17. Meta Prompting

flowchart TD
    subgraph MANUAL["Manual Iteration"]
        M1["Write prompt"] --> M2["Test"]
        M2 --> M3["Manual tweak"]
        M3 --> M4["Test again"]
        M4 --> M5["❌ Slow, 5-10 iterations"]
    end

    subgraph META["Meta Prompting"]
        MP1["Task description"] --> MP2["Ask LLM:\n'Write a good prompt\nto accomplish this'"]
        MP2 --> MP3["LLM generates\noptimized prompt"]
        MP3 --> MP4["Use it"]
        MP4 --> MP5["✅ Fast, 1-2 iterations"]
    end

    style MANUAL fill:#ef4444,color:#fff
    style META fill:#22c55e,color:#fff

flowchart LR
    META["Meta Prompting"] --> IMPROVE["Improve Prompt\nAnalyze and optimize"]
    META --> GENERATE["Generate Prompt\nCreate from description"]
    META --> CRITIQUE["Critique Prompt\nIdentify weaknesses"]
    META --> OPTIMIZE["Optimize for\nSpecific model"]
    META --> TRANSLATE["Translate Prompt\nAcross contexts"]

    style META fill:#8b5cf6,color:#fff
    style IMPROVE fill:#3b82f6,color:#fff
    style GENERATE fill:#22c55e,color:#fff
    style CRITIQUE fill:#f59e0b,color:#fff
    style OPTIMIZE fill:#ec4899,color:#fff
    style TRANSLATE fill:#14b8a6,color:#fff

I have this prompt:

[your prompt]

Please analyze it and suggest improvements:
1. What's missing?
2. What's unnecessary?
3. How could the structure be better?
4. How could the wording be clearer?
5. Are there any ambiguities?

Then provide an improved version of the prompt.

I need a prompt for the following task:

Task: [describe what you want the AI to do]
Context: [relevant background]
Target Audience: [who will use the output]
Output Requirements: [format, constraints]

Please generate an optimized prompt that will produce the best results.
Include:
- System prompt (persona, rules)
- User prompt structure
- Output format specification
- Example (if helpful)

Evaluate this prompt for potential issues:

[prompt]

Consider:
1. Clarity: Is the instruction clear?
2. Completeness: Is anything missing?
3. Bias: Does it lead the model?
4. Robustness: How does it handle edge cases?
5. Token efficiency: Is it concise?

Rate each aspect 1-10 and provide specific recommendations.

I'm using Chain of Thought prompting for this task. Here's my current prompt:

[CoT prompt]

How can I make the reasoning steps clearer or more effective?
Suggest 3 specific improvements.

I need the LLM to output in this format:

[describe format]

Here's my current prompt for this:

[prompt]

Is there a better way to specify this format to ensure
the model follows it consistently?

I'm using few-shot prompting for [task type].
Here are my current examples:

[examples]

Are these the right examples? Should I add, remove, or replace any?
What edge cases should I cover?

User:
"I need a prompt that helps junior developers debug React
performance issues. The prompt should guide them through
a systematic debugging process. Output should be actionable."

Meta-LLM Response:
System: You are a senior React performance engineer mentoring a junior developer.

Prompt:
"Walk through debugging this React performance issue step by step:

1. FIRST — Identify the symptom:
   [describe what's slow - renders? interactions? load time?]

2. CHECK — Common causes:
   - Is it unnecessary re-renders? (check React DevTools)
   - Is it expensive computations? (check useMemo opportunities)
   - Is it large lists? (check virtualization needs)

3. DIAGNOSE — Use React DevTools Profiler to:
   - Record a profile of the slow interaction
   - Identify which component re-rendered most
   - Check if props changed unexpectedly

4. FIX — Based on diagnosis:
   - memo() for stable props
   - useMemo for expensive calculations
   - useCallback for stable callbacks
   - Key props for list stability
   - Virtualization for long lists

5. VERIFY — Profile again to confirm improvement

Output format:
## Symptom
## Likely Cause
## Diagnosis Steps
## Recommended Fix
## Verification"

User:
"Here's my prompt for extracting data from invoices:

'Extract the invoice number, date, vendor name, and total amount
from this invoice text. Return as JSON.'

How can I improve it?"

Meta-LLM Response:
"Issues with current prompt:
1. ❌ No output format example — model may guess field names
2. ❌ No handling for missing fields — what if date is missing?
3. ❌ No instructions for amounts — should they be numbers or strings?
4. ❌ No constraints on extra text — model may add explanations

Improved version:

Extract the following fields from this invoice text.
If a field is not found, use null.
Return ONLY the JSON object with no additional text.

Fields to extract:
- invoice_number: string (e.g., "INV-001")
- date: string in YYYY-MM-DD format (or null)
- vendor_name: string (or null)
- total_amount: number (decimal, no currency symbol)

Example output format:
{
  "invoice_number": "INV-001",
  "date": "2024-01-15",
  "vendor_name": "Acme Corp",
  "total_amount": 1500.50
}

Invoice text:
[invoice text here]"