prp-research-team

PRP Research Team Planner

Input: $ARGUMENTS

---

Mission

Design a dynamic team of research agents and a structured research plan for any given question or topic. The plan targets Claude Code's experimental agent teams feature (TeamCreate, shared task list, delegate mode).

Core Principle: PLAN ONLY — no research is executed. Produce a comprehensive, executable research plan that enables a team of agents to deliver thorough findings.

Golden Rule: Every researcher must have a clear focus, measurable completion criteria, and a defined output format. No vague mandates.

Doctrine: The research question dictates the team — never force a fixed roster. A market research question demands different expertise than a codebase architecture question.

---

Variables

---

Phase 1: PARSE — Extract Research Question

1.1 Parse Arguments

Extract from $ARGUMENTS:

• Research question or topic: Everything that is NOT a flag
• Orchestration guidance: Value after --orchestration flag (if present)

Parsing rules:

1. Strip --orchestration "..." or --orchestration '...' from arguments → store as ORCHESTRATION
2. Remaining text = research question
3. If question is empty after parsing → STOP with error

1.2 Identify Scope Signals

Scan the research question for scope indicators:

1.3 Validate

If question is empty or unclear:

Research question required.

Usage:
  /prp-research-team "What are the best approaches for real-time collaboration?"
  /prp-research-team "Compare state management libraries for React" --orchestration "Focus on bundle size and DX"

PHASE_1_CHECKPOINT:

• Research question extracted and non-empty
• --orchestration flag parsed (or confirmed absent)
• Scope signals identified

GATE: If the research question is too vague to decompose into sub-questions → STOP and ASK user for clarification.

---

Phase 2: CLASSIFY — Domain & Complexity

2.1 Determine Research Domain

Classify the question into one or more domains:

For MIXED domains, identify the primary domain and supporting domains.

2.2 Assess Complexity

Complexity factors:

• Number of domains involved
• Breadth of the question
• Depth of analysis required
• Number of comparative dimensions
• Whether primary research vs. synthesis

2.3 Apply Orchestration Override

If ORCHESTRATION is set, adjust:

• Team composition emphasis
• Domain weighting
• Specific expertise requirements
• Any constraints on approach

PHASE_2_CHECKPOINT:

• Primary domain identified with rationale
• Supporting domains listed (if MIXED)
• Complexity assessed with team size determined
• Orchestration guidance applied (if provided)

---

Phase 3: DECOMPOSE — Sub-Questions

3.1 Break Down Research Question

Decompose into 3-7 independently investigable sub-questions.

Decomposition rules:

1. Each sub-question must be answerable by a single researcher
2. Sub-questions should cover the full scope of the original question
3. Identify which sub-questions can run in PARALLEL vs. which have DEPENDENCIES
4. Tag each sub-question with its primary domain

3.2 Map Dependencies

Create a dependency graph:

SQ-1 (foundational) ──┬──► SQ-2 (parallel)
                       ├──► SQ-3 (parallel)
                       └──► SQ-4 (parallel)
                                    │
                                    ▼
                              SQ-5 (synthesis, depends on SQ-2,3,4)

Dependency types:

• NONE: Can start immediately
• BLOCKED_BY: Must wait for specific sub-questions
• INFORMS: Benefits from but doesn't require other results

3.3 Validate Coverage

Check that sub-questions collectively:

• Cover the full scope of the original question
• Don't have significant overlap (some overlap at boundaries is acceptable)
• Include at least one synthesis/integration sub-question

PHASE_3_CHECKPOINT:

• 3-7 sub-questions defined
• Each sub-question is independently investigable
• Dependencies mapped (parallel vs. sequential)
• Full coverage of original question verified
• At least one synthesis sub-question included

---

Phase 4: COMPOSE — Design Team Roles

4.1 Design Researcher Profiles

For each researcher, define:

4.2 Spawn Prompt Requirements

Each spawn prompt MUST include:

1. Role statement: Who you are and what you're investigating
2. Research question(s): The specific sub-questions assigned
3. Methodology: How to approach the research (web search, code analysis, doc review, etc.)
4. Output format: Exact markdown structure for findings
5. Quality bar: What constitutes sufficient depth
6. Completion signal: How to indicate research is complete (update shared task)

4.3 Model Selection

4.4 Apply Orchestration to Team

If ORCHESTRATION is set, verify the team composition aligns with the guidance. Adjust roles, emphasis, or add/remove researchers as needed.

