spec-gathering（规格收集）

Requirements Gathering Skill

Overview

Gather user requirements through structured questioning and produce a validated requirements document. This skill transforms vague task descriptions into actionable, structured requirements.

Core principle: Ask smart questions, produce valid structured output. Nothing else.

When to Use

Always:

• Starting a new feature or project
• Clarifying ambiguous task descriptions
• When user provides high-level goals without specifics
• Before spec writing begins

Exceptions:

• Simple bug fixes with clear reproduction steps
• Single-file changes with obvious scope
• User explicitly provides complete requirements

The Iron Law

NO SPEC WRITING WITHOUT VALIDATED REQUIREMENTS FIRST

Requirements must be confirmed by the user before proceeding to spec creation.

Workflow

Phase 1: Load Project Context

Understand the project structure before engaging the user.

Steps:

1. Read project structure files if they exist
2. Identify services, tech stack, and ports
3. Understand existing patterns and conventions

# Read project structure
cat .claude/context/product.md 2>/dev/null || echo "No product context"
cat .claude/context/tech-stack.md 2>/dev/null || echo "No tech stack"

Phase 2: Understand the Task

If a task description was provided, confirm it:

"I understand you want to: [task description]. Is that correct? Any clarifications?"

If no task was provided, ask:

"What would you like to build or fix? Please describe the feature, bug, or change you need."

Wait for user response.

Phase 3: Determine Workflow Type

Based on the task, determine the workflow type:

Ask to confirm:

"This sounds like a [workflow_type] task. Does that seem right?"

Phase 4: Identify Services and Scope

Based on the project context and task, suggest affected areas:

"Based on your task and project structure, I think this involves: - [service1] (primary) - [why] - [service2] (integration) - [why] Any other services or areas involved?"

Wait for confirmation or correction.

Phase 4.5: Invoke Context-Compressor (Progressive Disclosure Mode / ECLAIR Pattern)

Before gathering requirements manually, use the context-compressor skill (progressive disclosure mode) to optimize the clarification process:

Skill({
  skill: 'context-compressor',
  context: {
    taskDescription: taskDescription,
    projectContext: projectContext,
    services: identifiedServices,
    mode: 'progressive-disclosure',
  },
});

Context-Compressor (progressive disclosure mode) performs:

1. E: Examine - Analyze ambiguities in the task description
2. C: Categorize - Prioritize by CRITICAL, HIGH, MEDIUM, LOW
3. L: Limit - Apply 3-5 clarification cap
4. A: Assume - Apply smart defaults from project patterns
5. I: Infer - Use existing code patterns and tech stack
6. R: Record - Document all assumptions with [ASSUMES: X]

Returns:

• Prioritized clarification questions (max 5)
• Smart defaults with [ASSUMES:] notation
• Updated understanding of requirements

Phase 5: Gather Detailed Requirements (with Context-Compressor)

Ask only the critical clarification questions identified by context-compressor:

Budget: 3-5 clarification questions maximum

Process:

1. Ask CRITICAL priority questions first (security, data loss, breaking changes)

- Example: "Does the system need role-based access control?" - Example: "Should users be able to delete data? Is this reversible?"

1. Ask HIGH priority questions if budget remains (UX, architecture, scalability)

- Example: "What's the expected user load?" - Example: "Should this support offline mode?"

1. For MEDIUM/LOW priorities, use [ASSUMES:] notation with defaults

- Example: "[ASSUMES: JWT tokens with 1-hour expiry]" - Example: "[ASSUMES: REST API following existing /api/v1/ pattern]"

Collect answers and document all assumptions.

Phase 6: Confirm and Output (with Assumptions)

Summarize what you understood, including clarified requirements and assumptions:

"Let me confirm I understand: Task: [summary] Type: [workflow_type] Scope: [list of affected areas] Clarified Requirements (questions asked): 1. ✅ [Question] → [Answer] 2. ✅ [Question] → [Answer] 3. ✅ [Question] → [Answer] Smart Defaults Applied (not asked due to budget or clarity): [ASSUMES: X] [ASSUMES: Y] [ASSUMES: Z] Success Criteria: 1. [criterion 1] 2. [criterion 2] Is this correct?"

Wait for confirmation. If user objects to any assumption, allow them to override or request adjustment.

