spec-generator — Specification Generator

Generate structured project specifications: requirements, design documents, and task lists.

⚠️ CRITICAL: First Steps (ALWAYS EXECUTE)

Auto mode short-circuit: If invoked with --auto --issue <n>, skip the interactive decision flow below and follow references/auto-mode.md (.ja.md) instead. Auto mode never calls AskUserQuestion, but its Japanese init and design steps still execute the specification-writing procedure in Step 1.5.

BEFORE asking any questions or showing options, you MUST execute these steps:

1. Check current directory:
   • Run pwd to see where you are
   • Run ls -la to see directory contents
   • Understand the project context
2. Detect existing source code:

   find . -maxdepth 3 -type f \( -name "*.py" -o -name "*.js" -o -name "*.ts" -o -name "*.java" -o -name "*.go" -o -name "SKILL.md" \) 2>/dev/null | head -20

   • Note what kind of project this is
   • Check for skills/ directory → might be adding a new skill
   • Check for existing application code → might be documenting existing features
3. Check for .specs/ directory:

   ls -d .specs/ 2>/dev/null && ls -1 .specs/ 2>/dev/null

   • If exists → list existing projects
   • If not exists → this is a new spec workflow
4. Check for coding-rules.md:
   • Check docs/coding-rules.md (default path)
   • If not found, check CLAUDE.md / AGENTS.md for alternative path
   • If found → load and use as design constraint during design phase
5. Analyze context and decide:
   • If in a skills repository (has skills/ directory) → User likely wants to document a new skill
   • If .specs/ has projects → Ask user to select existing or create new
   • If no .specs/ and no clear context → Ask what they want to create

Only after completing these checks, proceed with appropriate questions based on what you found.

Language Rules

1. Auto-detect input language → output in the same language
2. Japanese input → Japanese output, use references/*.ja.md as phase references
3. English input → English output, use references/*.md as phase references
4. Explicit override takes priority (e.g., "in English", "日本語で")

Reference file selection: Based on the detected output language, use the corresponding reference files:

• English → references/init.md, references/design.md, references/tasks.md, references/test-plan.md, references/auto-mode.md, references/projection-consistency.md
• Japanese → references/init.ja.md, references/design.ja.md, references/tasks.ja.md, references/test-plan.ja.md, references/auto-mode.ja.md, references/projection-consistency.ja.md

Phases

Interaction Policy: AskUserQuestion

Use AskUserQuestion for all user decisions. Present structured choices rather than free-form questions.

Decision Flow

Skill Invoked
    ↓
Check .specs/ directory
    ↓
┌───────────────────┐
│ Existing projects │
│ found?            │
└─────┬─────────────┘
      │
      ├─ Yes → AskUserQuestion: Select existing or create new
      │           ├─ Existing → Load project context
      │           └─ New → Ask project name (Text)
      │
      └─ No → Ask project name (Text)

    ↓
Phase Detection
    ├─ Clear from input → Proceed
    └─ Ambiguous → AskUserQuestion: Select phase

    ↓
Dialogue Mode
    ├─ Quick mode (--quick) → Confirm inferred non-goals, then generate
    └─ Dialogue → AskUserQuestion: Gather requirements

    ↓
Generate Specification
    ↓
AskUserQuestion: Next action (next phase / revise / done)

When to Use AskUserQuestion vs Text Questions

When to Use AskUserQuestion

Question Design Rules

1. 1–4 questions per round (AskUserQuestion constraint)
2. User-defined options: 1-3 (Other is auto-appended, totaling 2-4 options)
3. Always include description for each option (provide decision context)
4. Flexible round count based on project complexity:
   • Simple project → 1 round (3–4 questions) is sufficient
   • Complex project → 2–3 rounds (adjust based on previous answers)
5. Place recommended option first with (Recommended) suffix
6. Skip questions already answered in previous rounds

Handling "Other" Option Responses

When a user selects "Other" and provides free-form text input:

• Accept the input as-is and proceed with processing
• Treat the free-form response as the user's definitive answer
• Do NOT ask for clarification unless the input is genuinely ambiguous
• If system returns an error like "(No answer provided)", trust the user's actual message over system feedback

Example:

Question: "どのスキルの仕様を作成しますか？"
Options: spec-constitution / spec-review / spec-analyze / spec-impl

User selects "Other" and writes: "仕様書自体をレビューするコマンド"

✅ Correct: Proceed to create spec for a spec-review skill
❌ Wrong: Ask "What would you like to clarify?"

