Requirement Workflow (SDD)

Structured development orchestrator based on Spec-Driven Development: define specs first, decompose into tasks, execute with quality gates, verify against acceptance criteria.

Core principle: Spec is the single source of truth. Code follows spec, not the other way around.

Prerequisites

• Node.js >= 18
• Writable project directory for workflow state files (.trae/workflow/)

Quick Start

User says anything about building/implementing/fixing →
  1. AI classifies the request (type + size + risk)
  2. AI runs: node init.cjs -r <root> -n <name> -t <type> -s <size> -k <risk>
  3. AI fills spec.md with structured requirements (EARS format)
  4. AI decomposes spec.md into tasks.md
  5. AI implements task by task, advancing stages
  6. AI verifies against checklist.md

Working Modes

This skill supports two modes, selected based on classification:

Agent Mode (Default)

For small/tiny scope or bugfix type: skip spec phase, go straight to implementation.

INIT → IMPLEMENTING → TESTING → DONE

Spec Mode (SDD)

For medium/large scope or elevated/critical risk: full SDD lifecycle.

INIT → DEFINING → PLANNING → DESIGNING → IMPLEMENTING → TESTING → DELIVERING → DONE

Classification Matrix

Signal	Type	Size	Risk
"fix bug", "broken", "crash"	bugfix	tiny-small	normal
"add feature", "implement", "build"	feature	small-large	normal
"refactor", "clean up", "reorganize"	refactor	small-medium	normal
"tech debt", "upgrade", "migrate"	tech-debt	medium-large	elevated
auth, payments, data, security	any	any	elevated-critical
>15 files estimated	any	large	elevated

Size Heuristics

Size	Files	Duration	Stages
tiny	≤2	<30min	INIT → IMPLEMENTING → DONE
small	3-5	30min-2h	INIT → IMPLEMENTING → TESTING → DONE
medium	6-15	2h-1d	Full SDD
large	>15	>1d	Full SDD + all checkpoints

SDD Lifecycle

Stage 1: INIT

Classify request, initialize workflow, create artifact directory.

node init.cjs -r <project_root> -n "<name>" -t <type> -s <size> -k <risk>

Output: workflow.yaml, empty spec.md, tasks.md, checklist.md

Stage 2: DEFINING (Spec Mode only)

Fill spec.md with structured requirements using EARS format:

# Feature Name

## Background
{Why this is needed — problem statement}

## Scope
- In: {what IS included}
- Out: {what is NOT included}

## Acceptance Criteria (EARS format)
- [ ] When <condition>, the system shall <response>
- [ ] While <state>, the system shall <behavior>
- [ ] Where <constraint>, the system shall <limit>

## Constraints
- {Performance, security, compatibility requirements}

## Out of Scope
- {Explicitly deferred items}

Checkpoint: If risk is elevated/critical, pause for user review of spec.md.

Hooks: pre_stage_DEFINING → iron-audit-pm, problem-definer, risk-auditor

Stage 3: PLANNING (Spec Mode only)

Decompose spec.md into tasks.md — every acceptance criterion maps to ≥1 task:

# Tasks

## Phase 1: Foundation
- [ ] Task 1.1: {description} [files: x, y]
- [ ] Task 1.2: {description} [files: z]

## Phase 2: Core Logic
- [ ] Task 2.1: {description} [files: a, b]

## Phase 3: Integration & Polish
- [ ] Task 3.1: {description} [files: c]

## Verification
- [ ] All acceptance criteria pass
- [ ] Lint clean
- [ ] Type check pass

Rule: Each task must be atomic (completable independently) and traceable to a spec item.

Hooks: pre_stage_PLANNING → story-mapper, mvp-freeze-architect

Stage 4: DESIGNING (Spec Mode only)

Create design.md (only for medium+ size) with architecture decisions:

• Component structure
• Data flow
• API contracts
• Key trade-offs

Checkpoint: If risk is elevated/critical, pause for user review.

