prompt-engineering-rules

AI Prompt Engineering Rules

You are applying Eric Nowoslawski's five rules for efficient AI prompts, critical for running AI at scale in Clay workflows without wasting credits or getting bad outputs.

Context

These rules come from Eric Nowoslawski of Growth Engine X, who runs enrichment prompts at very high volume.

At scale, prompt quality directly impacts:

• Credit costs (bad prompts = wasted runs)
• Output quality (unclear prompts = unusable results)
• Filtering effort (no safeguards = manual cleanup)

The 5 Rules

Rule 1: 10-Minute Manual Research Rule

Only automate what you'd actually look up manually.

Principle: If you wouldn't spend 10 minutes researching this manually, don't ask AI to do it.

Why It Matters:

• AI research has costs (credits, latency)
• If the data isn't worth manual effort, it's not worth automated effort
• Prevents "nice to have" data bloat

Application:

Before adding an AI enrichment, ask:
- Would I manually research this for a prospect?
- Would this data change my approach?
- Is this worth 10 minutes per company?

If no → Don't automate it
If yes → Proceed to build the prompt

Rule 2: One Task Per Prompt

Have AI do one thing at a time. Break out enrichment and classification steps.

Principle: Each prompt should accomplish exactly one task.

Why It Matters:

• Multi-task prompts have lower accuracy
• Harder to debug when something fails
• Can't conditionally run parts

Bad (Multi-task):

Find the company's tech stack, determine if they're PLG,
score their ICP fit, and write a personalized opening line.

Good (Single-task):

Prompt 1: Identify technologies used (tech_stack)
Prompt 2: Is this a PLG company? (plg_status) [uses tech_stack]
Prompt 3: Calculate ICP score (score) [uses tech_stack, plg_status]
Prompt 4: Write opening line (opening) [uses score, relevant data]

Rule 3: Safeguards in Prompts

Output 'purple' for missing data. Reliably filter errors later.

Principle: Define explicit outputs for when data is missing or uncertain.

Why It Matters:

• AI will hallucinate if not told to abstain
• Consistent error outputs enable filtering
• Easier to spot-check quality

Implementation:

[Your prompt here]

IMPORTANT:
- If the information cannot be found, return exactly: "NOT_FOUND"
- If you are uncertain, return exactly: "UNCERTAIN"
- Do not guess or make assumptions
- Do not return empty strings

Why "Purple": Any distinctive, filterable string works. Eric uses "purple" because:

• It's visually obvious in data review
• Easy to filter in Clay/spreadsheets
• Never appears in real data

Rule 4: Examples in Prompts

Give AI plenty of examples. Use metaprompter ('what do you need to improve this prompt?')

Principle: Show, don't just tell. Include examples of desired output.

Why It Matters:

• Examples calibrate AI understanding
• Reduces ambiguity
• Improves consistency across runs

Implementation:

Determine if this company is B2B or B2C based on their website.

Examples:
- Salesforce.com → B2B (sells to businesses)
- Nike.com → B2C (sells to consumers)
- Shopify.com → B2B (sells to businesses who sell to consumers)
- Amazon.com → B2C (primarily consumer marketplace)

Now analyze: {{company_website}}
Return: B2B, B2C, or BOTH

Metaprompter Technique:

Here's my prompt: [YOUR PROMPT]

What additional context, examples, or clarifications would
help you produce better results for this task?

Rule 5: Chain of Thought

For complex prompts, ask AI to explain its reasoning before answering to improve accuracy.

Principle: Have AI show its work before giving the final answer.

Why It Matters:

• Reasoning improves accuracy
• Makes errors debuggable
• Better for complex classifications

Implementation:

Determine if this company would benefit from our solution.

Company: {{company_name}}
Data: {{enriched_data}}

First, explain your reasoning:
- What does this company do?
- What signals indicate they might need [solution type]?
- What signals indicate they might NOT need it?

Then provide your conclusion:
- FIT: High / Medium / Low
- KEY_REASON: [One sentence]

Putting It All Together

Well-Engineered Prompt:

TASK: Determine if {{company_name}} is a PLG (product-led growth) company.

CONTEXT: Check their website at {{company_website}} for self-serve signup,
free trials, or freemium tiers.

EXAMPLES:
- Notion.so → PLG (free tier, self-serve signup)
- Salesforce.com → NOT_PLG (contact sales required)
- HubSpot.com → HYBRID (free tools + sales motion)

REASONING: Before answering, note:
1. Is there a "Sign up free" or "Start trial" button?
2. Is pricing publicly available?
3. Can you access the product without talking to sales?

OUTPUT:
- Return: PLG, NOT_PLG, or HYBRID
- If website is unavailable or unclear, return: UNABLE_TO_DETERMINE

Anti-Patterns to Avoid

Anti-Pattern	Problem	Fix
Vague prompts	AI guesses intent	Be specific
Multi-task prompts	Lower accuracy	One task per prompt
No examples	Inconsistent output	Add 3-5 examples
No safeguards	Hallucinations	Define fallback outputs
No reasoning	Mysterious errors	Add chain of thought

Related Skills

• /metaprompter - Prompt improvement technique
• /data-point-research - Custom signals to detect
• /icp-scoring-dynamic - Complex scoring prompts
• /plg-company-detection - Example well-engineered prompt

Credits

The five rules are taught by Eric Nowoslawski (Growth Engine X). The explanations and examples here are ours.

---

Examples are illustrative. Company names, prices and numbers in them are placeholders or may be out of date, so check the live source before you rely on any detail.