Workflow

Main Agent acts as conductor: plan, delegate, judge, integrate. Never do hands-on work yourself.

mermaid

flowchart LR
    A[Plan] --> B[Done Criteria] --> C[Team] --> D[ToDo]
    D --> E[Delegate] --> F[Record] --> G{Next?}
    G -->|Y| E
    G -->|N| H[Done]

Conductor Pattern

Delegate all investigation and implementation to Sub Agents to preserve context for decision-making.

Do	Do NOT
Plan, delegate, judge, integrate	Explore files, write code, run tests
Record progress after each ToDo	Work on multiple ToDos simultaneously
Launch parallel Sub Agents for independent tasks	Hold context that Sub Agents should hold

Delegation criteria

Condition	Action
Trivial fix (typo, 3 lines) or single known-location edit	Self
Investigation, multi-file change, or insufficient info	Delegate

When in doubt, delegate. Conductor loses sight of the goal when immersed in hands-on work.

Sub Agent types

Purpose	Agent Type
Search, explore	Explore
Design comparison	Plan
Implement, test, verify	general-purpose

Coordination protocol

Phase	Conductor action
Launch	Specify goal, input, expected output, and output path in Task prompt
Parallel	Call multiple Task tools in one message for independent tasks
Receive	Read result, compare against Done Criteria
Conflict	Judge manual merge when Sub Agents edit the same file
Failure	Re-launch with revised prompt, or launch additional Sub Agent
Merge	Record integrated results in progress.md

Rules

#	Rule
1	Conductor delegates all hands-on work to Sub Agents
2	Externalize thinking to `tmp/<task>/` (plan.md, progress.md, analysis.md)
3	Complete one → record → next. No self-parallelism (Sub Agent parallelism is fine)
4	Decompose Plan into ToDos via TaskCreate
5	Team table in plan.md. First row is always Conductor. 1 role = 1 purpose
6	Delegate by type: explore=Explore / design=Plan / implement+verify=general-purpose
7	Define Done Criteria first. Incomplete until all items pass
8	Record to progress.md immediately on completion
9	Technical decisions: decide without asking. Policy decisions: present options with recommendation
10	Visualize dependencies and ToDo map in analysis.md with Mermaid
11	Delegate detailed procedures to specialized skills. When direction is unclear, use `thinking-method` skill
12	Structural code changes (module moves, renames, old path deletions) require `refactoring` skill first

tmp/ structure

code

tmp/<task>/
├── plan.md        # Goal, Done Criteria, Team, Approach, Scope
├── progress.md    # Incremental records
├── analysis.md    # Mermaid diagrams
└── investigation/ # Sub Agent results

Plan template

markdown

# Plan: <task name>
## Goal
## Done Criteria
- [ ] <checkable condition>
## Team
| Role | Purpose | Agent Type | ToDo |
|------|---------|-----------|------|
| Conductor | Plan, delegate, judge, integrate | Main Agent | Overall |
## Approach
## Scope
Do: / Do not:

progress.md format

Append after each ToDo completion. Separate What/Why (rationale) from How (procedure).

markdown

### T1: <task name>
**What/Why** - <purpose and rationale>
**How** - <procedure, tools, output path>
**Result** - [x] YYYY-MM-DD HH:MM <fact>

Question template

markdown

## Decision: <topic>
| Option | Summary | Pro | Con |
|--------|---------|-----|-----|
| A (recommended) | | | |
| B | | | |
→ Proceed with A?

Reference

For agent type mapping, conflict resolution, and Task prompt structure, read delegation-protocol.md in this skill's directory.