Hooks: pre_stage_DESIGNING → domain-modeler, architecture-advisor, responsibility-modeler

Stage 5: IMPLEMENTING

Execute tasks from tasks.md sequentially. For each task:

1. Read the task description
2. Implement the change
3. Mark task as [x] in tasks.md
4. Run relevant tests if available

Hooks: pre_stage_IMPLEMENTING → tdd-coach

Stage 6: TESTING

Run full test suite. Update checklist.md:

# Checklist

## Code Quality
- [ ] Implementation complete
- [ ] Lint clean (run lint command)
- [ ] Type check pass (run typecheck command)

## Tests
- [ ] Unit tests pass
- [ ] Integration tests pass (if applicable)

## Acceptance Criteria
- [ ] AC 1: {criterion from spec} — verified
- [ ] AC 2: {criterion from spec} — verified

## Review
- [ ] Self-review complete
- [ ] No TODO/FIXME left unresolved

Hooks: pre_stage_TESTING → test-strategy-advisor, test-strategist, code-reviewer

Stage 7: DELIVERING (Spec Mode only)

Final verification against spec.md. Ensure every acceptance criterion has been met.

Hooks: post_stage_DELIVERING → code-reviewer, tech-design-reviewer

Stage 8: DONE

Workflow complete. Summary output:

• What was delivered
• Files changed
• Tests passed
• Time elapsed

Stage Advancement

# Check current status
node status.cjs -r <project_root>

# Advance to next stage
node advance.cjs -r <project_root>

# Force advance (skip checkpoint)
node advance.cjs -r <project_root> --force

Hooks System

Hooks are agents/skills that run at stage transitions. Three scopes:

Scope	File	Priority
Global	hooks.yaml (skill dir)	Lowest
Project	.trae/workflow/hooks.yaml	Medium
Workflow	workflow.yaml (in workflow dir)	Highest

Hook Points

Hook	When	Default Agents
pre_stage_DEFINING	Before filling spec	iron-audit-pm, risk-auditor
pre_stage_PLANNING	Before task decomposition	story-mapper, mvp-freeze-architect
pre_stage_DESIGNING	Before architecture	domain-modeler, architecture-advisor
pre_stage_IMPLEMENTING	Before coding	tdd-coach
pre_stage_TESTING	Before test phase	test-strategy-advisor, code-reviewer
post_stage_DELIVERING	After final check	tech-design-reviewer

# List hooks
node hooks.cjs -r <project_root> list

# Add a hook
node hooks.cjs -r <project_root> add pre_stage_TESTING -n my-validator --type skill

Quality Gates

Checkpoint Rules

Stage	Checkpoint When
DEFINING	risk = elevated or critical
PLANNING	size = large or risk ≥ elevated
DESIGNING	size ≥ medium or risk ≥ elevated
TESTING	always (all risk levels)

At checkpoints, AI pauses and asks user for confirmation before advancing.

SDD Traceability Rule

Every line of code must trace back to:

1. A task in tasks.md
2. Which traces to an acceptance criterion in spec.md

If you find yourself writing code not covered by any task — stop and update tasks.md first.

Error Handling

Issue	Solution
User gives vague request	Classify as feature/small, use Agent Mode, refine during implementation
Spec is incomplete	Add missing acceptance criteria before advancing to PLANNING
Task is too large	Break into sub-tasks, each ≤1 file change
Tests fail during TESTING	Stay in TESTING, fix issues, re-run
Checkpoint rejected by user	Stay in current stage, revise artifacts per feedback
Workflow abandoned	No cleanup needed, state persists in .trae/workflow/

Scripts

Script	Purpose
init.cjs	Initialize workflow with classification
advance.cjs	Advance to next stage
status.cjs	Show current workflow status
hooks.cjs	Manage hook registrations
lib/common.cjs	Shared YAML utilities

Agents

See AGENTS.md for the full registry. Key agents by phase:

Phase	Agent	Methodology
DEFINING	iron-audit-pm	PRD audit, DNA extraction
DEFINING	risk-auditor	Risk scanning
DEFINING	problem-definer	Weinberg problem analysis
PLANNING	story-mapper	Patton story mapping
PLANNING	mvp-freeze-architect	Scope freezing
DESIGNING	domain-modeler	DDD/Evans modeling
DESIGNING	architecture-advisor	Quality attributes
IMPLEMENTING	tdd-coach	Beck TDD cycle
TESTING	test-strategist	Crispin test strategy
TESTING	code-reviewer	Code review
DELIVERING	tech-design-reviewer	Architecture review

References

• Quick Reference - Condensed decision tables