Agent Skills: Spec-Driven Planning Skill

Spec-Driven Planning Skill

Purpose

Guide feature planning in one of two modes:

• Quick mode — single docx/features/[NN-name]/plan.md with bite-sized tasks. No EARS, no RGR. For solo work, ≤3 days, no compliance/handoff. Derived from the superpowers writing-plans pattern.
• Full mode — three files (requirements.md EARS + design.md + tasks.md) with three approval gates and TDD enforcement downstream. For team work, multi-week, compliance/audit, or stakeholder review.

The skill picks a mode (or asks when ambiguous), then runs the matching playbook.

Activation Triggers

Activate this skill when:

• User says "create a new feature", "plan a feature", "I need to build [X]"
• User mentions "requirements", "specifications", "specs", "architecture", "technical design"
• User uses /dev-workflow:spec command (any sub-arg)

Mode Selection (ALWAYS DO THIS FIRST)

Before writing anything, decide Quick vs Full.

Auto-pick Quick when ALL true:

• Estimated effort ≤ 3 days OR ≤ 8 tasks
• Solo developer; no stakeholder approval needed
• No compliance, audit, or regulatory requirement
• No major architectural decision (no "X vs Y?" trade-off)
• Requirements clear from the user's request
• No cross-team handoff
• User said "quick", "lightweight", "simple plan", or /dev-workflow:spec quick

Auto-pick Full when ANY true:

• Effort > 1 week OR > 15 tasks
• Stakeholders ≠ engineers must approve scope
• Compliance / audit / regulatory traceability required
• Architecture choice with long-term consequences
• Requirements ambiguous; need elicitation
• Multi-developer or handoff likely
• User said "full spec", "EARS", "requirements doc", or /dev-workflow:spec full

Otherwise, treat as ambiguous

The thresholds above are deliberately gapped — 4-7 days of effort, 9-15 tasks, or any partial signal match falls in the middle zone. If you don't get a clean all-true Quick or any-true Full match, the feature is ambiguous by definition. Ask the user.

Ambiguous? Ask once with this menu:

This feature could go either way. Pick a planning mode:

1. Quick — single plan.md, no EARS, no RGR (recommended for solo, ≤3 days)
2. Full — 3-file spec with EARS + TDD enforcement (recommended for team, >1 week, compliance)

Default if you don't pick: Quick.

Announce the chosen mode:

"Using [Quick/Full] mode. [One-sentence reason from signals above.]"

Then run the matching playbook below.

Quick Mode Playbook

Output: docx/features/[NN-feature-name]/plan.md (single file)

Step 1: Create feature directory

1. ls docx/features/ to find next NN number
2. mkdir -p docx/features/[NN-feature-name]
3. Read dev-workflow/templates/plan.md
4. Write to docx/features/[NN-feature-name]/plan.md, replacing [Feature Name] with the actual name

Step 2: Fill in the plan

Audience assumption: an engineer with zero context for this codebase. Be exact.

Required sections:

• Goal — one sentence
• Architecture — 2-3 sentences, what goes where and why
• Tech Stack — key libraries / files
• Out of Scope — keeps scope honest
• Tasks — bite-sized (2-5 min steps), with:
  • Files: exact paths (Create / Modify with line ranges)
  • Steps: numbered, each one action. Include complete code, not "add validation here". Include exact verification commands with expected output. End with a commit step.

Granularity rule: if a step takes longer than 5 minutes, split it.

Step 3: One approval gate

"Quick plan saved to docx/features/[NN-feature-name]/plan.md. Review and approve to proceed to implementation. Run /dev-workflow:spec execute when ready."

That's it. No EARS, no design alternatives, no RGR. Implementation is handled by spec-driven-implementation (which auto-detects plan.md and runs in Quick execution mode).

Quick mode is wrong if you find yourself...

• Writing requirements that need approval from non-engineers → switch to Full
• Comparing 2+ architectural approaches → switch to Full
• Listing > 15 tasks → switch to Full
• Needing traceability from requirement → task → test → switch to Full

To upgrade, see "Upgrade path" at the bottom of this skill.

