Skip to main content

Refining your Play Agents

Test & refine your AI Play Instructions like an AI Expert

Updated over 2 weeks ago

Just like working with a new colleague, the likelihood that your Play Copywriting Agents will take your instructions and draft perfect outreach on the first attempt is slim.

In this article we'll show you how to run your own solo "workshop" to refine your instructions, preview and iterate until you get outreach which is consistent, compelling and ready for real prospects!

💡 Pro Tip

Think of this as teaching another version of yourself to do your writing for you

Once you've got this perfect, it'll be like you've got a new personal assistant, reading all the Account & Contact research, then drafting all of your outreach just like you would

If you haven't created any Plays yet - check out this article for a more thorough guide before you're ready to come back and start your Play-refining workshop


Setting up

While not overly complex, we recommend having a simple split-screen (or dual screen, if possible) view so you can both edit and test without lots of tab-switching or searching around for the Evergrowth screen you need.

This keeps the workflow much simpler and faster!

Screen 1 - Play List

This screen will be where you edit & save your Play Instructions.

If you are editing a Play which is currently marked as "Draft" (or another inactive status), you will have to amend the default status filter at the top of your list

Get here by going to Playbooks ➡️ Plays

Screen 2 - Previews

This screen will be where you launch tests and review the results

Get here by going to Playbooks ➡️ Plays, then switch to the Previews tab

Split-Screening

If you have 2 screens: great, you can skip this step and open the above 2 Evergrowth pages on a screen each!

Otherwise, you can easily arrange each window to take half of your screen by pressing and holding the Full screen icon:

The above instructions on for Mac

On Windows, use Win + ← / → to snap windows like this


Testing

To test your Plays, pick a handful of real contacts so you can see how the outreach would look in the wild.

This means you should have a handful of ICP-Yes Accounts fully qualified, plus IBP-Yes Contacts with research generated (and Digital Twins if you use them).

Step 1) Go to Playbook ➡️ Plays

Step 2) Click the Edit icon to update your Play's Instructions, then Save

Step 3) Click the Generate Preview icon

Step 4) Select a mix of contacts where you want to test different signals, personas and verticals

Don't worry - Play Previews do not count against your Credit usage totals, so you can test as much as you need!

Step 5) Switch to the Previews view & review the outputs

To see all the variants of your previews side by side, you can select multiple rows then click Compare

*We recommend selecting a mix of contacts when generating Previews

This is so you can check the Play Agent writes appropriately for different personas & industries, adapting which benefits & pains to reference for each!

Repeat Steps 1 - 5 until your play consistently picks the appropriate research, the appropriate sections from your Value Proposition, and the appropriate tone


Example of a Mini Workshop

In the below (condensed) video, James generates a new Play then tests and refines.

Get a glimpse into the structure that a self-run workshop should take, amending instructions, guardrails, guidance & tone of voice until the new Play writes how you would want your outreach to look:

Want to see the full version?

In the below recording, James explores all his decisions and updates to the play in more detail

With each update, he explains why he makes the amendments and what results he's looking to achieve with each iteration


"If -> Then" Fixes 🔧

Below are a series of changes to try based on what you’re seeing in Play previews.

Play not using the right customer stories / messaging?

Add explicit story selection rules

Tighten the way your stories are written in the Training Center.

In your Play instructions, tell the agent which story to use when (persona + pain + signal).

  • If it still picks randomly, it usually means your stories aren’t clearly differentiated. Update each story so it clearly states:

    • Who it was for (persona + segment)

    • The pain / goal (what they were trying to improve)

    • The outcome (what changed)

    • The evidence (metric, timeframe, or concrete before/after)

Your outreach mentions “ROI” but it’s vague (“drive revenue”, “save time”)

Require pain‑linked evidence

(and update your Value Prop's customer stories if the evidence isn’t there)

  • Add a rule like: “[If you mention ROI, you must tie it to a specific pain/KPI already stated in the email.]”

  • Add a guardrail: “Only use quantified results if they exist in the Customer Stories. Never invent metrics.”

If the agent can’t comply, that’s a signal your Training Center stories need a rewrite to include measurable outcomes (or at least a concrete “before → after” change)

The Play’s “proof” line feels disconnected from the prospect or feels chosen at random

Rebuild proof as “pain → outcome → evidence” (not just “logo & claim”).

  • In the Training Center, rewrite proof snippets so they don’t just say what happened, but why it mattered:

    • Average: “Reduced manual work by 30%.”

    • Better: “Reduced manual work by 30% so the team could hit [reference the goal from the customer story] with the same headcount.”

