Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Provides practical workflows, context engineering strategies, real-world execution examples, and debugging solutions for moai-adk. Covers JIT context management, efficient agent usage, SPEC→TDD→Sync execution, and common problem resolution.
JIT (Just-In-Time) Context Strategy
Principle: Load Only What's Needed Now
Traditional (overload):
Load entire codebase
→ Context window fills immediately
→ Limited reasoning capacity
→ Slow, inefficient
JIT (optimized):
Load core entry points
→ Identify specific function/module
→ Load only that section
→ Cache in thread context
→ Reuse for related tasks
→ Minimal context waste
Practice 1: Core Module Mapping
# 1. Get high-level structure
find src/ -type f -name "*.py" | wc -l
# Output: 145 files total# 2. Identify entry points (only 3-5 files)
find src/ -name "__main__.py" -o -name "main.py" -o -name "run.py"# 3. Load entry point + immediate dependencies
Glob("src/{**/}*.py")
# Load only files referenced by entry point# 4. Cache in Task() context for reuse
Task(prompt="Task 1 using mapped modules")
Task(prompt=)
"Task 2 reuses cached context"
Practice 2: Dependency Tree Navigation
Project Root
├─ src/
│ ├─ __init__.py ← Entry point #1
│ ├─ main.py ← Entry point #2
│ ├─ core/
│ │ ├─ domain.py ← Core models
│ │ ├─ repository.py ← Data access
│ │ └─ service.py ← Business logic
│ └─ api/
│ ├─ routes.py ← API endpoints
│ └─ handlers.py ← Request handlers
Load strategy:
1. Load main.py + __init__.py (entry points)
2. When modifying API → Load api/ subtree
3. When fixing business logic → Load core/service.py
4. Cache all loaded files in context
5. Share context between related tasks
Practice 3: Context Reuse Across Tasks
# Task 1: Understand module structure
analysis = Task({
prompt="Map src/ directory structure, identify entry points, list dependencies"
})
# Task 2: Reuse analysis for implementation
implementation = Task({
prompt=f"""Using this structure:
{analysis}
Now implement feature X...
"""
})
# Task 3: Reuse analysis for testing
testing = Task({
prompt=f"""Using this structure:
{analysis}
Write tests for feature X...
"""
})
# Result: No re-mapping, efficient context reuse
SPEC → TDD → Sync Execution Pattern
Step 1: Create SPEC with /alfred:1-plan
/alfred:1-plan "Add user authentication with JWT"# This creates:# .moai/specs/SPEC-042/spec.md (full requirements)# feature/SPEC-042 (git branch)# Track with TodoWrite
Step 2: Implement with /alfred:2-run SPEC-042
RED: Test agent writes failing tests
↓
GREEN: Implementer agent creates minimal code
↓
REFACTOR: Quality agent improves code
↓
Repeat TDD cycle for each feature component
↓
All tests passing, coverage ≥85%
Step 3: Sync with /alfred:3-sync auto SPEC-042
Updates:
✓ Documentation
✓ Test coverage metrics
✓ Creates PR to develop
✓ Auto-validation of quality gates
Debugging Pattern: Issue → Root Cause → Fix
Step 1: Triage & Understand
Error message: "Cannot read property 'user_id' of undefined"
Questions:
- When does it occur? (always, intermittently, specific scenario)
- Which code path? (which endpoint/function)
- What's the state? (what data led to this)
- What changed recently? (revert to narrow down)
Step 2: Isolate Root Cause
# Method 1: Binary search# Is it in API layer? → Yes# Is it in route handler? → No# Is it in service layer? → Yes# Is it in this function? → Narrow down# Method 2: Add logging
logger.debug(f"user_id = {user_id}") # Check where it becomes undefined# Method 3: Test locally# Reproduce with minimal example# Add breakpoint in debugger# Step through execution
Step 3: Fix with Tests
# RED: Write failing testdeftest_handles_missing_user_id():
"""Should handle case when user_id is undefined."""assert get_user(None) raises ValueError
# GREEN: Minimal fixdefget_user(user_id):
ifnot user_id:
raise ValueError("user_id required")
return fetch_user(user_id)
# REFACTOR: Improvedefget_user(user_id: int) -> User:
"""Get user by ID.
Args:
user_id: User identifier
Raises:
ValueError: If user_id is None or invalid
"""ifnot user_id or user_id <= 0:
raise ValueError(f"Invalid user_id: {user_id}")
returnself.user_repo.find(user_id)
5 Real-World Scenarios
Scenario 1: Feature Implementation (2-3 hours)
1. Create SPEC: /alfred:1-plan "Add user dashboard"
2. Clarify details: AskUserQuestion (which data to show?)
3. Implement: /alfred:2-run SPEC-XXX (TDD cycle)
4. Document: /alfred:3-sync auto SPEC-XXX
5. Result: Production-ready feature
Scenario 2: Bug Investigation (1-2 hours)
1. Reproduce: Create minimal test case
2. Isolate: Narrow down affected code
3. Debug: Add logging, trace execution
4. Fix: TDD RED→GREEN→REFACTOR
5. Validate: Ensure tests pass, regression tests
Scenario 3: Large Refactoring (4-8 hours)
1. Analyze: Map current code structure
2. Plan: Design new structure with trade-offs
3. Clone pattern: Create autonomous agents for parallel refactoring
4. Integrate: Verify all pieces work together
5. Test: Comprehensive test coverage
1. SPEC clarity: AskUserQuestion for ambiguous requirements
2. Agent routing: Delegate to specialist teams
3. Progress tracking: TodoWrite for coordination
4. Integration: Verify components work together
5. Documentation: Central SPEC as source of truth
Context Budget Optimization
Typical project context:
- Config files: ~50 tokens
- .moai/ structure: ~100 tokens
- Entry points (3-5 files): ~500 tokens
- SPEC document: ~200 tokens
→ Total: ~850 tokens per session
Reusable context:
- Load once per session
- Share across 5-10 tasks
- Saves: 3,500-8,500 tokens per session
- Result: More reasoning capacity
For detailed workflow examples: reference.md For real-world scenarios: examples.md Last Updated: 2025-11-12 Status: Production Ready (Enterprise v4.0.0)