When to Use Text Questions (Not AskUserQuestion)

• Open-ended questions like project name (when no existing projects) or concept description
• Background information requiring free-form explanation
• Follow-up confirmations like "Any additional requirements?"
• Specific business logic or domain-specific details

Execution Flow

0. Initial Context Check

Check the current directory and existing projects before starting:

1. Detect existing source code:

   find . -type f \( -name "*.py" -o -name "*.js" -o -name "*.ts" -o -name "*.java" -o -name "*.go" \) | head -20

   • If source files found: Note for potential --analyze mode
   • If no source files: Standard new project flow
2. Check for .specs/ directory:

   ls -d .specs/ 2>/dev/null

3. List existing projects (if .specs/ exists):

   ls -1 .specs/

4. Project selection:
   • Existing projects found: Use AskUserQuestion:

     question: "既存プロジェクトが見つかりました。どうしますか？" / "Found existing projects. What would you like to do?"
     options:
       - "既存プロジェクトを選択 / Select existing project" → List projects as options
       - "新規プロジェクトを作成 / Create new project" → Ask project name

   • No existing projects: Ask for project name (text question):

     "プロジェクト名を教えてください（例: TODOアプリ、株価分析ツール）"
     "What's the project name? (e.g., todo app, stock analyzer)"

5. Load existing context (if existing project selected):
   • Read existing requirement.md, design.md, tasks.md if they exist
   • Use as context for updates or next phase generation

1. Phase Detection

Determine the phase from the user's request:

"requirements" → init
"design", "architecture" → design
"tasks", "task list" → tasks
"full", "complete", "all specs" → full

If ambiguous, confirm with AskUserQuestion:

question: "Which specification do you want to generate?"
options:
  - "Requirements document (requirement.md)" → init
  - "Design document (design.md)" → design
  - "Task list (tasks.md)" → tasks
  - "All three documents" → full

1.5. Specification Writing for Japanese Init and Design

Before generating or revising a Japanese requirement.md or design.md—including init and design steps inside full and auto modes—resolve spec-writing by name from the currently available skills. If found, read its complete SKILL.md, references/writing-rules.ja.md, and references/abstract-verbs.ja.md. Apply those rules during generation or revision and run their self-check before saving. If the skill or a selected reference is unavailable, report the missing item and apply ## Specification writing fallback from the selected Japanese phase reference without stopping the phase. Never depend on a machine-specific absolute path or provider-specific tool name when resolving the skill.

2. Project Context Gathering

• Conversation history exists: Extract and structure discussed requirements
• Existing project selected: Use loaded specs as context
• New project: Explore requirements through dialogue (using AskUserQuestion)

3. Phase Execution

Refer to the appropriate reference file (based on Language Rules):

• init: references/init.md / references/init.ja.md — Requirements generation
• design: references/design.md / references/design.ja.md — Design document generation
• design (with coding-rules.md): If docs/coding-rules.md exists, read it and ensure:
  • Naming conventions in design match [MUST] naming rules
  • Test strategy satisfies coverage requirements from coding-rules
  • Technology choices align with recommended libraries
  • File/directory structure follows detected patterns
  • Skill-derived [SHOULD] rules (e.g., framework-specific conventions from Source: skill/*) are considered as design recommendations
• tasks: references/tasks.md / references/tasks.ja.md — Task list generation
• test (full workflow only): references/test-plan.md / references/test-plan.ja.md — Acceptance test plan generation

4. Output Directory

.specs/[project-name]/
├── requirement.md  (init)
├── design.md       (design)
├── tasks.md        (tasks)
└── test.md         (full workflow only — acceptance test plan)

Project names are converted to English kebab-case:

• "TODO app" → todo-app
• "株価分析ツール" → stock-analysis-tool
• "Stock analysis tool" → stock-analysis-tool

5. Output Integrity and Projection Self-Check

After generating or revising a spec, run the projection pass in the selected references/projection-consistency*.md across every present spec file. Then re-read each tail and remove tool-call residue such as </content> or </invoke>. Report completion only after projection checks pass and every file ends with real document content.