  • In your prompt, add: “Evidence must support the value claim you just made. No proof‑dumping.”

Output sounds like a product brochure / feature dump

Force an outside‑in structure (signal → pain/goal → value → evidence).

  • Add a simple structure requirement (and keep it consistent across previews):

    • In line 1: Pick a fixed number of relevant signals (1-2) to reference

      • Ensure there is fallback of an evergreen observation if no signal

    • In line 2: Expand on whey these matter for this persona (pain/goal)

    • In line 3: Value + proof (one sentence)

    • In line 4: CTA (low‑friction)

  • If the agent still defaults to features, your value prop content may be written too “inside‑out.” Rewrite it so it leads with outcomes and buyer goals, not capabilities.

Personalization is generic even though research exists

Force the agent to use exactly one strong signal, early.

This will ensue the key personalisation & pain/challenge you've uncovered garbs attention early (which is key for people's attention span!)

  • Add: “Use exactly 1 account signal in the first 2 lines. Don’t list multiple.”

  • Add: “If no strong signal exists, write an evergreen opener (don’t guess).”

It’s hallucinating signals or overstating what’s true

Add firm evidence requirements + a fallback paths.

  • Add: “[Only reference signals that appear in the provided research context]”

  • Add: “If research is thin, do not speculate — use an evergreen opener and move on.”

Tone feels “creepy” (“I noticed you…”, “I’ve been tracking…”)

Stipulate the need for neutral language.

  • Add: “[Use neutral phrasing: ‘It looks like… / It seems to me… / From what I can tell…’]”

  • Add: “Never imply monitoring, tracking, or inside access.”

Phrasing is too obviously AI / too polished

Hardcode a few “anchor” lines and keep your “do not” list concise.

  • Provide 1–2 example opener styles and 1 CTA style you like.

  • Add any 'banned phrases'

    • Don’t overwhelm the agent with a giant negative list, though

    • If you have a large list of approved/banned phrasing, consider adding it as a section to your value prop and instructing the Play Agent to conform to this section.

Outputs are inconsistent (sometimes prompt is followed perfectly, sometimes off)

Simplify + move “global rules” to the top.

  • Put your overarching constraints up front (tone, structure, length).

  • Put negative rules at the end.

  • If you keep adding rules and it gets worse, it’s often because instructions are fighting each other; organise your rules, then simplify where possible.

    • Don't just keep piling on more instructions to fix the bad ones.

Email is too long / rambly

Replace any subjective terms “be concise” with hard constraints.

  • Add: “[Max 110 words.” / "4 bullets points" / "12 -16 words]"

If it becomes too stiff, try loosening one variable at a time (word count or paragraphing structure) - don’t change everything at once!

CTA is weak or generic (“open to a quick chat?”) → design the CTA around a pain + proof combo.

  • Add a CTA which links to the email and offer light lift for the prospect: “[CTA should reference the pain + one proof point, then ask for a easy-to-action next step (e.g. 10–15 minutes).]”

  • Example instruction where a prospect will be offered value for their interaction: “[Ask one simple question that offers to share how teams solved [pain mentioned earlier in the message], backed by [reference evidence from customer stories or other proof found in the value proposition]].”

💡 Pro tip

If you keep tweaking Play instructions but the agent still keep repeating the exact same messaging mistakes, the source of the issue may be yoru Training Center content (value proposition, customer stories, approved messaging), not the prompt!


Editing Plays in Contact Records

It is possible to edit Plays directly from within a Contact record.

You may find this particularly useful for brand new Plays and Campaigns where the real-world results need updating on the fly

We allow this so you can make quick edits without having to:

➡️ Leave the Contact record

➡️ Head to to the Playbooks area

➡️ Find your play

➡️ Edit your play

➡️ then return to your contact to regenerate the updated Play

Step 1) Within a Contact record where a play has generated, click the edit icon

Step 2) Make your edits to the Instructions

Step 3) Click Save

NB. You have now amended the Play! Anywhere else this play is launched in future will use your new, updated instructions

Step 4) You can now click the Regenerate button to instantly relaunch the Play for your current Contact


📚 Additional Reading

📖 Play Agent Snippets - Access our bank of "snippet" instructions you can copy & paste into your plays to instantly achieve consistent, quality output, one play section at a time

📖 Play Drafting Best Practices - Learn all the pro tips we've accumulated from the thousands of plays we've written at Evergrowth so you can prompt AI like an expert

Coming Soon 👀

📖 Play Agent Application Best Practices - Once you've finished refining your Play, learn about all the best ways to put it into action and start executing your new intelligent outreach

Did this answer your question?