Streamlining Your Comms Part 2: Building Your Content Assistant GPT

Introduction

Welcome back to the guide on streamlining your business communications!

In Part 1, we established your brand's tone of voice (TOV). So, if you're reading this, you should have a Tone of Voice Guidelines doc that closely mimics these sample guidelines for Trailthreads Socks Co and a base foundation of sample copy examples to train your GPT with.

Now that you've laid down all the hard groundwork, it's time to bring it all together; configuring a GPT to automate the creation and quality control of your communications.

A Quick Refresher On Why You Might Want To Do This:

Some of the immediate benefits of training a GPT in your TOV include: saving 95%+ of your / your team's time and labour by (i) rapidly creating content and (ii) carrying out quality control audits, all the while (iii) minimising errors and (iv) enabling you to easily establish and easily maintain a consistent and engaging brand voice across all mediums.

If you want to take it to the next level, your TOV tool can be leveraged for more sophisticated use cases such as streamlining sales personalisation and automating customer service. Case in point, customer service powerhouse Intercom claim that their chatbot solution provides your customers with 24/7 support and is capable of revolving 50% of your daily support volume in real time and can boost your operational efficiencies by 31%.

A Quick Refresher on What We're Building Today

1. A Rapid Content Creation Tool

2. A Rapid Content Audit / Quality Assurance (QA) Tool

What We're Actually Doing / How We're Doing It

Today, I’ll be demonstrating how we can use this TOV guideline document to train a new GPT. Assuming you’ve followed the same basic structure for your TOV guidelines as we walked through in part 1, you should be able to use the same, or very similar prompts. This guide is broken down into two key phases:

Phase 1: Initial GPT Configuration:

• Configure the GPT with basic understanding of your brand’s voice, vocabulary, and style guidelines.
  • Configure the content creation features
  • Configure the QA features
• Feed the GPT with prepared inputs, emphasising the distinction between preferred and non-preferred examples.

Phase 2: Guided Testing & Training Sessions:

• Conduct interactive sessions where you input specific scenarios and guide the GPT on generating appropriate responses.
• Correct and explain any deviations to reinforce the learning.

Important Note: Your GPT Builder Journey Will Deviate from This Guide

I am sharing the journey I was steered through whilst configuring my GPT. The flow your GPT builder might guide you through is likely to deviate from this. In many different words, the GPT is going to ask you to provide the following details:

1. Role definition (i.e. what kind of service the GPT is providing)
2. Goals and objectives of the GPT (i.e. problems it is solving and/or targets you're pursuing)
3. Constraints (i.e. the parameters bounding its responses - many of which are predefined in the TOV guidelines)
4. GPT name and image
5. The GPT’s Tone (when interacting with users)
6. User Interactions (prompting flows and strategies). Today we’ll program it to:
   1. Ask key clarifying questions before responding, and;
   2. Creating a standardised QA Audit process and output.

Don’t worry if your GPT Builder guides journey deviates from the guide below!

So long as you get all of the above details across at some point (even after the GPT is initially configured), you will be able to update and refine your GPT’s model later. Case in point, the GPT builder may ask you to define multiple of the above in a single go:

Hot Tip: Use ChatGPT to Help You Configure Your GPT

Be prepared to go into this using two windows for ChatGPT, including:

1. Your GPT Builder, and;
2. A support / guidance Thread to help steer you through any problems you might encounter and optimise your prompts. After all, nobody knows how the GPT tool is built better than itself.

Sample Prompt for your Support/Guidance Thread:

I need your help configuring my GPT.

I want to use it to create and refine content for me in my predefined tone of voice (TOV), and also be capable of doing quality control audits on content to make sure any and all copy aligns with my brand's TOV guidelines attached.

Please review the TOV guidelines attached for reference.

Each time the GPT builder asks me something and I'm not entirely sure what to say, I might copy and paste or screenshot its question into this window, and I'd like your help optimising what I say back. 

Do you understand?

‍

Final Note Before Kicking Off: I Love Feedback

This guide is a living document, and I am always open to improving it. If you have any tips, tricks, or feedback, please share them with me.

Okay enough chit-chat, let's go!

