Docs
How can we help? 👋
All Categories
Get Smart
Prompt Writing Guide for Outbound Messaging

Prompt Writing Guide for Outbound Messaging

Give LLMs the info and instructions to write good copy

Last updated on December 26, 2025

Let’s start with what’s actually true about outbound:

  1. Poorly researched outbound doesn’t work
  1. Good outbound requires:
      • Strong segmentation
      • Clear observations
      • Correct persona → message matching
      • A differentiated product
      • A compelling reason to reply
  1. AI can help with all of the above if you tell it what to do

Here’s the problem:

  • Most BDRs and AEs don’t know how to prompt AI
  • Most RevOps, GTM engineers, and marketers who do know how to prompt don’t know how to outbound

This guide exists to bridge that gap.

You don’t need to be technical. You do need to be precise.

Understanding Good Outbound

IMPORTANT‼️

Before you touch prompting, you must understand what good outbound copy actually looks like.

Prompting doesn’t fix bad fundamentals, it just scales them.

🌐

Reference:

If you skip this, you’ll get AI-written spam faster.

What an LLM Needs to Write Good Copy

If you ask ChatGPT (or any AI) to “write an outbound message,” the result will be bad by default.

AI only performs well when you give it clear context and constraints.

At a minimum, it needs five things:

  1. Full context on your company
    1. What your product/services does
    2. Personas you sell to
    3. Problems they have which you solve
    4. How you uniquely solve those problems
    5. Existing customers and case studies to name drop
  1. Full context on the lead
    1. Name, job title, company, company website
    2. An observation about them that is making you reach out
  1. Web Access
    1. So it can lookup the lead, their LinkedIn, their website, their blogs, case studies, news, podcasts, 10-Ks, etc
  1. Clear Instructions
    1. What signals to look for and research to do and how to prioritize them
    2. Pain points for each persona to focus on and how to prioritize them
    3. How to pitch your product/service
    4. Acceptable CTAs to use
  1. Samples:
    1. Fully formed examples of what you’re shooting for

How Prompting Works in Letterdrop (At a High Level)

Letterdrop already handles a lot of the hard parts for you.

We automatically provide:

  • Details about your company
  • Details about the lead
  • A default prompt you can edit and improve

Helpful variables you’ll see:

  • {{your_company_description}} → what you sell, who you sell to, how you’re different
  • {{lead_linkedin_full_profile}} → full LinkedIn profile (about section, job history, etc.)

You don’t need to write everything from scratch.

How Our Prompt Is Structured (Plain English)

From top to bottom, the prompt tells the AI:

  1. Who you’re reaching out to
  1. What your company does
  1. What signals to research about the lead
  1. Which customers it’s allowed to reference
  1. How to pitch your product
  1. Which pain areas to focus on
  1. Which org function the lead belongs to (for referrals)
  1. Exactly how to use all of the above
  1. What a good outbound message looks like
  1. Full examples to copy from

Anything in ALL CAPS is meant for you to customize.

We’ll scrape your website and generate a first draft automatically—you just refine it.

I'm a BDR at {{your_company_name}}. I want to reach out to:

Name: {{lead_full_name}}
Company: {{lead_company_name}} ({{lead_company_website}})
Title: {{lead_job_title}}
LinkedIn: {{lead_linkedin_full_profile}}

Here is some information about our company:
<about_us>
{{your_company_description}}
</about_us>

Research some relevant signals:
<signals>
(LIST YOUR SIGNALS HERE)
</signals>

These are existing our company customers for references:
<existing_customers>
(LIST YOUR CASE STUDY CUSTOMERS HERE)
</existing_customers>

<product_pitch>
(PUT YOUR PRODUCT PITCH HERE. SEGMENT BY PERSONA IF NEEDED)
</product_pitch>

Possible pain areas:
<pain_areas>
(LIST THE PAIN AREAS YOU SOLVE HERE. SEGMENT BY PERSONA IF NEEDED)
</pain_areas>

Relevant functions for routing:
<relevant_functions>
(NAME THE ORG FUNCTION YOU SELL INTO SO WE CAN LOOK FOR OTHERS TO ASK FOR A REFERRAL TO)
</relevant_functions>

If {{lead_job_title}} is C-suite/VP, identify the most senior person reporting into them from <relevant_functions>. Output two variables: [backup_first_name] and [backup_title]. If no credible match, omit the handoff line.

Instructions
1. Research the company on the web using <signals> - look through recent news, their website, job postings, blogs, articles, podcasts, interviews, etc to find these signals
2. List observations and stack rank by urgency vs <pain areas>.
3. Pick a similar customer from <existing_customers>.
4. If the individual is C-suite or VP, look for the BACKUP in <relevant_functions> and populate [backup_first_name], [backup_title].

Style rules for a super-casual DM
- Max 85 words total.
- No sign-off
- Don't mention job titles.
- Allowed punctuation: . - ? , 
- Ban fluff words: innovative, cutting-edge, thrilled, delighted, synergy, leverage.
- No em dashes.

OPPS format (keep it tight)
<format>
Hi {{lead_first_name}}! {Make an observation by layering signals from above to guess their priorities or internal pain points}
{Show them how you can solve for that pain}
{Use a short, learning-focused CTA. For example: "Open to a quick chat to compare how others are tackling this?"}
{optional: if C-suite/VP and backup known: "Or sync with [backup_first_name], your [backup_title]?"}
</format>
 

How the Final Message Is Assembled

Anything inside {} tells the AI how to think, not what to copy verbatim.

Example structure:

Hi {{lead_first_name}}!

{Open with a specific observation based on research}

{Tie that observation to a likely priority or pain}

{Explain how we solve it, using a relevant customer example}

{End with a learning-focused CTA — not a demo pitch}
 

Using Tags to Keep Things Organized

When giving AI lists (customers, signals, pains), always wrap them in tags.

Example:

<customers>
- Acme: Manufacturing
- Northwell: Healthcare
- PetroChina: Energy
</customers>

ChatGPT can help

Paste the above prompt into ChatGPT and give it instructions on how to modify it.

Less is more

If there’s any information your prompt won’t use, get rid of it. Too much information confuses the LLM.

Picking Your Model

You want to pick the cheapest model that gets you the results you want.

If you instruct your prompt to do any net new research to find information you didn’t already provide it, you also need to give it access to a search engine.

So for example, once you write a prompt, test run it on a handful of leads with an expensive model like OpenAI’s GPT-5.1 and make sure you can get the results you want.

Then drop down to cheaper models and repeat the test run on the same leads so you can see if there is a noticeable quality drop off. If not, keep dropping to cheaper models and trying again. Stop once you’re no longer happy with the outputs.

Notion image
 

Final Takeaway

Prompting is not magic.

Good prompting is:

  • Clear context
  • Strong constraints
  • Real examples
  • A solid outbound foundation

If your inputs are good, AI will outperform your average rep.

If your inputs are bad, it will just fail faster.

That’s the tradeoff.