Phase 7: Map Requirements to Template Tokens (with Assumptions)

After confirming requirements and assumptions with user, map gathered data to template tokens:

const tokens = {
  // Required tokens
  FEATURE_NAME: gatheredRequirements.taskName,
  VERSION: '1.0.0',
  AUTHOR: 'Claude',
  DATE: new Date().toISOString().split('T')[0],
  STATUS: 'draft',

  // Required: Acceptance criteria (minimum 1, maximum 50)
  ACCEPTANCE_CRITERIA_1: gatheredRequirements.criteria[0] || '[Define acceptance criterion 1]',
  ACCEPTANCE_CRITERIA_2: gatheredRequirements.criteria[1] || '[Define acceptance criterion 2]',
  ACCEPTANCE_CRITERIA_3: gatheredRequirements.criteria[2] || '[Define acceptance criterion 3]',

  // Optional tokens (can be empty strings if not gathered)
  TERM_1: gatheredRequirements.terms?.[0] || '',
  TERM_2: gatheredRequirements.terms?.[1] || '',
  TERM_3: gatheredRequirements.terms?.[2] || '',

  HTTP_METHOD: gatheredRequirements.httpMethod || '',
  ENDPOINT_PATH: gatheredRequirements.endpointPath || '',
  PROJECT_NAME: gatheredRequirements.projectName || 'Agent Studio',

  // Assumptions from context-compressor (optional but recommended)
  ASSUMPTIONS_MADE: gatheredRequirements.assumptions.map(a => `- ${a}`).join('\n') || '',
  CLARIFICATIONS_ASKED: gatheredRequirements.clarifications || 0,
};

Validation Before Rendering:

• Check all required tokens are populated (FEATURE_NAME, VERSION, AUTHOR, DATE, STATUS)
• Check at least one ACCEPTANCE*CRITERIA** token is meaningful (not placeholder)
• Verify all assumptions are documented with [ASSUMES:] notation
• If missing required data, prompt user for missing information
• If clarifications > 5, warn user about potential requirement incompleteness

Phase 8: Render Specification via Template

Invoke the template-renderer skill to create the specification:

Skill({
  skill: 'template-renderer',
  args: {
    templateName: 'specification-template',
    outputPath: `.claude/context/artifacts/specifications/${featureNameSlug}-spec.md`,
    tokens: tokens,
  },
});

Output Location: .claude/context/artifacts/specifications/[feature-name]-spec.md

Post-Rendering Verification:

# Check file created
SPEC_FILE=".claude/context/artifacts/specifications/[feature-name]-spec.md"
test -f "$SPEC_FILE" && echo "✓ Spec created" || echo "✗ Spec creation failed"

# Check no unresolved tokens
grep -E '[{]{2}' "$SPEC_FILE" && echo "✗ Unresolved tokens found" || echo "✓ All tokens resolved"

# Check YAML frontmatter valid
YAML_COUNT=$(head -50 "$SPEC_FILE" | grep -E "^---$" | wc -l)
test "$YAML_COUNT" -eq 2 && echo "✓ YAML valid" || echo "✗ YAML invalid"

Verification Checklist

Before completing requirements gathering:

• Task description confirmed with user
• Context-compressor skill invoked in progressive disclosure mode (Phase 4.5)
• Clarification budget respected (3-5 questions maximum)
• CRITICAL priority questions asked and answered
• All assumptions documented with [ASSUMES:] notation
• Workflow type determined and confirmed
• Scope and affected areas identified
• Specific requirements captured
• Acceptance criteria defined (minimum 1, mapped to ACCEPTANCE_CRITERIA_1/2/3)
• Constraints documented
• User confirmed final summary (including assumptions)
• User allowed to override any assumptions before proceeding
• Token mapping complete (all required tokens populated, including ASSUMPTIONS_MADE)
• Template renderer invoked successfully
• Specification file created (no unresolved tokens)
• YAML frontmatter valid (2 delimiters, required fields present)
• [ASSUMES:] notation appears in specification output

Common Mistakes

Assuming Instead of Asking

Why it's wrong: Assumptions lead to building the wrong thing.

Do this instead: Ask clarifying questions. Confirm understanding.

Skipping Confirmation

Why it's wrong: User may have misunderstood your summary.

Do this instead: Always summarize and wait for explicit confirmation.

Vague Requirements

Why it's wrong: "Make it better" is not actionable.

Do this instead: Get specific: What behavior? What outcome? How to verify?