PHASE_4_CHECKPOINT:

• Each researcher has all 7 fields defined
• Spawn prompts are self-contained (no external context needed)
• Output formats are specific and structured
• Completion criteria are measurable
• Model selection is justified
• Team covers all sub-questions with no gaps

---

Phase 5: PLAN — Research Tasks

5.1 Create Task List

For each task, define:

5.2 Task Ordering

1. Wave 1: All tasks with no dependencies (parallel)
2. Wave 2: Tasks that depend on Wave 1 outputs
3. Wave 3: Synthesis and integration tasks
4. Final: Review and quality assurance

5.3 Define Cross-Cutting Concerns

Identify shared standards across all researchers:

• Citation format and requirements
• Confidence level tagging (HIGH / MEDIUM / LOW with rationale)
• Contradiction handling (when sources disagree)
• Scope boundary enforcement (when to stop digging)

PHASE_5_CHECKPOINT:

• Every sub-question has at least one task
• Dependencies form a valid DAG (no cycles)
• Parallel tasks identified for maximum throughput
• Synthesis task exists to integrate findings
• Cross-cutting concerns defined

---

Phase 6: GENERATE — Write Research Plan

6.1 Create Output Directory

mkdir -p .claude/PRPs/research-plans

6.2 Determine Output Filename

Convert the research topic to kebab-case, truncate to 50 chars max:

• "What are the best approaches for real-time collaboration?" → real-time-collaboration
• "Compare React vs Vue vs Svelte for enterprise apps" → react-vs-vue-vs-svelte-enterprise

Output path: .claude/PRPs/research-plans/{topic-slug}.research-plan.md

6.3 Write State Sentinel

Write the output path to .claude/prp-research-team.state so the Stop hook can validate:

.claude/PRPs/research-plans/{topic-slug}.research-plan.md

Just the file path, one line, no extra content.

6.4 Write Research Plan

Write the research plan to the output path using this exact template:

# Research Plan: {Research Question}

## Metadata

| Field | Value |
|-------|-------|
| Date | {YYYY-MM-DD} |
| Topic | {short topic name} |
| Domain | {PRIMARY / MIXED: list} |
| Complexity | {LOW / MEDIUM / HIGH} |
| Team Size | {N} researchers |
| Sub-questions | {N} |
| Tasks | {N} |

---

## Research Question

{The original research question, clearly stated and unambiguous.}

{If orchestration guidance was provided:}
**Orchestration**: {The orchestration guidance}

---

## Research Question Decomposition

| ID | Sub-question | Domain | Parallel | Dependencies | Assigned To |
|----|-------------|--------|----------|--------------|-------------|
| SQ-1 | {sub-question text} | {domain} | {yes/no} | {NONE or SQ-IDs} | {researcher name} |
| SQ-2 | ... | ... | ... | ... | ... |

### Dependency Graph

{ASCII dependency diagram showing parallel vs. sequential flow}

---

## Team Composition

### {Researcher 1 Name}

- **Focus**: {1-2 sentence description}
- **Sub-questions**: {SQ-IDs}
- **Model**: {sonnet / opus}
- **Output format**: {description of deliverable structure}
- **Completion criteria**: {measurable conditions}

**Spawn prompt**:
> {Complete, self-contained instructions for this agent. Must include:
> role statement, assigned sub-questions, methodology, output format,
> quality bar, and completion signal. The agent must be able to execute
> with ONLY this prompt — no external context.}

### {Researcher 2 Name}

{Same structure as above}

{Repeat for all researchers...}

---

## Research Tasks

### Wave 1: Foundation (Parallel)

| ID | Title | Assignee | Type | Dependencies | Acceptance Criteria | Effort |
|----|-------|----------|------|-------------|-------------------|--------|
| RT-1 | {title} | {name} | RESEARCH | NONE | {criteria} | {LOW/MED/HIGH} |

### Wave 2: Deep Analysis

| ID | Title | Assignee | Type | Dependencies | Acceptance Criteria | Effort |
|----|-------|----------|------|-------------|-------------------|--------|
| RT-N | {title} | {name} | ANALYSIS | RT-1, RT-2 | {criteria} | {LOW/MED/HIGH} |

### Wave 3: Synthesis

| ID | Title | Assignee | Type | Dependencies | Acceptance Criteria | Effort |
|----|-------|----------|------|-------------|-------------------|--------|
| RT-N | {title} | {name} | SYNTHESIS | RT-... | {criteria} | {LOW/MED/HIGH} |

### Cross-Cutting Concerns

- **Citations**: {format requirements}
- **Confidence levels**: Tag all findings as HIGH / MEDIUM / LOW with rationale
- **Contradictions**: When sources disagree, document both positions with evidence
- **Scope boundaries**: {when to stop investigating a thread}

