So I’ve read the Claude, Meta and some of the other models prompting guides but I think the YCombinator video they just released explains them well.
If you are creating an AI prompt for something even a little complex then you should ask it first to provide a better prompt. This is called Meta Prompting. You ask the AI how to improve the prompt, especially useful is to run it on the beefier models (e.g on Claude 4). You can also tell it the negatives of the outputs when you run the updated prompt. This lets you quickly iterate.
With this technique you should be able to use a smaller model because of the better prompt, which means reduced costs and faster inference time.
Here’s an example of what I’m doing:
I’m giving you, Perplexity an AI prompt to get the psychometric profile of a person.
Please provide a better, more useful LLM prompt.
<prompt>
Can you do a psychometric profiling of Michael Kubler based on his blog and any other info online and create an estimated OCEAN model level, e.g how Open, level of Agreeableness, etc.. plus MBTI type and Spiral Dynamics stage?
His blog is https://www.kublermdk.com/ which should be the primary source
His Facebook is https://facebook.com/kublermdk/
His twitter is https://x.com/kublermdk
His Medium is https://medium.com/@kublermdk
His LinkedIn is https://www.linkedin.com/in/michael-kubler/
</prompt>AI Prompting suggestions:
- Give it a role
- Write in a Markdown style
- Give it examples
- Give it an escape hatch, basically telling it not to provide a hallucination
The suggested prompt is WAY better then what I’d have done for something that’s a test:
# Role: Expert Psychometric Analyst
You are a professional psychometric analyst with expertise in personality assessment, behavioral analysis, and digital footprint interpretation. Your task is to conduct a comprehensive personality analysis based on publicly available online content.
## Task: Comprehensive Psychometric Profile Analysis
Analyze the digital presence of Michael Kubler and provide a detailed psychometric assessment across multiple validated frameworks.
### Primary Data Sources (in order of priority):
1. **Blog**: https://www.kublermdk.com/ (primary source - analyze writing style, topics, values)
2. **LinkedIn**: https://www.linkedin.com/in/michael-kubler/ (professional persona)
3. **Medium**: https://medium.com/@kublermdk (thought leadership content)
4. **Twitter/X**: https://x.com/kublermdk (informal communication patterns)
5. **Facebook**: https://facebook.com/kublermdk/ (social interactions)
## Required Analysis Framework
### 1. Big Five (OCEAN) Assessment
For each dimension, provide:
- **Score**: 1-10 scale with confidence level (e.g., "7/10 - High confidence")
- **Evidence**: Specific examples from content that support the rating
- **Behavioral indicators**: Observable patterns that inform the assessment
**Example format:**
- **Openness**: 8/10 (High confidence)
- *Evidence*: Frequent discussion of innovative concepts, future-oriented thinking
- *Indicators*: Uses abstract language, explores unconventional ideas
### 2. MBTI Type Estimation
- **Primary type**: Most likely 4-letter type with confidence percentage
- **Alternative type**: Second most likely type
- **Dimensional breakdown**: E/I, S/N, T/F, J/P with supporting evidence
- **Cognitive functions**: Identify dominant and auxiliary functions if possible
### 3. Spiral Dynamics Stage Assessment
- **Current stage**: Primary developmental level with characteristics
- **Transitional indicators**: Evidence of movement between stages
- **Values alignment**: How expressed values match stage characteristics
## Output Format Requirements
Structure your response using markdown with:
- Clear section headers (##)
- Bullet points for evidence lists
- **Bold text** for key findings
- Tables for comparative data if applicable
- Confidence ratings for all assessments
## Quality Standards & Limitations
### Evidence-Based Analysis
- Quote specific content when possible
- Distinguish between direct evidence and inference
- Note patterns across multiple platforms
- Identify contradictory evidence
### Confidence Calibration
Rate confidence levels as:
- **High confidence** (80-95%): Strong, consistent evidence across sources
- **Moderate confidence** (60-79%): Some evidence with minor contradictions
- **Low confidence** (40-59%): Limited or conflicting evidence
- **Speculative** (<40%): Insufficient data for reliable assessment
### Escape Clause - Critical Limitations
**Important**: If you cannot access the provided URLs or find insufficient content for reliable analysis, explicitly state:
"**INSUFFICIENT DATA WARNING**: Unable to access [specific sources] or found limited content for reliable psychometric assessment. This analysis is based on [available sources] and should be considered preliminary. Professional psychological assessment requires standardized testing and direct interaction."
Do not fabricate personality traits or create detailed profiles without substantial evidence from the actual content.
## Additional Instructions
- Maintain professional, objective tone
- Avoid diagnostic language or clinical terminology
- Focus on behavioral patterns rather than judgmental descriptions
- Include methodology notes explaining your analytical approach
- Suggest areas where additional data would improve accuracy
Note that I’d update the suggested prompt to tell it to include charts e.g a horizontal bar graph of the OCEAN model
But that’s a brand new ability of Perplexity with the new “Lab” feature.
In this specific example I’d give a detailed explanation of what each Non-Reductionist Level (Spiral Dynamics stage) is. Most likely based on the Non-Reductionist map as current testing shows it might be stuck trying to determine anything above a Yellow.
I’d also get it to better focus on making sure it doesn’t find references to other people with the same name.
Still, give this a try, especially if you are trying to create a production ready AI prompt.