Full Mode Playbook

Three phases, three approval gates. Use this when signals say Full.

Phase 1: Feature Creation

Goal: Establish feature structure and placeholder files

Process:

1. Parse feature name from user input
2. Check existing features using Bash tool: ls docx/features/
3. Determine next number (NN) for feature directory
4. Create directory using Bash tool: mkdir -p docx/features/[NN-feature-name]
5. Copy templates from plugin to feature directory:
   • Use Read tool: dev-workflow/templates/requirements.md
   • Use Write tool: docx/features/[NN-feature-name]/requirements.md (replace [Feature Name] with actual name)
   • Use Read tool: dev-workflow/templates/design.md
   • Use Write tool: docx/features/[NN-feature-name]/design.md (replace [Feature Name] with actual name)
   • Use Read tool: dev-workflow/templates/tasks.md
   • Use Write tool: docx/features/[NN-feature-name]/tasks.md (replace [Feature Name] with actual name)

Output:

Created feature: docx/features/[NN-feature-name]/
- requirements.md (from template)
- design.md (from template)
- tasks.md (from template)

Next step: Define requirements using EARS format

User Confirmation:

"Feature structure created. Ready to define requirements?"

Phase 2: Requirements Definition (EARS Format)

Goal: Capture clear, testable requirements using EARS methodology

Scope Decomposition Check (do FIRST):

Before eliciting requirements, scan the request. If it describes multiple independent subsystems (e.g., "auth + billing + admin dashboard"), STOP and decompose into separate features before refining any one. Don't burn elicitation questions on a feature that should be three.

🗣 Say: "This request covers [N] independent subsystems. I'll create separate features for each before eliciting requirements."

No Placeholders Rule:

Requirements and downstream plans MUST NOT contain:

• "TBD" / "to be determined" / "to be defined later"
• "add appropriate X" / "handle errors as needed" / "validate as required"
• "similar to [other thing]" / "see above" without specifics
• vague verbs without objects: "update", "improve", "fix", "enhance"
• "etc." or trailing ellipses in requirement text
• references to undefined names (REQ-IDs that don't exist, modules not yet specified) If you cannot specify a requirement concretely, ask the user. Don't write a placeholder.

Brainstorming Integration (Optional):

• If user has rough idea but unclear requirements, use Skill tool to invoke: dev-workflow:brainstorming
• Helps clarify what to build vs. what's out of scope
• Explores different feature scopes through collaborative questioning
• Determines must-haves vs. nice-to-haves

How to activate:

Use Skill tool: Skill(skill: "dev-workflow:brainstorming")

EARS Format Explained:

EARS (Easy Approach to Requirements Syntax) provides five templates for unambiguous requirements:

1. Ubiquitous Requirements - Always true

   • Template: "The system SHALL [requirement]"
   • Example: "The system SHALL validate all user inputs before processing"

2. Event-Driven Requirements - Triggered by events

   • Template: "WHEN [trigger] THEN the system SHALL [response]"
   • Example: "WHEN user clicks submit THEN the system SHALL validate form data"

3. State-Driven Requirements - Active during specific states

   • Template: "WHILE [state] the system SHALL [requirement]"
   • Example: "WHILE processing payment the system SHALL display loading indicator"

4. Conditional Requirements - Based on conditions

   • Template: "IF [condition] THEN the system SHALL [requirement]"
   • Example: "IF user role is admin THEN the system SHALL show management panel"

5. Optional Requirements - Feature toggles

   • Template: "WHERE [feature included] the system SHALL [requirement]"
   • Example: "WHERE premium subscription is active the system SHALL enable advanced analytics"

Research Protocol (Before Eliciting Requirements):

Before diving into requirement questions, gather context through research:

1. Prior Art Research

   • Use WebSearch to find similar features/products
   • Query: "[feature type] best practices 2025"
   • Query: "[feature type] common requirements"

2. Technical Documentation

   • Use WebFetch on relevant technical docs, APIs, or standards
   • Fetch competitor/similar product documentation

3. API Research (if applicable)

   • Use Bash with curl to explore API endpoints
   • Fetch API documentation and schemas

4. Document Findings

   • Summarize key insights in requirements.md under "## Research Summary"
   • Note patterns, anti-patterns, and industry standards discovered

🗣 Say: "Let me research similar implementations before we define requirements."

