pm%3aexplore（pm%3a 探索）

Compatibility: AskUserQuestion falls back to numbered list on non-Claude-Code platforms.

Skill Requirements

command -v npx >/dev/null 2>&1 || echo "MISSING: npx"

Critical Rules

• Never invent market data — every claim must come from a fetched source
• Each research agent writes to its own isolated artifact — never combine within the agent
• Synthesis from sub-artifacts happens only in Step 3, after all agents complete
• One research artifact per topic slug — reuse existing slug directories when extending
• Research artifacts feed pm:plan and pm:prd — deposit them faithfully at the specified paths
• No markdown tables in output — use bullets and definition lists
• Prefer Mermaid diagrams for any visual structure — timelines, competitive maps, process flows. Never use ASCII art for structured diagrams.
• See references/web-research-standards.md for source tiers, citation format, and agent output format

Agent Index

• internal-researcher (agents/internal-researcher.md) — scans project files for prior art, existing plans, and relevant skills
• ideation-researcher (agents/ideation-researcher.md) — surfaces unmet needs, market gaps, and JTBD signals via web search
• market-researcher (agents/market-researcher.md) — validates market size, growth, and existing solution landscape via web search
• competitive-researcher (agents/competitive-researcher.md) — maps the competitive landscape with per-competitor site fetching
• user-problems-researcher (agents/user-problems-researcher.md) — researches jobs-to-be-done and behavioral evidence via web search

Step 0: Parse arguments

TOPIC="${1:-}"
BG_FLAG=false; SILENT=false; DEEP=false
[[ "$*" =~ --bg|-b ]] && BG_FLAG=true
[[ "$*" =~ --silent ]] && SILENT=true
[[ "$*" =~ --deep ]] && DEEP=true

--deep escalates research: more searches, more URLs fetched, stricter source tier requirements.

Step 1: Mode and topic

Ask the user to select exploration mode(s):

AskUserQuestion:
  question: "What kind of exploration do you want to do?"
  header: "Explore mode"
  multiSelect: true
  options:
    - label: "Open-ended ideation"
      description: "Problem-seeking and idea generation — discover what's worth building"
    - label: "Feature/idea validation"
      description: "Validate a specific hypothesis with market research and competitive analysis"
    - label: "Competitor deep-dive"
      description: "Focused analysis of specific named competitors"
    - label: "User problem discovery"
      description: "JTBD-style research — surface jobs, pains, and gains from real sources"

If TOPIC is empty, ask for it. Derive SLUG (lowercase, hyphenated). Set:

• OUTPUT_PATH =.codevoyant/explore/{SLUG}/summary.md
• SUB_DIR =.codevoyant/explore/{SLUG}/research/

Ask mode-specific scoping questions (all in one AskUserQuestion call):

• Open-ended ideation: "What space or user segment are you curious about? Any early hunches to stress-test?"
• Feature/idea validation: "What is the hypothesis? Who is the target user? What would make this validation successful?"
• Competitor deep-dive: "Name the competitors to analyze. Dimensions that matter most: pricing, features, positioning, or all?"
• User problem discovery: "What job are users trying to get done? Which segments should be prioritized?"

Present a one-paragraph scope summary and confirm before launching agents.

Step 2: Launch parallel research agents

mkdir -p ".codevoyant/explore/{SLUG}/research"

Tell the user: "Starting parallel research on '{TOPIC}' (modes: {MODES}) — this will take a few minutes."

Spawn the following agents in a single message (run_in_background: true, model: claude-sonnet-4-6). Substitute all {PLACEHOLDERS} before spawning — agents receive no outer context.

Always spawn:

Agent:
  subagent_type: general-purpose
  model: claude-sonnet-4-6
  run_in_background: true
  description: 'pm-explore/internal: {SLUG}'
  prompt: |
    (full content of agents/internal-researcher.md with {TOPIC} → "{TOPIC}", {SLUG} → "{SLUG}" substituted)

Spawn based on selected modes:

For each agent, substitute into the agent file's prompt: {TOPIC}, {SLUG}, {DEEP}, and any mode-specific variables ({HYPOTHESIS}, {TARGET_USER}, {NAMED_COMPETITORS}, {DIMENSIONS}, {JOB_DESCRIPTION}, {USER_SEGMENTS}) gathered in Step 1. If "Competitor deep-dive" and "Feature/idea validation" are both selected, spawn competitive-researcher only once.

Do not send agent calls across separate messages — all must be in one message to run in parallel. Wait for all agents to complete before continuing.

Step 3: Synthesize and write artifact

Read all sub-artifacts in .codevoyant/explore/{SLUG}/research/. Write a unified research summary to OUTPUT_PATH:

# Research: {TOPIC}

**Modes:** {MODES}
**Date:** {DATE}

## Summary
{2–3 sentence synthesis — what do we now know that we didn't before?}

## Key Findings
{5–8 most important findings across all sub-artifacts, each cited}
- **{Finding}** — {source} [Tier N] [High/Medium/Low confidence]

## Problem Space / JTBD
{Include if ideation or user problem discovery mode — bullets from ideation.md and user-problems.md}

## Market Landscape
{Include if validation mode — bullets from market.md}

## Competitive Analysis
{Include if validation or competitor deep-dive mode — profiles from competitive.md}

## Internal Context
{bullets from internal.md — relevant prior work, plans, PRDs in this repo}

## Gaps and Open Questions
{bullets from all sub-artifacts — questions research couldn't answer}
- [UNVERIFIED] {claims attempted but not sourced}

## Suggested Next Steps
- Use this research with `/pm:plan` to inform roadmap priorities
- Use this research with `/pm:prd {SLUG}` to draft a PRD

Step 4: Notify

if [ "$SILENT" != "true" ]; then
  npx @codevoyant/agent-kit notify \
    --title "pm:explore complete" \
    --message "Research for '{TOPIC}' written to {OUTPUT_PATH}"
fi

Report: "Research summary written to {OUTPUT_PATH}. Sub-artifacts in .codevoyant/explore/{SLUG}/research/. Use /pm:prd {SLUG} or /pm:plan to continue."