A 5-step workflow that systematically carries out feature implementation.
Core Principles
- Context Isolation: Each Step is recommended to be executed in a new conversation
- Human in the Loop: Proceed after user confirmation
- Document as Interface: Communication between Steps is done through documents
- Git as History: Create a checkpoint with a commit upon completion of each Step
Workflow Overview
| Step | Role | Input | Output | Details |
|---|---|---|---|---|
| 1 | Requirements Analyst | 00-user-prompt.md | 10-output-plan.md | step-1.md |
| 2 | System Designer | 10-output-plan.md | 20-output-system-design.md | step-2.md |
| 3 | Task Analyzer | 10+20 | 30-output-task.md + todos/ | step-3.md |
| 4 | Coordinator (Team Lead) | todos/*.md | 40-output-implementation.md | step-4.md |
| 5 | Reviewer | 40-output-impl.md | 50-output-review.md | step-5.md |
Load Rules
At the start of each Step, assets/rules/AGENTS.mdread and load the rules:
- Required:
MUST/workflow-rule.md(always) - Domain: Dynamically loaded according to the task context (react/, testing/, api/, etc.)
Start a New Task
- Determine Task ID: Request from the user (e.g.,
PROJ-001) - Initialize Task:
./scripts/task.sh init <TASK_ID> - Write Input:
.ai/tasks/<TASK_ID>/00-user-prompt.mdedit - Load Rules: assets/rules/AGENTS.md read
- Run Step 1: references/step-1.md refer to
Resume Task
When the user mentions a TASK_ID or requests "continue the task":
status.yamlread → identify the current Step/status- Load rules → proceed with the reference document for the corresponding Step
Detailed procedure: references/resume-guide.md
Check Progress Status
./scripts/task.sh status <TASK_ID>
./scripts/task.sh list
Step Summary
Step 1: Requirements Analysis
- Role: Requirements Analyst — analyze and structure user requirements
- Input:
00-user-prompt.md→ Output:10-output-plan.md - Details: references/step-1.md
Step 2: Design & Planning
- Role: System Designer — establish design and implementation plan based on requirements
- Input:
10-output-plan.md→ Output:20-output-system-design.md - Details: references/step-2.md
Step 3: Task Analysis
- Role: Task Analyzer — decompose the design into work units, plan parallelization
- Input:
10 + 20→ Output:30-output-task.md+todos/*.md - Details: references/step-3.md
Step 4: Implementation (Agent Team)
- Role: Coordinator — spawn an Agent Team for parallel implementation
- Input:
todos/*.md→ Output:40-output-implementation.md+ code - Method:
TeamCreate→TaskCreate→Task(Spawn Worker) →SendMessage(Monitoring) →TeamDelete(Cleanup) - Details: references/step-4.md / Detailed Team Spawn Guide
Step 5: Review (code-review-team)
- Role: Reviewer — expert review based on the code-review-team skill
- Input:
40-output-implementation.md→ Output:50-output-review.md - Method: Phase 1 expert review → user confirmation → Phase 2 team improvement → PR
- Details: references/step-5.md
Detailed Guide
| Document | Description |
|---|---|
| references/step-1.md | Requirements Analysis Procedure |
| references/step-2.md | Design and Planning Procedure |
| references/step-3.md | Task Decomposition Procedure |
| references/step-4.md | Agent Team Implementation Procedure |
| references/step-5.md | Code Review and PR Procedure |
| references/team-spawn.md | Step 4 Team Spawn Pattern |
| references/resume-guide.md | Detailed Procedure for Resuming Work |
| assets/templates/ | Output Document Template |
| scripts/task.sh | Task Management Script |