Systematic Questioning Approach:

Ask the user these questions to elicit requirements:

1. Core Functionality

   • "What is the primary purpose of this feature?"
   • "What problem does it solve?"

2. Event-Driven Requirements

   • "What user actions trigger this feature?"
   • "What system events are involved?"

3. State-Driven Requirements

   • "Are there different states or modes?"
   • "What should happen during each state?"

4. Conditional Requirements

   • "Are there different behaviors for different users/roles?"
   • "What conditions affect functionality?"

5. Performance Requirements

   • "Are there response time requirements?"
   • "What's the expected load/scale?"

6. Security Requirements

   • "What data needs protection?"
   • "Who should have access?"

7. Error Handling

   • "What can go wrong?"
   • "How should errors be handled?"

8. Edge Cases

   • "What are the boundary conditions?"
   • "What happens at extremes?"

Best Practices:

• Use "SHALL" for mandatory requirements
• Be specific and measurable (avoid "quickly", use "within 2 seconds")
• One requirement per statement
• Avoid ambiguous terms ("appropriate", "reasonable", "user-friendly")
• Use active voice

Requirement IDs & Traceability:

• Assign unique IDs to every requirement using a consistent prefix (e.g., REQ-001).
• Keep numbering sequential across all requirement types (functional + non-functional).
• Record the IDs directly in each requirement line so later tasks can reference them.
• Add a short traceability note indicating how tasks/design will map back to these IDs.

Output Format: Update docx/features/[NN-feature-name]/requirements.md with:

• Overview section
• Functional requirements (organized by EARS type)
• Non-functional requirements (performance, security, usability)
• Constraints
• Acceptance criteria (checkboxes)
• Out of scope items

User Confirmation:

"Requirements complete. Ready for design phase?"

Phase 3: Technical Design

Goal: Create comprehensive technical design with architectural decisions

Research Protocol (Before Design):

Before proposing architectural approaches, research solutions:

1. Architecture Research

   • Use WebSearch: "[technology] architecture patterns 2025"
   • Use WebSearch: "[problem domain] implementation approaches"

2. Library/Framework Research

   • Use WebFetch on documentation for potential libraries
   • Compare approaches used by similar projects

3. API Research (if applicable)

   • Use WebFetch on external API documentation
   • Use Bash with curl to test API endpoints
   • Understand integration requirements and constraints

4. Document Findings

   • Add "## Technical Research" section to design.md
   • Include links to sources and key insights

🗣 Say: "Let me research technical approaches before proposing architecture options."

Process:

1. Brainstorming Integration

   • Use Skill tool to invoke: dev-workflow:brainstorming for collaborative design exploration
   • Explore 2-3 different architectural approaches
   • Discuss trade-offs for each approach

   How to activate:

   Use Skill tool: Skill(skill: "dev-workflow:brainstorming")
   

UltraThink for Complex Designs: Before proposing technical approaches, activate deep thinking when:

• Architecture involves multiple services or complex data flows
• Trade-offs between approaches aren't obvious
• Design impacts security, performance, or scalability
• Requirements seem contradictory or incomplete

🗣 Say: "This design requires deep thinking. Let me ultrathink the architectural fundamentals before proposing approaches."

During UltraThink, question:

• Are we solving the right problem?
• What are we assuming that might be wrong?
• What could break at scale?
• What's the simplest architecture that works?
• What are the hidden costs of each approach?
• What would we do differently if starting from scratch?

After UltraThink: Present approaches with explicit reasoning about architectural trade-offs and scalability considerations.

