During Task

@during-task - Lightweight Tactical Guidance

Use DURING:

• Implementing tasks from tasks.md
• Before coding each subtask
• When switching implementation areas
• When unsure which MCP for specific subtask

CAN be called MULTIPLE times (designed for this!)

What It Does (Lightweight!)

1. Get MCP suggestions for subtask (no full Graphiti search)
2. Quick gotcha lookup (optional, only if stuck)
3. Brief workflow guidance

Token cost: ~300 tokens
Frequency: 5-10 times per spec
ROI: Stays aligned = prevents rework

Execution

Default: Just MCP Suggestions

suggest_mcps({
  task: "[specific subtask]",
  include_examples: false
});

Returns:

• Vibe (might change if switching domains)
• Top 3 MCPs for this subtask
• Brief purpose

When Stuck: Quick Gotcha Lookup

search_memory_facts({
  query: "[specific tech/component] [specific issue]",
  group_ids: ["screengraph"],
  max_facts: 3  // Just top 3, not 10!
});

Returns:

• Known workarounds
• Quick fixes
• Past solutions

Output Format (Brief!)

**Task**: [subtask]
**MCPs**: [mcp1], [mcp2], [mcp3]

Do this:
1. [action 1]
2. [action 2]

~5 lines. That's it.

Integration

With Spec-Kit Implementation

# Already ran @before-task during discovery
# Now implementing tasks.md

# Task 1: Create database schema
@during-task Create user table schema
→ MCPs: encore-mcp, context7
# Code it

# Task 2: Add API endpoint  
@during-task Add user registration endpoint
→ MCPs: encore-mcp, sequential-thinking
# Code it

# Task 3: Build UI component
@during-task Build registration form
→ MCPs: svelte, browser
# Code it

# All tasks done → Run @after-task

Domain Switching

# Working on backend
@during-task Add database migration
→ Vibe: backend_vibe, MCPs: encore-mcp

# Now switching to frontend
@during-task Update UI to show new field
→ Vibe: frontend_vibe, MCPs: svelte, browser
# ✅ Vibe changed automatically!

Token Efficiency

Good Usage (Specific)

@during-task Add password validation logic      → 300 tokens ✅
@during-task Create login form component        → 300 tokens ✅
@during-task Write unit test for auth endpoint  → 300 tokens ✅

Bad Usage (Too Broad)

@during-task Implement entire authentication feature → 2000 tokens ❌
# This should be @before-task, not @during-task!

Rules

✅ Call for each subtask - Designed for frequent use
✅ Be specific - "Add validation" not "implement feature"
✅ Skip Graphiti re-search - Already have context from @before-task
❌ Don't use for discovery - That's @before-task
❌ Don't call for trivial changes - Changing a variable name doesn't need context

When NOT to Call

Skip @during-task for:

• Trivial changes (typo fixes, variable renames)
• Copy-paste work (adapting existing code)
• Following exact instructions from plan.md

Use @during-task for:

• New implementation work
• Switching between domains
• Unsure which MCP to use
• Implementing complex logic

Comparison

Without @during-task:
  → Implement blindly
  → Use wrong MCP
  → Waste time

With @during-task:
  → Quick guidance (300 tokens)
  → Right MCP immediately
  → Stay on track

300 tokens to avoid 30 minutes of wrong direction = 100x ROI.

Purpose: Provide lightweight, frequent check-ins during implementation without burning tokens on redundant searches.