Skeleton-of-Thought (SoT) Skill

Achieves up to 2x speedup through parallel content generation.

Core Concept

Instead of sequential generation, SoT:

1. First generates a skeleton (outline) of the answer
2. Then expands each skeleton point in PARALLEL
3. Finally assembles the complete response

When to Use

• Long-form responses (>500 words expected)
• Structured content (lists, tutorials, explanations)
• Time-sensitive queries
• Batch processing scenarios

When NOT to Use

• Math/reasoning problems (need sequential thought)
• Simple factual queries
• Creative writing requiring flow
• Code generation requiring context

Process

Step 1: Skeleton Generation

Given question, generate ONLY the skeleton outline:
- Point 1: [brief description]
- Point 2: [brief description]
- Point 3: [brief description]
...
Do NOT expand. Just the skeleton.

Step 2: Parallel Expansion

Launch parallel expansions for each point:

Point 1 → Agent 1 → Expanded content
Point 2 → Agent 2 → Expanded content
Point 3 → Agent 3 → Expanded content
(all run simultaneously)

Step 3: Assembly

Combine expanded points with transitions:
[Introduction]
[Point 1 expanded]
[Transition]
[Point 2 expanded]
[Transition]
[Point 3 expanded]
[Conclusion]

Implementation

python
# Skeleton generation prompt
SKELETON_PROMPT = """
For the question: {question}

Generate ONLY a skeleton outline with 3-8 key points.
Format:
1. [Point]: [5-10 word description]
2. [Point]: [5-10 word description]
...

Do NOT expand the points. ONLY the skeleton.
"""

# Point expansion prompt
EXPAND_PROMPT = """
Context: Answering "{question}"
Skeleton: {skeleton}

Expand ONLY point {point_number}: "{point_description}"

Write 2-4 sentences expanding this point.
Do not include other points.
"""

Performance

| Query Type | Sequential Time | SoT Time | Speedup | |------------|-----------------|----------|---------| | Tutorial | 10s | 5s | 2.0x | | Explanation| 8s | 4.5s | 1.8x | | List-based | 12s | 6s | 2.0x | | Analysis | 15s | 9s | 1.7x |

Quality Preservation

• Skeleton ensures coherent structure
• Each expansion has full question context
• Assembly adds smooth transitions
• Quality remains high despite parallelism

Combination with Other Skills

• Self-Consistency: Generate multiple skeletons, vote on best
• ToT: Use ToT for skeleton generation, SoT for expansion
• LATS: Apply LATS scoring to skeleton options

Reference: "Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding" (2023)