Part 1: Initial GPT Configuration

1.01 Input Preparation

• Gather an extensive list of your brand voice examples: In the TOV guides you created in part 1, make sure you have a wide variety of good, almost-good, and bad examples for different channels (e.g., Twitter). Explain why each example works or doesn't.
• Export your TOV guidelines: Export your TOV guidelines including all examples in a word file or pdf. (Note: Connecting natively to Google Docs via API is beyond the scope of this).

1.02 Create a New GPT and Get Oriented.

Once you've followed the menu to create the new GPT, you find yourself inside the 'GPT Builder Tool.'

1.03 Initial prompt: Role Definition

Let’s set the stage and give the GPT builder a clear idea of what we’re trying to achieve. Feel free to copy and paste the prompt below into your GPT builder window after you've tailored it to your brand/business. Also, don't forget to upload and attach your Tone of Voice Guidelines here (Word / PDF format)!

Prompt

I want to create a content generation and quality assurance GPT for my business.

My company is called {insert brand name} and we {make/sell/do X}.

This GPT will help generate content for various communication mediums (e.g., social media posts, EDMs, press releases amongst others) and perform quality control audits to ensure all content aligns with our Tone of Voice guidelines (see attached).

‍

1.04 Define Name and Image

You’ll be prompted to assign a name and image to the GPT. Don’t worry, it’s very easy to change either of these later.

‍

Some ideas for names:

1. {brand name} TOV
2. {brand name} Content Assistant
3. {brand name} Content Expert
4. {brand name} Content Maestro
5. {brand name} TOV Copy Expert

‍

1.05 Define Constraints (AKA Things to ‘Focus On’, ‘Emphasise’ or ‘Avoid’)

The GPT may ask for 'constraints' in a variety of ways. Basically, these are the parameters which define the responses it can and cannot make. Fortunately, we’ve already defined these at length in the TOV guidelines, and the GPT builder is more than capable of extracting these from the document.

I used a separate GPT thread to request assistance here. It gave me the following output, which I have updated slightly towards the bottom.

At a high level, the following aspects should be emphasised in accordance with the previously attached tone of voice guidelines:

1. Alignment with the Key Audience Segments
2. Brand Values and Mission Statements
3. Brand Personalitys
4. Voice
5. Writing Style Guidelines 
6. Vocabulary Guideliness
7. Medium/Channel Guidelines

Additionally, The following aspects should be avoided:

1. Inconsistent Tone: Avoid shifts in tone that do not align with my brand's voice.
2. Passive Voice: Use active voice to maintain clarity and engagement.
3. Terms identified as being important to Avoid
4. Jargon

Please note, the above two lists are not exhaustive. Comprehensive details and additional clarifications about things to emphasise and avoid can be found in the previously attached Tone of Voice guidelines.

In general, Content should consistently reflect our brand’s mission, values, and personality while being tailored to the specific medium and audience.

‍

1.06 Interaction with Users

At a certain point, you’ll be asked how you want the GPT to interact with its users. It is important to note here that we’re talking about the voice of the GPT ASSISTANT itself. We’re NOT talking about your brand’s voice.

Again, you might receive any number of responses for this. Here are some of the different ways I’ve been asked this:

‍

My typical response is as follows:

Somewhat formal and friendly. It should sound like a friendly service provider.

‍

1.07 Ignore prompt saying that it is ready

It’s around about here in the configuration journey that the GPT will tell you it is ready for testing. Feel free to have a play in the window on the right.

However, we’re not quite ready to move on. We still need to define several key user interactions - specifically:

1. the asking of clarifying questions, and;
2. the QA Audit tool behaviour.

1.08 Interactivity Guidelines 1: Ensure the GPT asks clarifying questions

In Machine Learning speak, asking for content without giving context is the same as just asking for generic content.

In my experience, providing the following clarifying information is the difference between your GPT producing average and fantastic content.

1. Context & Purpose. What is the purpose of the communication? Are you announcing a new product, promoting a sale, sharing a company update, releasing your quarterly performance summary?
2. Medium. Is it an EDM, Twitter post, press release, blog post, or something else?
3. Goals. What are the goals of this communication? Are you trying to boost traffic to your website, or to a physical store location? Promote a sale? Simply alert existing users that a change is coming up?
4. Audience. Who is the intended audience? Is it something intended for your pre-defined key target segments, or another audience (i.e. investors, newsletter subscribers, internal staff etc.)
5. Existing Copy [If applicable]. Is there existing copy that needs improvement or auditing?

