Prompting Toolkit: Intent-First Design

Use this toolkit to craft and evolve prompts across models and modalities. Output: 1 prompt library per use case.

What Is Intent-First Prompting?

Intent-first prompting starts with what you want the AI to do and why, then works backward to the words. It's model-agnostic—the same intent strategy works for Claude, Gemini, ChatGPT, or any multimodal system.

Key principle: The prompt is a contract. You specify what success looks like; the model delivers to spec.

Step 1: Define Intent (The Why)

Intent Statement Template

**Task:** [What is the AI doing? Be specific.]
Example: "Classify customer support tickets into 1 of 8 predefined categories."

**Success Criteria:** [How will you know it worked?]
Example: "Accuracy ≥95%, no hallucinations about categories not in the list, confidence score <0.7 triggers human review."

**Failure Mode:** [What's the worst error?]
Example: "Confidently assigning a safety-critical issue (injury, fraud) to 'general inquiry'."

**Context:** [Who uses this output? Why does it matter?]
Example: "Support team routes tickets. Wrong routing wastes 20 min/ticket and frustrates customers."

Step 2: Craft the Instruction (Model-Agnostic Core)

Use this template regardless of which model you're using:

Core Instruction Template

You are a [ROLE: expert ticket classifier / data analyst / etc.].

**Your task:** [TASK: classify the following support ticket into 1 of these categories]

**Categories (use only these):**
1. Billing & Payment
2. Technical Support
3. Account Access
4. Product Feedback
5. Complaint / Escalation
6. General Inquiry
7. Other
8. Safety / Fraud (escalate immediately)

**Instructions:**
- Read the customer message carefully.
- Identify the PRIMARY issue (if multiple topics, choose the most urgent).
- Respond in JSON format only: {"category": "...", "confidence": 0.0-1.0, "reasoning": "..."}
- If the message doesn't fit any category, respond: {"category": "Other", "confidence": ..., "reasoning": "..."}
- If ANY mention of safety (injury, fraud, threats), immediately classify as "Safety / Fraud" and STOP.

**IMPORTANT:**
- Do NOT invent categories.
- Do NOT explain beyond the JSON response.
- Confidence must reflect your certainty (0.7+ = high confidence, <0.5 = uncertain, trigger human review).

**Ticket:**
[INSERT TICKET TEXT HERE]

Step 3: Model-Specific Adjustments

For Claude (Anthropic)

Strengths: reasoning, nuance, instruction-following
Adjustments:

• Use clear structural language ("You are X. Your task is Y.")
• Ask for reasoning; Claude excels at explanation.
• Add a "think step-by-step" phrase for complex tasks.

Tweaked instruction for Claude:

You are an expert customer support analyst. Your task is to classify support tickets.

[Same categories and instructions as above, plus:]

Before responding, think through your reasoning:
1. What is the PRIMARY issue the customer is describing?
2. Which category fits best?
3. How confident are you (0.0-1.0)?

Then respond in JSON format: {"category": "...", "confidence": ..., "reasoning": "..."}

For GPT-4 / ChatGPT

Strengths: speed, diverse modalities
Adjustments:

• Be more explicit about JSON formatting (ChatGPT can be loose with format).
• Use system/user message structure if available.

Tweaked instruction for ChatGPT:

System: You are a customer support ticket classifier. You classify tickets into exactly one category.

User: 
[Same core instruction, with emphasis on JSON structure]

You MUST respond with valid JSON only:
{
  "category": "[exact category name from list]",
  "confidence": 0.0-1.0,
  "reasoning": "[2-3 sentence explanation]"
}

For Gemini (Google)

Strengths: multimodal, real-time data
Adjustments:

• Gemini handles multimodal well; if you have images/docs, include them.
• Be explicit about safety boundaries.

Tweaked instruction for Gemini:

Role: You are a content classifier.
Task: Classify the following input into one of these categories: [list].
Safety rule: If any input mentions harm, injury, or fraud, classify as "Safety / Fraud" immediately.
Output format: JSON only.

Step 4: Test & Iterate (The Evaluation Loop)

Test Dataset

Create 5-10 labeled examples that cover:

• Straightforward case (clear category)
• Ambiguous case (could fit 2 categories)
• Rare/edge case (unusual request)
• Safety/fraud case
• Out-of-scope case (should be "Other")

Evaluation Matrix

Iteration Rules

If accuracy <95%:

1. Add clarifying examples to the prompt (show the model a category it confused).
2. Adjust the instruction language (e.g., "Safety is ANY mention of injury, fraud, or threat").
3. Try a different model (Claude for nuance, GPT-4 for speed).

If confidence scores are wildly different across models:

1. Increase the specificity of the prompt.
2. Ask the model to "explain first, then classify" (forces reasoning).

Step 5: Deploy & Monitor (Continuous Learning)

Monitoring Dashboard

**Prompt Performance (Weekly)**

| Metric | Target | Actual | Status |
|--------|--------|--------|--------|
| Accuracy | ≥95% | 94.2% | ⚠️ Declining |
| Confidence (avg) | ≥0.85 | 0.78 | ⚠️ Drift |
| Hallucinations | 0 | 2/500 | ✓ Acceptable |
| Human override rate | <5% | 3.2% | ✓ Good |

**Recent Changes:**
- Model upgraded to Opus 4.5 (2/1). Accuracy dropped 1.5%; investigating prompt compatibility.

**Next Action:**
- Retrain prompt for Opus 4.5 using updated test set.

Quarterly Prompt Refresh

Every 3 months, review:

• Model accuracy drift (new model versions may need prompt tweaks).
• Category distribution changes (new ticket types emerging?).
• User feedback on misclassifications (add examples to prompt).
• Cost/latency (should we switch models?).

Prompt Library Template

Keep all production prompts in a single document (or database) for version control and comparison:

# Production Prompts (2026 Q1)

## Prompt v2.3: Support Ticket Classification
**Created:** 2026-01-15  
**Last Updated:** 2026-02-01  
**Model(s):** Claude 3.5 Sonnet (primary), GPT-4 (backup)  
**Accuracy:** 94.8% (100 test cases)  
**Status:** ACTIVE

**Full Prompt:**
[Paste complete prompt here]

**Test Results:**
[Link to evaluation matrix]

**Known Issues:**
- Gemini 3.0 struggles with "Technical Support vs. Product Feedback" distinction; not recommended for this task.

**Next Review:** 2026-05-01

---

## Prompt v2.2: Support Ticket Classification (DEPRECATED)
**Retired:** 2026-02-01  
**Reason:** Accuracy <92% with new model releases.  
**Archive:** [Link to archived version]

Foundational Skills Checklist

• Prompting: Intent clearly stated, instruction model-agnostic, test cases documented
• Critical Evaluation: Accuracy measured, confidence scores validated, edge cases tested
• Workflow Integration: Prompt versioned, handoff to human review defined (e.g., confidence <0.7)
• AI Strategy: Prompt supports business goal (ROI, risk threshold, compliance requirement)

Quick Reference: Prompt Dos & Don'ts