alivirgo Ask AI How to Improve Your Prompt (Meta-Prompting) (AI Skill)
Overview
Prompt engineering is not guessing magical keywords - it is providing the right balance of context, task constraints, and format specifications.
Instead of struggling through trial-and-error when a response falls flat, you can use Meta-Prompting: asking the AI itself to diagnose why your prompt was underspecified and write a significantly upgraded version.
The Meta-Prompting Feedback Loop
┌─────────────────────────────────────────────────────────────┐
│ The Prompt Optimizer Engine │
│ │
│ User Initial Prompt ──► [ AI Output Was Subpar / Generic ] │
│ │ │
│ Meta-Audit Directive ◄───────────┘ │
│ "Diagnose my prompt's missing context & rewrite it" │
│ │ │
│ ▼ │
│ [ Upgraded Prompt Template with Constraints & Exemplars ] │
└─────────────────────────────────────────────────────────────┘
Master Meta-Prompt Templates
Pattern 1: The Prompt Doctor & Upgrader
Use this whenever an answer feels mediocre, generic, or off-target:
Review the prompt I gave you earlier: "[PASTE YOUR ORIGINAL PROMPT]"
Act as a Principal AI Prompt Engineer. Do the following:
1. **Gap Analysis**: What vital context, constraints, or formatting instructions was my prompt missing?
2. **Ambiguity Check**: Where could an LLM easily misunderstand my intent?
3. **Upgraded Prompt**: Rewrite my prompt into a state-of-the-art, production-grade template with explicit roles, guidelines, and output schemas.
4. **Demonstration**: Run the upgraded prompt right now and show the improved output.
Pattern 2: The Reusable Template Generator
Use this to turn a one-time success into a permanent team workflow:
I loved the answer you just generated.
Now, write the master prompt template that produced this result.
- Replace specific variables with bracketed placeholders (e.g., `[TOPIC]`, `[TARGET_AUDIENCE]`, `[MAX_WORDS]`).
- Include the exact system rules and output formatting instructions.
- Save it in a clean copy-paste markdown code block.
Real-World Case Study
Scenario: Writing Customer Support Responses
The Vague Initial Prompt
"Help me reply to an angry customer whose shipment was lost."
❌ Result: Overly formal, stiff email full of apologies without clear compensation rules or tracking steps.
Running the Prompt Doctor Directive
User: "How should I reword that prompt to get a world-class customer service response?"
AI Prompt Doctor Response:
Diagnosis: Your prompt lacked 4 crucial parameters:
- Tone level (Empathic vs. Formal).
- Resolution policy (Refund vs. Reshipment vs. Store credit).
- Required placeholders (Order #, tracking link).
- Guardrails (Do not admit legal liability or make unrealistic delivery promises).
The Upgraded Prompt Template:
Act as a Senior Customer Delight Specialist for a premium e-commerce brand.
Draft a reply to a customer whose order has been lost in transit for >7 business days.
Key Information:
- Customer Name: [NAME]
- Order ID: [ORDER_NUMBER]
- Resolution: We have issued an immediate free expedited reshipment AND a $20 store credit code.
Guidelines:
1. Tone: Deeply empathetic, accountable, proactive, and warm.
2. Structure:
- Acknowledge and apologize sincerely in sentence 1 (no corporate excuses).
- State the 2-part resolution clearly in bullet points.
- Provide the new tracking number placeholder and expected arrival window.
- Offer the $20 gift code for their next order.
3. Length: Under 150 words.
Anatomy of a Perfect Prompt
Every high-performing prompt generated by the Meta-Prompting skill contains these 5 pillars:
| Pillar | Purpose | Example |
|---|---|---|
| 1. Role | Sets tone and domain expertise | "Act as a Lead Python Backend Engineer..." |
| 2. Context | Background situation and problem | "We are running FastAPI with AsyncPG on Kubernetes..." |
| 3. Task | The exact atomic deliverable | "Write a database connection pool manager..." |
| 4. Constraints | Guardrails, banned items, word counts | "Must handle reconnection retries, no third-party ORMs..." |
| 5. Format | The exact schema and structure | "Output as a single code block with type hints..." |