Why does this matter?

Drafting the Instructions for your Play Agents is equal parts Art and Science.

When done well, you are giving your Play Agents the perfect structure, guidance and context they need to create outreach that aligns with your tone of voice and picks the perfect combination of research and value messaging.

Getting this wrong can lead to the typical, boring AI spam which completely fails to break through the noise of generic cold outreach.

It also risks getting your email and phone numbers flagged as spam, which is a costly and frustrating issue to overcome.

A typical generic spam email: No personalization; Nothing that directly appeals to me or shows any knowledge of my situation. No appeals to my pains or goals.

Result:

Nothing compelling me to respond.
Automatically flagged by Gmail as spam.

This article will share our best practice guidance for drafting instructions for your Play Agents to stick the landing.

Following these steps, you'll ensure you're putting all your Qualification, Research, Value Proposition and Playbook content to work so your Sales Reps trust and love working with your Evergrowth Plays!

Basics: How Evergrowth Play Agents Work

Before diving into instruction-writing best practices, it helps to understand what Play Agents use when drafting outreach.

How a Play Agent Approaches drafting a Play

What Play Agents Have Access To While writing, Play Agents can call on: Your Value Proposition Account Qualification results Account Research results Contact Research results Account Plans The Play Agent Instructions	How They Use Your Instructions They combine these to write drafts: The guidelines: Tone, constraints, mandatory phrasing Intelligence: Value proposition, Account & Contact Research, etc. called upon for content The instructions: the structure you want them to interpret for their content
Hardcoded Text vs. Instruction Fields Hardcoded text: When they must use a specific phrase, guarantee compliance, or require exact wording. [replace me] instructions: When you want the Agent to interpret structure, apply research, or write original copy.	Advanced Settings If your instructions reference a specific signal, persona detail, or expertise insight, you should add conditions so the Agent only generates that section if: a) the required research is available, or b) the contact matches the intended persona/expertise type This keeps outputs aligned with context and prevents inappropriate references!

Our best practice guidance for Play Agent Instructions

Many of these notes are most useful when it comes to Email-writing Play Agents, where the text is the final output that gets shared with your prospects

(As opposed to call talk tracks, for instance, which allow for your reps to adapt and pivot on the fly if a line of messaging is not working)

However, these tips can be applied broadly to any type of writing

Avoid Overwhelming the Agent

Break your instructions into several bite-sized chunks.

Just like a human writer, an Agent is more reliable when focusing on a small set of instructions at a time. When everything is lumped together, the Agent is more likely to miss or drop instructions.

Practical tip:
Use separate sections for:

tone of voice
structure
each distinct section of a play
opening line
research/evidence/value messaging
CTA

Bad: Giant paragraph with vague instructions about the desired output

Good: Step-by-step instructions. Consistent structure coming from hard-coding combined with AI instructions.

Put Overarching Instructions Up Front

If a rule applies to the entire Play (such as "Write this in British English" or "Use a warm, confident tone") put it at the very top.

If added at the end, the Agent must revise everything it has already drafted, which increases the chance of missing corrections, or overwriting existing sections causing them to drift from their specific scope.

Use Negative Rules Sparingly and Place Them at the End

Too many “don’t do X” instructions at the beginning can get ignored as the Agent shifts focus towards drafting a coherent piece of content.

For better compliance:

Like the above tip, put your positive/active rules first (“Use X tone. Keep it value-first”),
Put "avoid"/warning rules at the end (“Avoid harsh phrasing… avoid sounding accusatory…”).

Pro Tip: This mirrors how AI Agents prioritize instructions in general

**KISS (Keep It Simple, Seriously)**

If you have a phrase or exact wording you want every rep to use:

Hardcode it
Then, surround the hardcoded section with instructions the Agent can interpret

Think of this method like modernizing an old-school template or snippet...but with smarter flexibility for an AI writer to work with.

When to Hard-code vs. Describe

For most of an email Play, you should describe what/how you want your Agent to write: tone, structure, what to mention.

For critical lines that must be consistent (e.g., disclaimers, legal language, a specific sales messaging framework), hard-coding the text directly ensures consistency.