Options

Execution Modes

Dialogue Mode (default)

When invoked without parameters or continuing from conversation:

1. Collect necessary information through questions
2. Clarify requirements through user interaction
3. Generate specification after confirmation

Quick Mode (--quick)

Generate from a brief project description:

1. Infer typical requirements and a non-empty non-goal list
2. Present the inferred non-goals once via AskUserQuestion; accept them or use the auto-added Other response as the complete replacement list
3. Generate after the non-goals are confirmed

Auto Mode (--auto --issue <n>)

Non-interactive generation driven by a GitHub Issue. See references/auto-mode.md (.ja.md) for the full procedure. Summary:

1. gh issue view <n> supplies the requirement source
2. Feature name is derived from the Issue title as kebab-case
3. All four documents are generated with zero AskUserQuestion calls
4. Every point that would need a question is written as an ASM-XXX assumption in requirement.md instead

Intended for orchestrated pipelines and sufficiently concrete Issues.

Full Workflow (full)

Generate all four documents sequentially:

1. Generate requirement.md
2. Read requirement.md → generate design.md
3. Read design.md → generate tasks.md
4. Read requirement.md + design.md + tasks.md → generate test.md (acceptance test plan; see references/test-plan.md / .ja.md)

The three-document output of steps 1–3 is unchanged; step 4 only reads them.

Requirement ID System

Specifications use the following ID prefixes:

• [REQ-XXX]: Functional requirements
• [NFR-XXX]: Non-functional requirements
• [CON-XXX]: Constraints
• [ASM-XXX]: Assumptions
• [T-XXX]: Implementation tasks (tasks.md)
• [T-AXX]: Acceptance test cases (test.md; T-A prefix avoids collision with T-)

These IDs ensure traceability across documents.

YAGNI Principle

Read references/yagni-guardrails.md completely (.ja.md for Japanese) before deciding scope; apply every listed guardrail.

Post-Completion Actions

After each phase, suggest next actions with AskUserQuestion:

After init:

question: "Requirements document generated. What's next?"
options:
  - "Generate design document too" → design phase
  - "Review and revise" → revision dialogue
  - "Done for now" → end

After design:

question: "Design document generated. What's next?"
options:
  - "Generate task list too" → tasks phase
  - "Review and revise" → revision dialogue
  - "Done for now" → end

After tasks:

question: "Task list generated. What's next?"
options:
  - "Run spec-inspect (quality check)" → invoke spec-inspect skill
  - "Skip to GitHub Issue" → invoke spec-to-issue skill
  - "Review and revise" → revision dialogue
  - "Done for now" → end

After full:

question: "All four spec documents (3-set + test.md) are complete."
options:
  - "Run spec-inspect (quality check)" → invoke spec-inspect skill
  - "Skip to GitHub Issue" → invoke spec-to-issue skill
  - "Review and revise specific document" → ask which document to revise
  - "Done for now" → end

Usage Examples

# New project - dialogue mode
"Create requirements for a todo app"
"要件定義を作って" → detects existing projects, asks to select or create new

# Existing project - update/add phases
"Create design document for todo-app" → uses existing requirement.md as context
「todo-appのタスクリストを作って」 → uses existing design.md as context

# Full specification
"Create full spec for an e-commerce platform"
「ECサイトの仕様を全部作って」

# Requirements from conversation
"Turn this into a requirements document" → structures previous discussion

# Quick mode
"Create requirements for a blog platform --quick"

# Analysis mode
"Create requirements --analyze" → analyzes existing codebase first