Prompt tools effectively

When you give an agent access to tools, the quality of your prompt determines whether the agent uses those tools correctly. A well-crafted prompt tells the agent exactly when to use each tool, what to look for, and how to format the results.

Start with the Tool Reference

Before writing your prompt, read the reference documentation for the tool you're using. Each tool in the individual tools reference includes:

Parameters: What inputs the tool accepts
Returns: What data the tool provides back
Common Use Cases: Practical applications

Understanding what a tool can and can't do helps you write prompts that work with its capabilities rather than against them.

For example, the SimilarWeb tool requires just a domain (like example.com), not a full URL. Knowing this prevents errors.

The Anatomy of a Good Tool Prompt

Effective tool prompts share a consistent structure. Here's the pattern used by experienced Cotera users:

1. Define the Task

Start by explaining what the agent should accomplish and which tools to use. Reference tools with the @ symbol:

Use @Google Search to find recent news articles about the company I give you.

2. Specify the Input

Tell the agent what information it will receive:

The user will provide a company name and optional location.

Example: "Stripe" or "Blue Bottle Coffee in San Francisco"

3. Explain What to Search For

Be specific about what information matters:

From the search results, look for:
- Recent funding announcements
- Product launches in the last 6 months
- Executive hires or departures
- Partnership announcements

4. Define a Search Strategy

For multi-tool workflows, explain the sequence:

1. Search Google for "[Company name] news 2024"
2. If you find their website, use @Scrape Website to get their About page
3. Search for "[Company name] funding" to find investment news

5. Specify What Counts as Valid

Help the agent filter results:

Only include results from the last 12 months. Ignore:
- Job postings
- User reviews
- Forum discussions

6. Structure the Output

Define exactly how you want results formatted:

Return your findings as:
- Company: [name]
- Latest News: [1-2 sentence summary]
- Key Insight: [most important finding]
- Source: [URL]

Example: A Complete Tool Prompt

Here's a full prompt that chains multiple tools together:

## Task

Use @Google Maps/Search Locations to find a local business, then use
@Cotera Site Scraper/Scrape Website to find their contact email from
their website.

## Input

The user will provide a business name and city.

Example: "Blue Bottle Coffee in San Francisco"

## What to Search For

From Google Maps:
- Business phone number
- Website URL
- Full address

From their website:
- Email addresses (check Contact page and footer)
- Social media links

## Search Strategy

1. Search Google Maps for the business with location
2. Get their website URL from the listing
3. Scrape their website's contact page
4. If no contact page, scrape the homepage footer

## What Counts as Valid

- Only report contact info you actually find—never guess
- If no email exists, say "No email found on website"
- Prefer general contact emails over personal ones

## Output

**Business:** [name]
**Address:** [full address]
**Phone:** [number]
**Email:** [email or "Not found"]
**Website:** [URL]

Common Mistakes to Avoid

Being Too Vague

❌ "Search for information about the company"

✅ "Search for the company's recent funding rounds, product launches,
   and executive changes from the past 12 months"

Not Specifying Output Format

❌ "Tell me what you find"

✅ "Return a JSON object with fields: company_name, funding_amount,
   funding_date, lead_investor"

Forgetting to Handle Edge Cases

❌ "Get their email from the website"

✅ "Get their email from the website. If no email is found,
   return 'Not available' and note where you looked"

Using Wrong Input Formats

❌ "Search SimilarWeb for https://www.example.com/about"

✅ "Search SimilarWeb for example.com"

Check the tool reference to understand what input format each tool expects.

Tips for Multi-Tool Prompts

When chaining tools together:

Order matters: Tell the agent which tool to use first and why
Pass data explicitly: Explain how output from one tool feeds into the next
Set fallbacks: Define what to do if a tool returns no results

First, use @Apollo/Search People to find the prospect's LinkedIn URL.
Then use @LinkedIn/Get Profile to get their work history.

If Apollo doesn't find them, search @Google Search for
"[Name] [Company] LinkedIn" instead.

Testing Your Prompts

Before running a tool prompt across your dataset:

Test in chat first: Try the prompt with a few sample inputs
Check tool calls: Verify the agent is calling tools with correct parameters
Review edge cases: Test inputs that might fail (uncommon names, missing data)
Refine and repeat: Adjust your prompt based on what you observe