---

## Team Orchestration Guide

### Prerequisites

This research plan is designed for execution using Claude Code's experimental **agent teams** feature. Before executing:

1. Ensure agent teams is enabled (experimental feature)
2. Review the team composition and adjust if needed
3. Confirm the research question and scope

### Execution Steps

1. **Create team**: Use `TeamCreate` to spawn all researchers defined in Team Composition
2. **Create shared tasks**: Use the shared task list to create all tasks from the Research Tasks section
3. **Set dependencies**: Link tasks with their dependencies so agents pick up work in the correct order
4. **Monitor progress**: Use delegate mode or direct messaging to check on researcher progress
5. **Collect outputs**: Each researcher posts findings to their assigned tasks
6. **Run synthesis**: The synthesis researcher integrates all findings into the final report

### Display Mode

Use **delegate mode** for autonomous execution:
- Researchers work independently on their assigned tasks
- The lead researcher monitors progress and resolves blockers
- Use `SendMessage` to communicate between researchers when dependencies complete

### Communication Patterns

- **Handoff**: When a Wave 1 researcher completes, notify dependent Wave 2 researchers via task updates
- **Clarification**: Researchers can message the lead for scope questions
- **Contradiction**: If two researchers find conflicting information, escalate to lead for resolution

### Plan Approval

Before execution, review:
- [ ] Team composition matches the research domain
- [ ] Spawn prompts are detailed enough for autonomous execution
- [ ] Task dependencies are correct
- [ ] Acceptance criteria are measurable

---

## Acceptance Criteria

Research is complete when ALL of the following are met:

- [ ] Every sub-question (SQ-*) has been investigated and answered
- [ ] Every research task (RT-*) has been completed and meets its acceptance criteria
- [ ] Findings are cited with sources and confidence levels
- [ ] Contradictions are documented with both positions
- [ ] A synthesis document integrates all findings into a coherent answer
- [ ] The original research question is directly answered with evidence

---

## Output Format: Final Report Structure

The final research report (produced during execution, not in this plan) should follow:

1. **Executive Summary** — Direct answer to the research question (2-3 paragraphs)
2. **Key Findings** — Bulleted list of major discoveries
3. **Detailed Analysis** — Section per sub-question with evidence
4. **Comparative Matrix** — If applicable, structured comparison table
5. **Recommendations** — Actionable next steps with confidence levels
6. **Sources** — All references with URLs and access dates
7. **Appendix** — Raw data, extended quotes, additional context

PHASE_6_CHECKPOINT:

• Output directory exists
• State sentinel file written with output path
• Research plan file written with ALL required sections
• All researcher spawn prompts are self-contained
• All tasks have acceptance criteria
• Template sections are filled (no placeholders remain)

GATE: Do NOT proceed to Phase 7 until the research plan file passes validation — all 6 required sections must be present:

1. ## Research Question
2. ## Research Question Decomposition
3. ## Team Composition
4. ## Research Tasks
5. ## Team Orchestration Guide
6. ## Acceptance Criteria

---

Phase 7: OUTPUT — Report to User

Display a summary to the user:

## Research Plan Created

**File**: `{output path}`
**Question**: {research question}

### Team Composition ({N} researchers)

| Researcher | Focus | Model |
|------------|-------|-------|
| {name} | {1-line focus} | {model} |

### Plan Overview

- **Domain**: {domain classification}
- **Complexity**: {LOW/MEDIUM/HIGH}
- **Sub-questions**: {N}
- **Tasks**: {N} ({W1} parallel → {W2} analysis → {W3} synthesis)

### Execution

To execute this research plan with agent teams:
1. Review the plan: `read {output path}`
2. Create the team and start execution using the orchestration guide in the plan

### Manual Execution Alternative

If agent teams is not available, execute sequentially:
1. Work through Wave 1 tasks in parallel using Task tool subagents
2. Feed Wave 1 outputs into Wave 2 tasks
3. Synthesize in Wave 3

PHASE_7_CHECKPOINT:

• Summary displayed to user
• Team composition table shown
• Execution instructions provided
• Output file path clearly communicated

---

Success Criteria

• QUESTION_PARSED: Research question extracted and validated
• DOMAIN_CLASSIFIED: Primary and supporting domains identified
• DECOMPOSED: 3-7 independent sub-questions with dependency mapping
• TEAM_DESIGNED: Each researcher has name, focus, spawn prompt, output format, completion criteria
• TASKS_PLANNED: All tasks have IDs, assignees, dependencies, acceptance criteria
• PLAN_WRITTEN: Research plan file created with all required sections
• SENTINEL_SET: State file written for stop hook validation
• USER_INFORMED: Summary with execution instructions displayed