Setting up a workflow where ChatGPT consistently follows a client’s “do’s and don’ts” for content creation (e.g., avoiding certain words, using specific phrases, maintaining a particular tone) is fairly simple using ChatGPT and Microsoft’s Power Platform. Here is a quick walkthrough of how that’s done.
Collect and Document Client Guidelines
- Initial Briefing: Start with a detailed discussion with the client to gather their “do’s and don’ts,” including tone, preferred terminology, structure, and prohibited words.
- Create a Guideline Document: Store this information in a structured format using a SharePoint list or a Microsoft Dataverse table within the Power Platform. Include columns for tone, terminology, content structure, and restricted words. SharePoint lists tend to be easier to use and have less overhead than Dataverse, so I recommend SharePoint.
Prompt Engineering
- Create a Master Prompt Template: Use the client’s guidelines to craft a master prompt template. This template should be stored in a format that can be accessed and modified within the Power Platform, such as a SharePoint document.
- Embed Guidelines in the Prompt: Make the prompt modular and adaptable. For instance, you can include placeholder tags in the prompt (e.g., “{tone}”, “{preferred_terms}”) that can be dynamically replaced based on data stored in Dataverse.
⠀Workflow Automation Using Power Automate
- Automate Prompt Preparation: Create a Power Automate flow that triggers when a new content request is submitted (e.g., through a Microsoft Form or a SharePoint list). This flow can pull the specific guidelines from Dataverse or SharePoint and use them to dynamically build the prompt.
- Call OpenAI API: Use the HTTP action in Power Automate to send a request to the OpenAI API with the constructed prompt. The flow will pass the generated content back into the workflow for further processing.
Content Generation and Review Loop
- Store Generated Content: Once the content is generated, store it in SharePoint for initial review.
- Automated Initial Review: Implement automated checks using Power Automate to scan the content for compliance with guidelines. For example, you can use the AI Builder Text Recognition model to analyze the text and flag prohibited words or phrases.
- Manual Review with Power Apps: Create a Power App where editors can review the generated content. This app can display the content alongside the client’s guidelines, making it easier to ensure compliance. Editors can make adjustments directly within the app if needed. Or use a tech writer/content strategist to scan words on a periodic basis depending on content turnaround times.
Feedback and Iteration
- Feedback Collection: Use Microsoft Forms to gather feedback from the client on the generated content. Integrate the form responses into Dataverse to keep a record of feedback for each content piece.
- Prompt Refinement: Create a Power Automate flow to use this feedback for prompt refinement. For instance, if a specific word is flagged frequently, automatically update the “restricted words” list in Dataverse.
⠀Quality Assurance and Reporting
- Quality Checks: Use Power BI to create a dashboard that monitors the quality of the generated content. This can include metrics like the frequency of manual edits, compliance rate with the guidelines, and client satisfaction scores.
- Reporting: Set up scheduled Power BI reports to track the success of the workflow in maintaining adherence to the client’s “do’s and don’ts.” Share these reports with stakeholders to identify areas for improvement.
Continuous Improvement
- Regular Updates: Use Power Automate to schedule regular checks for any updates in the client’s guidelines stored in Dataverse. If updates are detected, send notifications to the team to ensure everyone is aware of the changes.
- Training and Documentation: Store training materials and documentation in a SharePoint library and link them within the Power App to ensure that team members have easy access to the latest guidelines and workflow processes.
This can be adapted and modified based on templates and client need. I have created several of these instance in the past of customers. Some have had the ability to work with what I provided and needed no further enhancement, others have needed me to come in and tweak on occasion.