2. Approach Comparison Present options with trade-offs:

   Option A: [Approach Name]

   • Pros: [Advantages]
   • Cons: [Disadvantages]
   • Complexity: Low/Medium/High
   • Timeline: [Estimate]

   Option B: [Approach Name]

   • Pros: [Advantages]
   • Cons: [Disadvantages]
   • Complexity: Low/Medium/High
   • Timeline: [Estimate]

3. Recommendation

   • State recommended approach
   • Provide clear reasoning
   • Explain why it best fits requirements

4. Design Document Structure Create comprehensive design.md covering:

   Architecture Overview

   • How feature fits into system
   • High-level component diagram (ASCII art)

   Component Structure

   • List components with responsibilities
   • Define dependencies between components
   • Specify public interfaces

   Data Flow

   • Step-by-step data movement
   • Diagram showing flow

   API Contracts

   • Input/output schemas
   • Error responses
   • Example requests/responses

   Error Handling

   • Error scenarios and handling strategy
   • Fallback behaviors

   Security Considerations

   • Authentication/authorization
   • Input validation
   • Data protection

   Performance Considerations

   • Optimization strategies
   • Caching approach
   • Database indexing needs

   Testing Strategy

   • Unit test areas
   • Integration test scenarios
   • E2E test workflows

Approval Gate:

"Design complete. Ready for task breakdown?"

Wait for explicit user approval before proceeding.

Next Steps

After Quick mode (plan.md saved):

Run /dev-workflow:spec execute. The implementation skill auto-detects plan.md and runs Quick execution: follow the steps, run verification commands, commit. No RGR enforcement.

After Full mode (design.md approved):

Run /dev-workflow:spec tasks then /dev-workflow:spec execute. The implementation skill detects tasks.md and runs Full execution: RED-GREEN-REFACTOR per task, traceability to requirement IDs, quality gates.

Both modes integrate with dev-workflow:review, dev-workflow:git-workflow, and dev-workflow:test-driven-development (the last one is mandatory in Full mode, optional in Quick).

Upgrade Path: Quick → Full

If a Quick feature outgrows its plan (scope expands, stakeholders enter, architecture decisions emerge):

1. Read existing docx/features/[NN-name]/plan.md
2. Seed the three Full-mode files from it:
   • requirements.md — extract Goal + implicit requirements from tasks; rewrite in EARS; assign REQ-### IDs
   • design.md — extract Architecture + Tech Stack; add "Alternatives Considered" by interviewing the user
   • tasks.md — re-break tasks into RED/GREEN/REFACTOR with linked requirement IDs
3. Rename plan.md → plan.archived.md for traceability
4. Resume from Phase 2 (Requirements) or Phase 3 (Design) of the Full playbook

Downgrade (Full → Quick) is not supported. Once EARS requirements exist, throwing them away is wasteful — just continue in Full mode.

Examples

Example A: Quick mode (auto-picked)

User: "Add a CSV export button to the reports page"

Assistant:

Signals: solo, ~1 day, no compliance, no architectural decision → Quick mode.

Created docx/features/03-csv-export/plan.md with 4 tasks.
Review and approve, then run /dev-workflow:spec execute.

Example B: Full mode (auto-picked)

User: "Add user authentication across web + mobile, must meet SOC 2"

Assistant:

Signals: multi-platform, compliance required, JWT vs session decision → Full mode.

Created docx/features/04-user-authentication/{requirements,design,tasks}.md.
Phase 1 complete. Phase 2: I'll elicit requirements in EARS format.

1. What authentication methods should be supported?
2. What's the session lifetime requirement?
[...]

Example C: Ambiguous (asks user)

User: "Add a notifications system"

Assistant:

Scope is ambiguous between Quick and Full. Pick one:

1. Quick — single plan.md, no EARS, no RGR (≤3 days, solo)
2. Full — 3-file spec with EARS + TDD (>1 week, team, compliance)

Default if you don't pick: Quick.

Notes

• Always request explicit approval between phases
• Use EARS format strictly for requirements
• Focus on "what" and "how", not "doing"
• Explore multiple architectural options before recommending one