Prompt Anatomy
The 5-part structure AutoClaygent generates for every prompt
The 3-Task Rule
Every Claygent prompt should have at most 3 distinct tasks. More than that, and quality drops significantly.
"Find the platforms they use, the owner's name, the business model, their tech stack, and whether they serve B2B or B2C" — This is 5 tasks. Split it into 2-3 Claygents instead.
AutoClaygent automatically enforces the 3-task rule. When you describe what you need, it will recommend splitting complex requests into multiple focused Claygents—and explain why.
Anatomy of a Good Prompt
Every production-ready Claygent prompt has these 5 parts:
1. Input Context
Tell Claygent what data it's working with. Use Clay's variable syntax:
Given the company domain: {{domain}}2. Clear Goal
One sentence describing what you want to find:
Goal: Detect what SaaS platforms this company uses by analyzing their portal URLs.3. Step-by-Step Instructions with Decision Logic
Numbered steps with specific sources and explicit decision trees:
Research steps:
1. Look for customer-facing portals:
- Check {{domain}}/login, {{domain}}/portal, {{domain}}/app
- Look for "Client Login", "Patient Portal" in footer
2. When you find a portal link, analyze the URL pattern:
**Subdomain Patterns (highest confidence):**
- {company}.salesforce.com → Salesforce
- {company}.hubspot.com → HubSpot
- {company}.janeapp.com → Jane AppNotice the explicit URL pattern mapping. This is what separates 70% accuracy prompts from 95% accuracy prompts. Don't say "find what CRM they use" — tell the model exactly how to detect it.
4. Fallback Instructions
What to do when the primary approach fails:
If no portal found on website:
- Check if login redirects to a third-party domain
- Look for embedded widgets (chat, booking, support)5. JSON Output Specification with Evidence
Explicit structure with field descriptions and evidence requirements:
Output as JSON:
{
"platforms_detected": [
{
"platform_name": "Name of platform",
"platform_category": "CRM | Support | Scheduling | EHR",
"detection_method": "subdomain | redirect | widget",
"evidence_url": "URL that revealed this"
}
],
"portal_url": "Customer portal URL if found" | null,
"confidence": "high" | "medium" | "low"
}Bad vs. Good: A Comparison
Bad Prompt (Score: ~4.0)
Find what CRM or booking platform this company uses.Problems: No context, no detection method, no output format, no fallback, no confidence levels.
Good Prompt (Score: ~8.5)
Given the company domain: {{domain}}
Goal: Detect what SaaS platforms this company uses by analyzing their portal URLs.
Research steps:
1. Look for customer-facing portals:
- Check {{domain}}/login, {{domain}}/portal, {{domain}}/app
- Look for "Client Login", "Patient Portal", "Customer Portal" in footer
- Check navigation menu for portal links
2. When you find a portal link, analyze the URL pattern:
**Subdomain Patterns (highest confidence):**
- {company}.salesforce.com → Salesforce
- {company}.hubspot.com → HubSpot
- {company}.zendesk.com → Zendesk
- {company}.janeapp.com → Jane App
- {company}.mindbody.io → Mindbody
3. Check for scheduling/booking tools:
- Look for "Book Now", "Schedule", "Appointments" buttons
- These often reveal Calendly, Acuity, Cal.com, etc.
If no portal found on website:
- Check if login redirects to a third-party domain
- Look for embedded widgets (chat, booking, support)
Output as JSON:
{
"platforms_detected": [
{
"platform_name": "Name of platform",
"platform_category": "CRM | Support | Scheduling | EHR",
"detection_method": "subdomain | redirect | widget",
"evidence_url": "URL that revealed this"
}
],
"portal_url": "Customer portal URL if found" | null,
"confidence": "high" | "medium" | "low"
}
IMPORTANT:
- URL patterns are 95%+ accurate — trust subdomains over text
- If no platform detected, return empty array (don't guess)
- "high" confidence = subdomain pattern match
- "medium" confidence = redirect or widget detection
- "low" confidence = text mentions only7 Anti-Patterns to Avoid
- "Search the web for..." — Be specific about sources (URL patterns, navigation, footer)
- More than 3 tasks — Split into multiple Claygents
- No detection method — Explain HOW to find it (subdomain patterns, redirects)
- No JSON output format — Specify exact structure with required fields
- No evidence requirement — Always ask for the URL/source that proved the finding
- Asking for guesses — Instruct to return null or empty when uncertain
- No confidence gradients — Define what makes high vs medium vs low confidence
The best prompts include the detection method accuracy. "URL subdomain patterns are 95%+ accurate" tells the model to prioritize that signal over text scraping (70% accurate).
Try It Yourself
Type a natural language goal and see how it transforms into a structured prompt:
Click a pattern to see how AutoClaygent transforms it into a structured prompt:
Select a pattern to see the structured prompt...
AutoClaygent generates prompts like these automatically based on your goals.
AutoClaygent generates all 5 parts automatically—Input Context, Goal, Steps with Decision Trees, Fallbacks, and JSON Output—based on your plain English description. It also generates a valid JSON schema and iteratively improves the prompt until it scores 8.0+.
Key Takeaways
- Limit each Claygent to 3 or fewer distinct tasks
- Every prompt needs: Input, Goal, Steps with Decision Logic, Fallback, JSON Output
- Include explicit detection methods — don't just say "find X"
- Require evidence URLs for every finding
- Define confidence levels with specific criteria
- Instruct to return null/empty when data can't be verified