Task Decomposition

Task Decomposition

When to Use

Use this skill when given a high-level goal that needs to be broken into specific, assignable tasks for a team of agents.

Procedure

1. Read and Understand the Goal

# Read the goal from the task or input
GOAL_FILE="${1:-/home/shared/goal.md}"
cat "$GOAL_FILE" 2>/dev/null

# Or read from task board
bash /home/shared/scripts/task.sh get "$TASK_ID" | jq -r '.description' 2>/dev/null

Write down:

• End state: What does "done" look like? What artifact(s) exist when complete?
• Constraints: Deadlines, technology limits, available tools, team size
• Assumptions: What are we taking for granted?

2. Discover Available Team Members

echo "=== Available Agents ==="
# List all agent users in the system
getent group agents 2>/dev/null | cut -d: -f4 | tr ',' '\n' | while read agent; do
  if [ -n "$agent" ]; then
    # Get their role from passwd GECOS field
    ROLE=$(getent passwd "$agent" 2>/dev/null | cut -d: -f5)
    echo "  $agent — $ROLE"
  fi
done

echo ""
echo "=== Agent Capabilities ==="
# Check each agent's persona/skills
for agent_home in /home/*/; do
  agent=$(basename "$agent_home")
  [ "$agent" = "shared" ] && continue
  if [ -f "$agent_home/.claude/CLAUDE.md" ]; then
    echo "--- $agent ---"
    head -5 "$agent_home/.claude/CLAUDE.md" 2>/dev/null
  fi
done

3. Decompose into Tasks

Break the goal into tasks following these rules:

1. Each task has exactly one owner. If two personas need to collaborate, create separate tasks with a dependency between them.
2. Each task has a clear deliverable. Not "work on X" but "produce X and write it to /home/shared/path."
3. Each task has explicit inputs and outputs. Where does the input come from? Where does the output go?
4. Maximize parallelism. Tasks that can run concurrently should have no dependency between them.
5. No cycles in the dependency graph. Task A depends on B, B depends on A is forbidden.

4. Define Each Task

For each task, specify:

### Task [ID]: [Subject]
- **Owner:** [persona/agent name]
- **Description:** [What to do — specific and actionable]
- **Input:** [What to read, where to find it]
- **Output:** [What to produce, where to put it]
- **Depends on:** [Task IDs that must complete first, or "none"]
- **Estimate:** [S/M/L — Small: <30min, Medium: 30min-2hr, Large: 2hr+]

5. Build the Dependency Graph

# Visualize as text DAG
cat <<'EOF'
Goal: [High-level goal]

  T1 (researcher) ──┐
                     ├──> T3 (architect) ──> T5 (coder) ──> T7 (qa)
  T2 (researcher) ──┘                        │
                                              └──> T6 (writer) ──> T8 (editor)
  T4 (security) ─────────────────────────────────────────────────> T7 (qa)
EOF

6. Validate the Decomposition

Run these checks:

# Check 1: Every task has exactly one owner
echo "=== Owner check ==="
# (Manually verify each task has one owner field)

# Check 2: No dependency cycles
echo "=== Cycle check ==="
# List all edges and check for cycles
cat <<'EOF' > /tmp/deps.txt
T3 T1
T3 T2
T5 T3
T6 T3
T7 T5
T7 T4
T8 T6
EOF
# Simple cycle detection: if tsort fails, there is a cycle
tsort /tmp/deps.txt 2>&1
if [ $? -ne 0 ]; then
  echo "ERROR: Dependency cycle detected!"
else
  echo "OK: No cycles"
fi

# Check 3: Clear end state
echo "=== End state check ==="
echo "Final deliverable tasks (no task depends on them):"
# Tasks that appear in column 1 but never in column 2
comm -23 <(cut -d' ' -f1 /tmp/deps.txt | sort -u) <(cut -d' ' -f2 /tmp/deps.txt | sort -u)

# Check 4: Parallelism
echo "=== Parallelism check ==="
echo "Tasks with no dependencies (can start immediately):"
# Tasks in column 2 that never appear in column 1
comm -23 <(cut -d' ' -f2 /tmp/deps.txt | sort -u) <(cut -d' ' -f1 /tmp/deps.txt | sort -u)

7. Write to Task Board

# Add each task to the shared task board
bash /home/shared/scripts/task.sh add \
  --subject "Research authentication options" \
  --description "Survey available auth libraries for Node.js. Compare JWT vs session-based. Write findings to /home/shared/research-auth.md" \
  --owner "researcher" \
  --depends ""

bash /home/shared/scripts/task.sh add \
  --subject "Design auth architecture" \
  --description "Based on research, design auth system. Produce design doc at /home/shared/design-auth.md" \
  --owner "architect" \
  --depends "T1"

# ... continue for all tasks

8. Write the Spec to Shared Workspace

SPEC_FILE="/home/shared/taskplan-$(date +%Y%m%d).md"

cat > "$SPEC_FILE" <<'EOF'
# Task Plan: [Goal]

**Date:** YYYY-MM-DD
**Planner:** [agent name]

## Goal
[High-level goal statement]

## End State
[What "done" looks like — specific deliverables]

## Task Breakdown

### T1: [Subject] (owner: [persona])
- Description: ...
- Input: ...
- Output: ...
- Depends: none

### T2: ...

## Dependency Graph

T1 ──> T3 ──> T5 T2 ──> T3

## Parallelism
- Wave 1 (parallel): T1, T2, T4
- Wave 2 (after wave 1): T3
- Wave 3 (after wave 2): T5, T6
- Wave 4 (final): T7, T8

## Risks
- [What could go wrong and how to mitigate]
EOF

# Register as artifact
bash /home/shared/scripts/artifact.sh register \
  --name "task-plan" \
  --type "plan" \
  --path "$SPEC_FILE" \
  --description "Task decomposition plan for: [goal]"

Quality Checklist

• End state is specific and testable
• Every task has exactly one owner
• Every task has specific input and output locations
• No dependency cycles (tsort succeeds)
• Parallelism is maximized (tasks that CAN run in parallel DO)
• Owner assignments match persona capabilities
• Tasks are written to the task board
• Spec is written to shared workspace and registered as artifact