Write tool to create docs/rejected/{YYYY-MM-DD}-{idea-slug}.md:

   ---
   status: {rejected|deferred}
   idea: "{Feature Name}"
   date: {YYYY-MM-DD}
   score: {total_score}
   reason: "{reason}"
   ---
   
   ## Reason
   
   {User's explanation or generated summary based on reason category}
   
   ## Original Evidence Summary
   
   {2-3 bullet points from the Phase 3 synthesis for this feature}
   

   The {idea-slug} should be the feature name in lowercase-kebab-case (e.g., "Real-time Collaboration" → "real-time-collaboration").

3. Commit — Stage and commit all new rejection files:

   git add docs/rejected/
   git commit -m "docs: record rejected/deferred ideas from PM research"
   

4. Confirm — Tell the user which ideas were recorded and where the files are saved.

Key Rules

• Always launch 3 agents in parallel — never sequentially
• One question at a time in Phase 1 — don't overwhelm
• 5-10 features — broad enough for ideation, scored and ranked
• Cache research daily — Write to docs/research/, reuse same-day same-topic results
• Evidence-based — every recommendation must cite specific findings
• No implementation — this skill discovers WHAT to build, not HOW
• Check rejected ideas first — Always run Phase 1.5 before research; never skip it
• Record every decision — Every idea that isn't accepted must be recorded in Phase 4.5 as rejected or deferred