To be an expert Play Agent instructor, you should aim to use hard-coding sparingly: overusing it will make the output rigid and hard for your Play Agent to adapt across Accounts or Verticals.

Even better is to only use them to guide your Play Agent down a predictable path, while leaving the flexibility to switch up the text based on what fits.

A perfect example is the below template where the word "with" is hardcoded to force the Agent to structure a sentence in a specific way, but the Agent has complete control

Make It Human

Agents default to very formal English.
If you want modern outreach that writes more how a human would:

Use k / m / % instead of writing out “ thousand”, “million”, or “percent”
Keep phrasing conversational
Use shorter, lighter sentences

This leads to drafts that sound more natural and less robotic.

Avoid Subjective Terms

Subjective rules (“make this shorter than the previous sentence”) are hard for an Agent to interpret consistently.

If you know what you want, be explicit:

Instead of: “shorter than above” or "not too long"
Use: “11–13 words”

This reduces the Agent’s reasoning load while making the output predictable.

Word Count Matters (A Lot)

The entire length of a play of course matters to readers.

Too long and people will lose interest. Too short and you can't really tell any compelling story and convey the richness of the context needed to use your research and value proposition.

But there's another way this applies: in how constrained your Agent is with regards to structure.

Defining word count ranges is always recommended, but this often leads to 2 common pitfalls:

1. Too strict

If you restrict the Agent to something like “<10 words” (especially when combined with hardcoded phrases) you give it almost no room to produce a natural-sounding sentence.

If previews feel awkward or clipped, increase the word budget.

2. Too loose

If outputs feel inconsistent or meandering, tighten:

sentence length
total word count
or increase your ratio of hardcoded phrases

Small adjustments dramatically improve consistency.

Example Prompt:

It seems [continue the sentence with a fact found in public data/research criteria that is independently useful to the prospect]

6-8 words: Too short. Not enough room for context or smooth phrasing.

"It seems your hiring jumped last quarter suddenly."

10-14 words: Balanced. Enough space for richer context and range for flexible phrasing.

"It seems your hiring activity picked up last quarter after that leadership change.

15-20: Too loose. Leads to meandering and 'fluff'

"It seems your hiring activity started picking up quite a bit last quarter, especially once that new leadership team settled in."

Design CTAs around a Pain + Proof combo

In your instructions, your CTA can just be a simple message with a low-barrier to entry...

However, if you want to appeal to technical buyers or really lean into value-based message, you can instead specify a pain you want the agent to reference in the CTA.

Example instruction:

[End with a short CTA that invites a conversation about reducing a pain obtained from the persona research, and mention that similar customers have achieved an improvement (e.g., ‘Some of our customers have cut this work by 60%’)

Avoid Suspicious Tone

(This is more of a Sales tip than an AI tip), using neutral observational language is less confrontational and invites the prospect to correct or agree with your assumptions

Prefer phrases like: “It seems…”, “From what I can see…”, or “It looks like…”
Instead of: “I noticed…” or “I’ve been tracking…”, which can sound creepy or overly surveillant.

In your instructions, you can even hard-code this pattern into your agent instructions, forcing your Play Agent to use it as their structure:

Example: From what I can see [add a contextual insight from enrichment criteria—connect it to a business pain or opportunity, using actual names, numbers, shorten dates, or locations]

How to clearly reference the prospect vs yourself

If you do not do this well, you may end up with your Play Agents writing outreach as if you would be the one receiving

For Example, if we at Evergrowth got this wrong, our plays would say something like the below, where our attempt to demonstrate understanding of the customer via our research actually just references our own value proposition:

To prevent this, avoid saying "you", "I", "them" when telling the Play Agent whom to reference.

You clearly need to tell the Play Agent in the 3rd person whom to reference.