My Prompt:

Before performing any content QA (AKA 'Audit' AKA 'Quality Control' AKA 'Review'), creating new content or refining content, you must always ask the following clarifying questions:

1. Context. What is the purpose of the communication? (e.g., announcing a new product, promoting a sale, sharing a company update)
2. Medium. Through which channel is the communication being delivered? (e.g., EDM, Twitter post, press release, blog post)
3. Goals. What are the primary goals of this communication? (e.g., increasing awareness, driving newsletter signups, boosting sales)
4. Audience. Who is the intended audience? Specify key target segments or other specific audience groups.
5. Existing Copy [If applicable]. Is there existing copy that needs improvement or auditing? If yes, please provide it.

Do not proceed with any task until you have asked these questions and received answers. It is crucial to gather this information first.

1.09 Interactivity Guidelines 2: Define QA Audit Flows and Outputs

Essentially what we’re doing here, is making sure that the QA process

1. Asks clarifying questions;
2. Carries out key QA tests as defined by all key sections of the TOV guidelines;
3. Summarises the results of the test in a quick reference table;
4. Provides a summary list of issues for the user to review at the bottom, and;
5. if requested, suggests ways to improve the copy.

‍

Notes here:

• The below prompt is long and may cause your GPT builder to get confused. You may seek to break it out over 3 prompts, by finishing each prompt with something like “in the next prompt, I’ll define the tests you need to run” etc.
• Full disclosure: This prompt just builds a base foundation for the tests that your QA tool should run. We will need to append more comprehensive rules later during the testing phases (See 2.02d) and as your TOV guidelines evolve.

‍

Here's the Prompt I used for the initial configuration:

When performing a QA support request, follow these steps:

___

1. Clarifying Questions: Ensure all necessary clarifying information is obtained:
    - What is the purpose of the communication?
    - What type of content is needed?
    - What are the primary goals of this communication?
    - Who is the intended audience? Specify key target segments or other specific audience groups.
    - Is there existing copy that needs improvement or auditing? If yes, please provide it.

___    
    
2. Run Tests: Perform the following tests:

    - Key Audience Segments
        - 1.1 Audience Segment Targeted: Is the content tailored to the intended audience?
        - 1.2 Demographic Appropriateness: Does the content resonate with the demographic characteristics of the target audience?
        - 1.3 Psychographic Alignment: Does the content align with the interests, values, and attitudes of the target audience?
        - 1.4 Media Consumption Habits: Is the content suitable for the preferred communication channels of the target audience?
        
        
    - Brand Values and Mission Statement
        - 2.1 Message Aligned with Brand Values: Does the content reflect the core values of sustainability, adventure, community, quality, and innovation?
        - 2.2 Mission Statement Consistency: Is the content consistent with the brand's mission statement?
        
        
    - Brand Personality
        - 3.1 Personality Traits Respected: Does the content embody the brand’s personality traits?
    - Voice
        - 4.1 Voice Traits Respected: Are the selected top voice characteristics reflected in the content?
        - 4.2 Consistency in Tone: Is the tone of the content consistent with the defined brand voice?
        
        
    - Writing Style Guidelines
        - 5.1 Grammar and Punctuation Rules: Are the established grammar and punctuation rules followed?
        - 5.2 Syntax Preferences Respected: Does the content adhere to the specified syntax preferences?
        - 5.3 Active Voice: Is active voice predominantly used, avoiding passive constructions?
        - 5.4 Formatting Standards: Are the document formatting standards met?
        - 5.5 Language and Spelling Conventions: Are Australian English spelling and grammar rules used correctly?
        
        
    - Vocabulary Guidelines
        - 6.1 Preferred Terms Used: Are the preferred terms appropriately used in the content?
        - 6.2 No Terms to Avoid Used: Are the terms to avoid absent from the content?
        
        
    - Medium/Channel Guidelines
        - 7.1 Identify Key Channels: Is the content appropriate for the identified key channels?
        - 7.2 Platform Norms Respected: Are the platform-specific norms and expectations adhered to?
        
        
    - Examples: Our Voice in Practice
        - 8.1 Scenario-Specific Appropriateness: Given the training examples, is the content appropriate for the specific scenario it addresses?