Integration with Other Skills

This skill works well with:

• context-compressor: Invoked in Phase 4.5 (progressive disclosure mode) to optimize clarification process with ECLAIR pattern (3-5 limit, smart defaults, [ASSUMES:] notation)
• template-renderer: Used automatically after Phase 8 to render specification template with gathered requirements and assumptions
• spec-critique: Use to validate the generated specification
• complexity-assessment: Assess complexity after requirements are clear
• brainstorming: Use for creative exploration before requirements gathering

Workflow Chain:

spec-gathering (→ context-compressor in Phase 4.5) → template-renderer → spec-critique → planner

Context-Compressor Integration Details (Progressive Disclosure Mode):

• When: Invoked at the start of Phase 5 (before manual questioning)
• Why: Limits clarifications to 3-5 questions (reduces cognitive load), applies smart defaults, documents all assumptions
• Output: Prioritized questions + default assumptions with [ASSUMES:] notation
• Impact: Specification includes documented assumptions that can be overridden during implementation

Examples

Example 1: Feature Request (End-to-End with Progressive-Disclosure & Template Rendering)

Input: "Add user authentication to the app"

Process:

1. Confirm task: "You want to add user authentication. Is this correct?"

- User confirms

1. Invoke context-compressor in progressive disclosure mode (Phase 4.5): Skill({skill: 'context-compressor', context: {taskDescription: "Add user authentication", projectContext: {...}, mode: 'progressive-disclosure'}});
2. Context-compressor returns:

- CRITICAL questions (3 asked of 5 budget): 1. "Should users be able to reset passwords via email?" → YES 2. "Do we need role-based access control (RBAC)?" → YES (Admin, User, Guest) 3. "Is single sign-on (SSO) required?" → NO - Assumptions applied (with [ASSUMES:]): - [ASSUMES: JWT tokens with 1-hour expiry] - [ASSUMES: bcrypt for password hashing, cost factor 12] - [ASSUMES: HTTPS required for all endpoints]

1. Identify scope: "This will touch the backend API, database, and frontend login page."

- User confirms scope

1. Define success criteria based on clarifications and defaults
2. Map to tokens:

{
  FEATURE_NAME: "User Authentication",
  VERSION: "1.0.0",
  AUTHOR: "Claude",
  DATE: "2026-01-28",
  STATUS: "draft",
  ACCEPTANCE_CRITERIA_1: "Users can sign up with email and password",
  ACCEPTANCE_CRITERIA_2: "Email verification required before login",
  ACCEPTANCE_CRITERIA_3: "Users can reset password via email link",
  ACCEPTANCE_CRITERIA_4: "Role-based access control working (Admin/User/Guest)",
  ASSUMPTIONS_MADE: "- JWT tokens with 1-hour expiry\n- bcrypt cost factor 12\n- HTTPS required",
  CLARIFICATIONS_ASKED: 3
}

1. Invoke template-renderer:

Skill({
  skill: 'template-renderer',
  args: {
    templateName: 'specification-template',
    outputPath: '.claude/context/artifacts/specifications/user-authentication-spec.md',
    tokens: tokens,
  },
});

Output:

• Rendered specification at .claude/context/artifacts/specifications/user-authentication-spec.md
• All tokens resolved
• Assumptions clearly documented with [ASSUMES:] notation
• YAML frontmatter valid
• Ready for spec-critique review (user can challenge any assumption)

Example 2: Bug Investigation

Input: "The page loads slowly"

Process:

1. Clarify: "Which page specifically? What do you consider slow?"
2. Context: "When did this start? Any recent changes?"
3. Metrics: "What load time would be acceptable?"

Output: Investigation requirements with specific pages, metrics, and success criteria.

Troubleshooting

Issue: User gives one-word answers

Symptoms: "Yes", "No", "That's fine" without detail.

Solution: Ask more specific questions. Provide options: "Would you prefer A, B, or C?"

Issue: Scope keeps expanding

Symptoms: Every answer adds new features.

Solution: Document "Out of Scope" explicitly. Confirm: "Should we include this in this task or save for later?"

Iron Laws

1. NEVER begin implementation before a validated specification exists for STANDARD+ complexity
2. ALWAYS capture non-functional requirements (performance, security, scalability) explicitly
3. NEVER accept vague success criteria — require measurable, testable numbers
4. ALWAYS document what is explicitly out of scope, not just what is in scope
5. NEVER hand off to planner without explicit user confirmation of requirements

Anti-Patterns

Memory Protocol

Before starting: Read .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: If it's not in memory, it didn't happen.