When speaking about the Prospect

Use the exact term: [the prospect]

e.g. [use your knowledge of the prospect's company business model to write out a niche main benefit that the prospect's company offers]

When mentioning the Prospect, you can also go a level further and differentiate between

a) The Prospect's Company (i.e. calling out the account name, referencing the Account Research), or

b) The Prospect individual (i.e. the person receiving the outreach, where the Play Agent should reference context based on their mapped Persona or the Contact Research)

💡 Since Plays are written by the same agent that generates the Digital Twin, when a play references [the prospect], it will default to assuming you are talking about the person, not the company

When talking about Yourself

Use the exact term: [the salesperson]

e.g.

[write a single sentence that complements the pains mentioned in the previous paragraph, uses only research-sourced details about the prospect's company that the salesperson's value proposition can address]

Asking AI to translate

What to avoid doing. What to NEVER do. And what to do if only if you must.

Just adding [write this in the prospect's main spoken language] might technically work, but readers will be able to spot it a mile away.

In order to authentically preserve your copy, AI may use anachronisms, odd phrasing and overl-formal tone/conjugations that make drafts feel like advertisements, not direct human-to-human messages.

To test what we mean by this, take an email play and paste into ChatGPT (or Gemini) and say "translate this to [french]" (or pick another language)...

Then copy the result and paste it into a new chat window and say "translate this to [your language]"

It might technically cover the same bases, but even with great AI, tone and nuances of expression can be lost in translation.

Not to mention, such a simple play instruction will probably conflict with any set phrases or messaging examples you've used in your Play.

There are 3 methods of storing multi-language Plays in Evergrowth:

Best: Language variants + conditional rules (scale, consistency)
OK: Set language at the agent/instructions level using contact language, while keeping the whole play flexible and prompt-based (no hard-coded phrases)
Fallback: Translate a finished Play version

🥇 Best Practice: Draft Plays in multiple languages

The simplest way to make a play that sounds good in a specific language is to write it in this language.

Not only will you not run into the language nuance issues mentioned above, but this method will mean any hardcoded words, set-phrases & sentence-length guardrails are much less unstable when applied to multiple languages.

i.e.

Step 1) Create your Play, referencing the signal and structure you desire

Step 2) Add the language suffix at the end of the play name (e.g. Cold Email | EN)

Step 3) Translate the play instructions yourself (or with a colleague) to the desired language you also wish to be able to send it

Step 4) Add this language as a suffix (e.g. FR / DE / ES)

Step 5) For each of the Plays, add a conditional rule that it should only generate if a Contact Research agent finds the language spoken by a contact matches (see below)

💡 It's possible to enable every language variant of a play for a prospect but only have their language version generate.

Here are the steps:

Step 1) Create a "Contact Languages" Contact Research Agent

👇 Expand this section for a prompt you can copy 👇

task: discover_contact_languages
question: Does the contact publicly indicate what language(s) they speak that can be captured for segmentation and personalization?

output:
  format: semicolon_separated
  values: language_names_in_english
  multiple: true
  empty_value: n/a

rules:
  - Do not include proficiency levels in the output
  - Do not infer languages from name, nationality, location, employer, or posting behavior
  - Accept only explicit personal declarations of spoken languages
  - Normalize language names to English when necessary
  - Reject UI language selectors or interface language artifacts

process:
  step_1:
    name: linkedin_languages_primary
    actions:
      - Retrieve the contact profile using the contact_linkedin tool
      - If languages are not present in the default profile payload, explicitly retrieve the LinkedIn languages details page
      - If a LinkedIn Languages section is found, treat it as authoritative
      - Capture all listed languages regardless of proficiency
      - If languages are found at this step, skip remaining steps

  step_2:
    name: personal_site_secondary
    actions:
      - Check for a personal website, CV, or bio linked from LinkedIn
      - Scan About, Bio, Resume, or CV sections for explicit language lists or statements

  step_3:
    name: employer_and_speaker_bios
    actions:
      - Review employer team pages and leadership bios
      - Review conference, webinar, podcast, and author bios for explicit language declarations

  step_4:
    name: social_profiles_supporting
    actions:
      - Review public social profiles only for explicit self-statements of spoken languages
      - Do not infer languages from content language or posting behavior

  step_5:
    name: targeted_search
    actions:
      - Run 2 to 6 targeted public search queries combining the contact name with language-related terms
      - Use employer or title for disambiguation when needed

decision:
  success:
    output: semicolon_separated_language_names
  failure:
    output: n/a

quality_control:
  - Output must contain only the final value with no additional text
  - Output must strictly match the defined format to prevent downstream data errors

Step 2) For your Play/s, within Advanced Settings, add a Conditional Rule:

If [Contact Languages] Contains [Language of the play]

Step 3) Now you can enable all language variants of your Play within your workflows - only those which are matched to the prospect will be generated when that workflow is run, the rest will be skipped!

