alivirgo Batch Small Questions Together (Token Batching) (AI Skill)
Overview
Every time you send a new message in an ongoing AI conversation, the entire previous chat history is re-sent as input tokens. Sending 5 separate, one-sentence questions sequentially creates quadratic context accumulation ($O(N^2)$ token costs), triggering rate limits and wasting time.
Token Batching is the practice of combining 3 to 5 related micro-questions into a single structured prompt, receiving a clean, multi-part answer in one fast round-trip.
The Economics of Context Accumulation
┌─────────────────────────────────────────────────────────────┐
│ Sequential vs. Batched Calls │
│ │
│ Sequential (5 separate messages): │
│ Turn 1: Base Context (500 tokens) │
│ Turn 2: Base Context + Turn 1 (1,200 tokens) │
│ Turn 3: Base Context + Turns 1-2 (1,900 tokens) │
│ Turn 4: Base Context + Turns 1-3 (2,600 tokens) │
│ Turn 5: Base Context + Turns 1-4 (3,300 tokens) │
│ Total Tokens Billed: ~9,500 tokens (5 round trips) │
│ │
│ Batched (1 message with 5 sub-questions): │
│ Turn 1: Base Context + 5 Sub-questions (750 tokens) │
│ Total Tokens Billed: ~1,500 tokens (84% Savings, 1 Trip) │
└─────────────────────────────────────────────────────────────┘
Master Batching Prompt Templates
Pattern 1: The Numbered Q&A Batch (Research & Fact-Finding)
Based on the attached documentation / topic of [TOPIC], answer these 4 specific questions in order:
1. **Prerequisites**: What specific SDK version and runtime is required?
2. **Pricing**: What is the free-tier allowance vs. overage cost per 1k requests?
3. **Authentication**: Does this use OAuth2 Bearer tokens or static API keys?
4. **Rate Limits**: What are the requests-per-minute (RPM) limits for standard accounts?
Format each answer as:
**[Number]. [Topic Name]**: [1-2 concise sentences]
Pattern 2: The Multi-Variable Debug Batch (Technical)
Review the function below and answer these 3 diagnostic points:
[PASTE FUNCTION]
1. What is the time complexity in Big-O notation?
2. Is there an unhandled null pointer or memory leak?
3. What is the exact one-line fix to make it thread-safe?
Real-World Case Study
Scenario: Reviewing a Freelance Contract
Wasteful Sequential Flow (5 Back-and-Forths)
- Turn 1: "What is the payment term in section 3?" $\rightarrow$ AI responds (200 tokens).
- Turn 2: "Is there an intellectual property assignment clause?" $\rightarrow$ AI re-reads entire doc (1,500 tokens).
- Turn 3: "What is the notice period for termination?" $\rightarrow$ AI re-reads entire doc (1,800 tokens).
- Turn 4: "Is there a non-compete clause?" $\rightarrow$ AI re-reads entire doc (2,100 tokens).
❌ Total Time: 4 minutes. Total Tokens: ~6,000 tokens.
The Single Batched Audit Prompt & Output
User: *"Analyze this contract and answer in 4 bullet points:
- Payment schedule & late penalty terms
- IP ownership assignment timing (upon creation vs. upon full payment)
- Termination notice period required
- Non-compete duration and geographic restrictions"*
AI Response:
- Payment Terms: Net-30 from invoice date; 1.5% monthly late fee applies after 30 days (Section 3.2).
- IP Ownership: IP transfers only upon full payment received (Section 5.1 - favorable to freelancer).
- Termination: 14 days written notice required by either party without cause (Section 8.3).
- Non-Compete: None. Only a standard 1-year client non-solicitation clause (Section 9.2).
(Completed in 1 turn, 4 seconds total generation time, 1,200 tokens total).
Best Practices for Batching
- Keep questions related: Batch questions that draw from the same source document or context.
- Number each item: Numbering ($1, 2, 3$) ensures the model doesn't accidentally skip an item.
- Set a ceiling of 3-5 items: Asking 15 questions in one turn can lead to shallow answers or truncation.