___

3. Provide Output:
    - Summarise the tests in a table with columns for Test Number, Title, Description, and Pass? (✅ Pass, 🟡 Review and 🔴 Fail).
    - Underneath the table, summarise all items requiring review or have failed, explaining why and providing directions for how to fix the issue.
    - Ask the user if they want support drafting alternative copy. If yes, generate two alternative pieces of content that satisfy the TOV guidelines.

1.10 Create Predefined Prompts

Now, let’s make your life, and your stafff/collaborators’ lives easier by pre-populating some suggested prompts. Navigate to ‘Configure’ and update the ‘Conversation Starters’.

Suggested ‘Conversation Starters’

1. Help me draft new content
2. Help me fix some existing content
3. I have some copy. Help draft some alternatives
4. Help me QA some content

‍

The results should look something like this:

You might find later on that when you make a change in your GPT’s configuration, that these get automatically changed. I used the below prompt to counteract this:

‍

Every time I give you a new prompt, you update the standard conversation starters. Moving forward, DO NOT CHANGE THESE, EVER, unless I instruct you to.

I want these to be permanently set to the following:
Suggested ‘Conversation Starters’

1. Help me draft new content
2. Help me fix some existing content
3. I have some copy. Help draft some alternatives
4. Help me QA some content

1.11 Take a look around inside 'Configure'

Before we properly jump into testing, please take a quick look around the ‘Configure’ menu. Notice that in here, you can:

1. change the profile picture by clicking on it and uploading something else;
2. update the GPT’s name;
3. review the current, up to date summary of comprehensive instructions;
4. Update the conversation starters, and;
5. update the Knowledge file contents.

‍

1.11 Train the Model in Good / Bad Prompts

This is arguably the most important part of the entire training configuration process. We're going to ask the GPT to review all of the good and bad examples you've listed in section 8, and use these learnings to update its models.

Enter the following prompt to ensure your GPT is well adept at understanding what is good and bad copy.

I'd now like to refine your model by training you with examples of good and bad practice.

In the previously attached TOV guidelines, I first want you to review 8.2i.

Once you've reviewed them and updated your model, I'd like you to progress to the next section and update your model, until we've completed all of the examples in section 8.

Do you understand? If so, let's start.

Once you're done, ask me for confirmation to continue to 8.2ii and so on.

‍

Please just be aware, depending on how many examples you have defined, this might take some time.

‍

Part 2: Guided Testing and Training Sessions

Now that we’ve completed the initial configuration of the GPT, it’s time to test and refine how it actually responds. At this stage, there are only a few tests we really need to vet

1. The Content QA process Works
2. The Content Creation / refinement process works

Hot Tip: Get Comfortable Taking Screenshots!

‍This essentially allows you to streamline the reporting of poor ‘As Is’ behaviour: While you can drag files/images into your GPT, it's way easier to just screenshot things to your clipboard and paste them straight into the dialogue box. To save a screen snippet to your clipboard:

• Mac: Control + Command + Shift + 4
• PC: Windows + Shift + S

Similarly, when the GPT gives a good ‘best practice’ response, it’s worth having a screenshot of this for later training. Sometimes, the GPT will forget an important directive, so being able to remind it of ‘best practice’ is helpful.

2.01 Accessing and Reconfiguring Your GPT

Let's kick this off by getting you oriented and setting you up for success. If you ever get lost inside ChatGPT, you can find your GPT by navigating to 'Explore GPTs' > MyGPTs:

Then, when you're in a chat with your GPT, please note that you can:

1. Hot Tip: Pin it permanently in your side bar, and;
2. Access the Builder/Configuration window by clicking ‘Edit GPT’.

2.02 Testing the QA Flow

‍

2.02a) QA Flow Audit: Define Test Posts

Let’s start by defining some test posts. Defining these once at the start will save you time manually writing these out each time you're prompted. Here are some sample test posts I used.

Try writing your own prompt to ChatGPT to define these test examples. Think about the fields you need to ask it to write (i.e. purpose ... audience), and the desired output you'd like it to give you (i.e. list or table format?).

If you want to get even more sophisticated, you could specify for each of its examples to clearly violate 3-5 distinct rules, and to also list these rules out.

2.02b) QA Flow Audit: End-to-end Flow

The process should look a little like this:

Tests:

• Did the GPT immediately ask clarifying questions?
• Did all of the checks you wanted to happen actually happen?
• Did all of the output appear as expected?
• Is there anything you'd like to improve?

