Gather Requirements Skill

Gather Requirements Skill

Purpose: Invoke all stakeholder agents (architect, tester, formatter) in parallel for requirements gathering, enforcing proper coordination and verification.

Performance: Reduces requirements phase from hours (sequential) to minutes (parallel)

When to Use This Skill

✅ Use gather-requirements When:

• Task is in CLASSIFIED state
• Ready to gather stakeholder requirements
• Need to transition to REQUIREMENTS phase
• Want to ensure proper parallel coordination

❌ Do NOT Use When:

• Task not yet initialized (use task-init first)
• Still in INIT state (transition to CLASSIFIED first)
• Requirements already gathered (check for requirement reports)
• Ad-hoc work outside task protocol

What This Skill Does

1. Validates Preconditions

# Checks before invocation:
- Task exists
- Task in CLASSIFIED state
- Agent worktrees exist
- No existing requirement reports (prevents duplicates)

2. Invokes Stakeholder Agents in Parallel

CRITICAL: All three agents MUST be invoked in a single message with multiple Task tool calls.

Single assistant message containing:
- Task tool call #1: architect agent
- Task tool call #2: tester agent
- Task tool call #3: formatter agent

This ensures true parallel coordination (all agents start concurrently).

3. Monitors Completion

# Waits for all agents to complete
# Verifies all requirement reports created:
- {task-name}-architect-requirements.md
- {task-name}-tester-requirements.md
- {task-name}-formatter-requirements.md

4. Validates Reports

# Checks each report:
- File exists
- Non-empty content
- Contains required sections
- Proper markdown format

Required Agent Prompts

Architect Agent Prompt Template

You are the architect stakeholder agent for task: {task-name}

**PHASE**: REQUIREMENTS (NOT implementation)

**YOUR ROLE**: Analyze and document requirements only. DO NOT write code.

**TASK DESCRIPTION**:
{task-description-from-todo.md}

**WORKING DIRECTORY**:
/​workspace/​tasks/​{task-name}/​agents/​architect/​code

**MANDATORY OUTPUT REQUIREMENT**:
Write your analysis to: `/​workspace/​tasks/​{task-name}/​{task-name}-architect-requirements.md`

**ANALYSIS SCOPE**:
1. Dependencies and integration points
2. Design patterns applicable
3. Architecture considerations
4. API design recommendations
5. Module structure recommendations

**CRITICAL**: You are in REQUIREMENTS mode. Write ONLY the requirements report.
DO NOT implement code. DO NOT create source files.

Report file MUST be created at the specified path.

Tester Agent Prompt Template

You are the tester stakeholder agent for task: {task-name}

**PHASE**: REQUIREMENTS (NOT implementation)

**YOUR ROLE**: Analyze testing requirements only. DO NOT write tests yet.

**TASK DESCRIPTION**:
{task-description-from-todo.md}

**WORKING DIRECTORY**:
/​workspace/​tasks/​{task-name}/​agents/​tester/​code

**MANDATORY OUTPUT REQUIREMENT**:
Write your analysis to: `/​workspace/​tasks/​{task-name}/​{task-name}-tester-requirements.md`

**ANALYSIS SCOPE**:
1. Test coverage needs
2. Test strategy (unit, integration, edge cases)
3. Business logic validation requirements
4. Edge cases to cover
5. Test data requirements

**CRITICAL**: You are in REQUIREMENTS mode. Write ONLY the requirements report.
DO NOT implement tests. DO NOT create test files.

Report file MUST be created at the specified path.

Formatter Agent Prompt Template

You are the formatter stakeholder agent for task: {task-name}

**PHASE**: REQUIREMENTS (NOT implementation)

**YOUR ROLE**: Analyze documentation and style requirements only. DO NOT implement yet.

**TASK DESCRIPTION**:
{task-description-from-todo.md}

**WORKING DIRECTORY**:
/​workspace/​tasks/​{task-name}/​agents/​formatter/​code

**MANDATORY OUTPUT REQUIREMENT**:
Write your analysis to: `/​workspace/​tasks/​{task-name}/​{task-name}-formatter-requirements.md`

**ANALYSIS SCOPE**:
1. Documentation requirements (JavaDoc, comments)
2. Code style standards to enforce
3. Naming convention requirements
4. File organization requirements
5. Checkstyle/​PMD considerations

**CRITICAL**: You are in REQUIREMENTS mode. Write ONLY the requirements report.
DO NOT implement code. DO NOT create source files.

Report file MUST be created at the specified path.

Usage

Basic Requirements Gathering

# Step 1: Ensure task in CLASSIFIED state
cd /​workspace/​main
TASK_NAME="implement-formatter-api"

jq '.state = "CLASSIFIED" | .phase = "requirements"' \
  /​workspace/​tasks/​$TASK_NAME/​task.json > tmp.json
mv tmp.json /​workspace/​tasks/​$TASK_NAME/​task.json

# Step 2: Invoke skill
# Skill will use Task tool to invoke all 3 agents in parallel

With Task Description

# Extract task description from todo.md
TASK_NAME="implement-formatter-api"
DESCRIPTION=$(grep -A 2 "^- \[ \] $TASK_NAME" /​workspace/​main/​todo.md | tail -n 1)

# Pass to skill for agent prompts
# Skill will incorporate description into agent prompts

Parallel Invocation Pattern

✅ CORRECT Pattern (True Parallel)

Assistant: Invoking all stakeholder agents for requirements gathering...

[Single message with 3 Task tool calls:]

Task tool call #1:
- subagent_type: architect
- prompt: [architect prompt with REQUIREMENTS emphasis]
- model: opus

Task tool call #2:
- subagent_type: tester
- prompt: [tester prompt with REQUIREMENTS emphasis]
- model: opus

Task tool call #3:
- subagent_type: formatter
- prompt: [formatter prompt with REQUIREMENTS emphasis]
- model: opus

Result: All agents start work concurrently, requirements gathered in parallel.

❌ WRONG Pattern (Sequential) - DO NOT DO THIS

The following shows what NOT to do - invoking agents one at a time in separate messages:

# ❌ ANTI-PATTERN - Sequential invocation wastes time

Message 1:
Assistant: Invoking architect agent...
[Task tool call: architect]

[Wait for architect to complete - could take minutes]

Message 2:
Assistant: Now invoking tester agent...
[Task tool call: tester]

[Wait for tester to complete - more minutes wasted]

Message 3:
Assistant: Finally invoking formatter agent...
[Task tool call: formatter]

[Wait for formatter to complete]

Problem: Each agent runs sequentially, tripling the total time. If each agent takes 2 minutes, sequential = 6 minutes vs parallel = 2 minutes.

---

Verification After Completion

After all agents complete, verify reports exist:

TASK_NAME="your-task-name"
ls -la /​workspace/​tasks/​$TASK_NAME/​*-requirements.md

# Expected output:
# {task-name}-architect-requirements.md
# {task-name}-tester-requirements.md
# {task-name}-formatter-requirements.md

Related Skills

• task-init: Initialize task before gathering requirements
• select-agents: Determine which agents are needed
• synthesize-plan: Next step after requirements gathered
• verify-requirements-complete: Validate all reports exist