We appreciate though that this may lead to more upkeep over time - making updates to the same play multiple times (for each language you store the play in) is obviously more work than doing it once...

🥈 OK Practice: Keep plays flexible and set the language near the start

If you can start your Play prompt to begin writing in the language of the prospect (using the same Contact Research Agent above works well for this), that will guarantee what follows is written with the prospect's language in mind, not written and then rewritten to translate retroactively.

To do this add (or amend) a writing language instruction at the start of your play, like:

[Writing language: Write this message in the prospect's first listed spoken language, found using the "Contact Language" research]

In this scenarios, hardcoding set phrases should be avoided since it can confuse the Agent whose main goal is to satisfy you.

If you tell an agent:

"[start the email with 'hello my name is Tom'. Write the email in Swedish]", you are technically giving conflicting instructions.

Generally, Play Agents will keep the hardcoded 'hello my name is Tom' since this is the most unambiguous part of a play to write and can never be 'wrong' based on how they've interpreted instructions.

This will obviously cause issues and lead to undesirable play drafts.

🥉 Last Resort: Set firm instructions to translate the entire play

If you are still hard-coding set phrases, technically you can begin the Play instructions with:

[Writing language: strictly write the message in the prospect's first listed spoken language, found using the "Main Language" contact research. Override any hardcoded wordings or set phrases in order to facilitate this translation. Prioritize this instruction above all others.]

However, you should ensure that tone-of-voice guardrails are reiterated in generalised terms, such as [do not use overly formal or clinical language], [keep the tone light, positive and ensure there is approachable confidence in what is said to the prospect]

Since your want the translation to take precedence, this should also be moved to the top of the play (as covered in the above section ⬆️ Keep overarching rules up front)

Email Example: Signal–Value–Evidence

Email Example

Below is all the text from our "Signal-Value-Evidence" Expert Play Agent template

We'll break down why we've drafted the instructions the way we have:

Preface: Why This Works

The email instructions are broken into smaller sections ("Avoid Overwhelming the Agent" principle)
Style/tone rules are clear and placed upfront
Each section contains only a few instructions
Rules are placed in order: overarching → structure → specific line-level constraints → negative rules last
Word limits are explicit

Intro Specifications

This section defines the non-negotiable guardrails.
By locking in tone and language first, you remove ambiguity and ensure every downstream instruction builds on a consistent foundation.

Opening Line + Signal

This is where you shape the email's first impression.
The instructions create enough structure to guarantee clarity and allow for personalization, giving the Agent room to adapt to each prospect's research insights.

We are very prescriptive about elements to avoid and the tone to strike in this opening line for a simple reason: people judge cold emails very quickly.

If the first line feels AI-generated, or irrelevant, it's game over!

Evidence

This block reinforces credibility.
It guides the Agent to weave in proof that supports the opening message without scripting the exact example, building up the relevance and authenticity.

CTA

The final instruction aims to close with purpose.

It guides the Agent toward a confident, low-pressure action that feels natural and aligned with the prospect’s context.

Additional Reading 📚

📖 Plays - Get a complete overview of how to configure and use Play Agents

📖 Play Templates - Learn how to add and use Play Agent Templates (including how to find our tried-&-tested Expert Play Templates library)

📖 Play Testing - Learn the simplest way to test and refine your Play Agents, like running a mini workshop with a new colleague!

Article Coming Soon ⏳

📖 Best Practice Leveraging Plays - Once your Play Agents draft world-class copy, learn the best ways to put them to work for maximum efficiency and autonomy.

⚡️ Play Drafting Best Practices