2.02c) QA Flow Audit: Handling Output Issues

After I deployed my GPT, I noticed that the output was is missing a couple of key elements, including:

1. The 'Pass?' status (✅ Pass, 🟡 Review and 🔴 Fail).
2. A description of the problem, and in a separate sentence, a suggestion for improvement.

So, I described the issue with screenshots to my prompt-support GPT. It recommended the following prompt:

‍

‍

‍2.02d) QA Flow Audit: If your GPT Isn't Enforcing Your TOV Rules

When I was evaluating my QA checks, I noticed that my GPT did not detect an issue with a heading syntax that clearly violated my TOV formatting standards (Section 5).

So, I checked My GPT's Configuration Instructions (see screenshot below) and realised that a lot of TOV guideline rules had not been persisted to my GPT.

Fortunately, this was an easy fix. It's just a matter of making the builder refresh its knowledge of the rules as per the GPT guidelines, before appending these rules to its list of existing rules.

Here’s the prompt I used to achieve this:

I don't think the standard tests you're performing as part of the QA Support Request Process fully capture all of the writing styles rules and conventions detailed in section 5.

Please review all of the individual rules and preferences described in section 5, including those described in 5.1, 5.2, 5.3, 5.4 and 5.5.

Rewrite each rule / preference as a test you can perform.

Append these rules to the current list of QA Tests you normally run.

Note: This will create a lot more rules, so please update your numbering convention as well, so the first rule in section 5 can be 5.1.01.

Let's go one section at a time starting with 5.1.

Once you've appended the new rules from 5.1 to your standard tests, we can move on to 5.2.

‍

Some additional notes:

• If you ask the GPT builder to update too much at once, it will break and flip out. Better to ask it to update its configuration in smaller, more digestible chunks.
• It's also worth paying attention to the outputs the GPT gives you in real time, because it can often miss important instructions. In the case below, the screengrab below, you can see the GPT Builder missed the instruction to append rules immediately after reviewing each sub section (5.1, then 5.2 and so on).

Here's that prompt in case you need it:

Before we move on to 5.2, please update your configuration to append the above rules to your existing tests.

2.03 Some Useful Prompts for Correcting Poor Responses

This is where it comes in handy to have access to a live document with your tone of voice guidelines. When your GPT gives you an output that’s not ideal, you can use these templates.

Template 1: Language and Vocabulary Corrections"

"When providing responses, your recent output included the phrase '{incorrect phrase}', which does not align with our preference to use {correct term/language}. Please remember to adhere to Australian English standards and avoid using words like '{banned word}'. Going forward, consistently apply the guideline that emphasises {specific instruction or preference}.

‍

Template 2: Tone and Style Adjustments

"In your last message, the tone used was {describe issue}, which differs from our brand’s desired tone of {desired tone}. Remember, we aim to engage our audience with a tone that is {specific desired tone characteristics}. Please revise your approach to ensure it aligns more closely with our brand personality by incorporating elements like {specific elements to include}."

‍

Template 3: Specific Content Instructions

"Your recent content mistakenly included {specific content mistake}, which we should avoid. It is important to remember that our brand does not promote {undesired content}. Please refer back to our content guidelines, specifically the section on {relevant section}, and ensure that all future content strictly follows these directives. Focus on {what to focus on} and ensure to reflect our core values of {brand values}."

2.04 Updating the Model on an Ongoing Basis

Simply update the TOV guidelines, edit the GPT, upload the new guidelines alongside the following prompt:

I have updated the Tone of Voice (TOV) Guidelines document. Please use the new document to reconfigure your settings and ensure all future content generation and QA audits align with the updated guidelines.

Here is the updated TOV Guidelines document: [Attach the Word document]

Let's sit down and go through the new TOV guidelines section by section, to update your model accordingly. We'll start with section 1.1, then 1.2, then go all the way through to the final section.

Please update your model after each section, and ask me if we can proceed to the next item.

‍

Again, if you ask it to update everything all at once, it will either skip over a lot of your new rules, or it will break and throw an error out. In my experience, it's best to just sit down with it and take the time to make sure each section is updating appropriately.

Thank you for taking the time to read through this guide on streamlining your business communications. I hope it has helped and provided you with useful insights and practical steps to automate and enhance your content creation and quality control processes. Remember, this is just a starting point!

Additionally, if you need any support with web design and web development, feel free to contact me!