The Spray Strategy Guide — Multi-Prompt, Multi-Model Testing
How to design prompt variations, run multi-model comparisons, and systematically find the best AI output.
The Strategy Guide 🎯
Systematic variation beats random hope.
Designing Prompt Variations
The key to effective spraying is varying one dimension at a time so you can attribute quality differences to specific changes.
Variation 1: Framing
Same question, different frame changes the response style:
| Frame | Prompt | Output Tendency |
|---|---|---|
| Expert | "As a senior data scientist, explain..." | Technical, detailed |
| Teacher | "Explain to a university student..." | Educational, structured |
| Journalist | "Write a news article about..." | Concise, quote-heavy |
| Consultant | "Advise a client on..." | Strategic, actionable |
| Sceptic | "Challenge the assumption that..." | Critical, balanced |
Variation 2: Structure
Same content request, different format instruction:
- "Write a paragraph about X"
- "Create a bullet-point summary of X"
- "Build a comparison table of X"
- "Draft an FAQ about X"
- "Write X as a step-by-step guide"
Variation 3: Constraints
Same question with different tightness:
| Constraint Level | Example |
|---|---|
| Loose | "Tell me about renewable energy" |
| Medium | "Explain the top 5 renewable energy sources for UK homes" |
| Tight | "Compare solar panels vs heat pumps for a 3-bed semi in Manchester, budget under £10k, including ROI timeline" |
Multi-Model Comparison Strategy
When to Use Which Model
| Task | Best First Choice | Best Second Choice | Why |
|---|---|---|---|
| Long-form writing | Claude | ChatGPT | Claude handles nuance and tone better |
| Code generation | ChatGPT / Copilot | Claude | ChatGPT has deeper code training |
| Factual research | Perplexity | Gemini | Both have search grounding |
| Creative/brainstorm | ChatGPT | Claude | ChatGPT is more willing to be creative |
| Data analysis | ChatGPT (Code Interpreter) | Gemini | Code execution matters |
| Summarisation | Claude | Gemini | Claude handles long contexts well |
The Three-Model Sprint
For important tasks, run the same prompt through three models simultaneously:
- Send to Claude, ChatGPT, and Gemini
- Read all three outputs (2-3 minutes)
- Pick the best overall
- Use the best output as a base and iterate in that model
Total time: 5 minutes. Quality improvement over single-model: typically 25-40%.
Temperature Spraying
Temperature controls randomness in AI outputs. Most people leave it at default. Spraying across temperature settings reveals surprising quality differences.
| Temperature | Output Character | Best For |
|---|---|---|
| 0.0-0.3 | Conservative, predictable, factual | Code, data, factual writing |
| 0.4-0.7 | Balanced (default range) | General content, business writing |
| 0.8-1.0 | Creative, varied, surprising | Brainstorming, creative writing |
| 1.0-1.5 | Wild, unpredictable, occasionally brilliant | Ideation, creative exploration |
Run the same prompt at 0.3, 0.7, and 1.0. You will see three notably different outputs. Pick the one that best suits your purpose.
Evaluating Spray Results
The Quick Evaluation Matrix
For each output, score 1-5 on four dimensions:
| Dimension | What to Evaluate |
|---|---|
| Accuracy | Are the facts correct? Claims verifiable? |
| Relevance | Does it answer what was actually asked? |
| Quality | Is it well-written, well-structured, clear? |
| Usefulness | Can you use this directly with minimal editing? |
Total score out of 20. The winner is usually clear within 30 seconds of scanning.
Cost Management
Spraying costs more than single-shot — but less than most people think.
| Strategy | Cost per Task (API) | Cost per Task (Subscription) |
|---|---|---|
| Single shot | $0.01-0.05 | Included |
| 3-model spray | $0.03-0.15 | Included (if subscribed to all 3) |
| 3-variant x 3-model | $0.09-0.45 | Included |
For subscription users (ChatGPT Plus, Claude Pro, Gemini Advanced), spraying is essentially free — you are paying flat rate regardless. There is no cost reason not to spray on subscriptions.