Workflow Templates

Summarize YouTube Videos & Chat About Content with GPT-4o-mini via Telegram Description This n8n workflow automates the process of summarizing YouTube video transcripts and enables users to interact with the content through AI-powered question answering via Telegram. It leverages the GPT-4o-mini model to generate summaries and provide insights based on the video’s transcript. How It Works Input: The workflow starts by receiving a YouTube video URL. This can be submitted through: A Telegram chat message. A webhook (e.g., triggered by a shortcut on Apple devices). Transcript Extraction: The URL is processed to extract the video transcript using the custom youtubeTranscripter community node (available here). The transcript is concatenated into a single text and stored in a Google Docs document. Summarization: The GPT-4o-mini AI model analyzes the transcript and generates a structured summary, including: A general overview. Key moments. Instructions (if applicable). The summary is then sent back to the user via Telegram. Interactive Q&A: Users can ask questions about the video content via Telegram. The AI retrieves the stored transcript from Google Docs and provides accurate, context-based answers, which are sent back through Telegram. Setup Instructions To configure this workflow, follow these steps: Import the Workflow: Download the provided JSON template and import it into your n8n instance. Install the Community Node: Install the youtubeTranscripter community node via npm: npm install n8n-nodes-youtube-transcription-kasha Important: This node requires a self-hosted n8n instance due to its external dependencies. Configure Nodes: Webhook: Set up the webhook to receive YouTube URLs. Alternatively, configure the Telegram node if using Telegram as the input method. Google Docs: Provide valid credentials to enable writing the transcript to a Google Docs document. AI Model: Set up the GPT-4o-mini model for summarization and Q&A functionality. Test the Workflow: Send a YouTube URL via your chosen input method (Telegram or webhook) and confirm that the summary is generated and delivered correctly. Customization Language**: Adjust the AI prompts to generate summaries and answers in any desired language. Output Format**: Modify the summary structure by editing the prompt in the summarization node. Input Methods**: Replace the Telegram node with another messaging or input node to adapt the workflow to different platforms. Who Can Benefit? This template is perfect for: Content Creators**: Quickly summarize video content for repurposing or review. Students and Researchers**: Extract key insights from educational or informational videos efficiently. General Users**: Interact with video content via AI without needing to watch the full video. Problem Solved This workflow simplifies video content consumption by: Automating the extraction and summarization of key points. Enabling interactive Q&A to address specific questions without rewatching the video. Additional Notes Disclaimer**: The youtubeTranscripter community node is required and only works on self-hosted n8n instances due to its reliance on external services. Apple Users**: Enhance your experience with a custom shortcut to share YouTube videos directly to the workflow. Download the shortcut here.

Who is this for? This template is designed for internal support teams, product specialists, and knowledge managers in technology companies who want to automate ingestion of product documentation and enable AI-driven, retrieval-augmented question answering via WhatsApp. What problem is this workflow solving? Support agents often spend too much time manually searching through lengthy documentation, leading to inconsistent or delayed answers. This solution automates importing, chunking, and indexing product manuals, then uses retrieval-augmented generation (RAG) to answer user queries accurately and quickly with AI via WhatsApp messaging. What these workflows do Workflow 1: Document Ingestion & Indexing Manually triggered to import product documentation from Google Docs. Automatically splits large documents into chunks for efficient searching. Generates vector embeddings for each chunk using OpenAI embeddings. Inserts the embedded chunks and metadata into a MongoDB Atlas vector store, enabling fast semantic search. Workflow 2: AI-Powered Query & Response via WhatsApp Listens for incoming WhatsApp user messages, supporting various types: Text messages: Plain text queries from users. Audio messages: Voice notes transcribed into text for processing. Image messages: Photos or screenshots analyzed to provide contextual answers. Document messages: PDFs, spreadsheets, or other files parsed for relevant content. Converts incoming queries to vector embeddings and performs similarity search on the MongoDB vector store. Uses OpenAI’s GPT-4o-mini model with retrieval-augmented generation to produce concise, context-aware answers. Maintains conversation context across multiple turns using a memory buffer node. Routes different message types to appropriate processing nodes to maximize answer quality. Setup Setting up vector embeddings Authenticate Google Docs and connect your Google Docs URL containing the product documentation you want to index. Authenticate MongoDB Atlas and connect the collection where you want to store the vector embeddings. Create a search index on this collection to support vector similarity queries. Ensure the index name matches the one configured in n8n (data_index). See the example MongoDB search index template below for reference. Setting up chat Authenticate the WhatsApp node with your Meta account credentials to enable message receiving and sending. Connect the MongoDB collection containing embedded product documentation to the MongoDB Vector Search node used for similarity queries. Set up the system prompt in the Knowledge Base Agent node to reflect your company’s tone, answering style, and any business rules, ensuring it references the connected MongoDB collection for context retrieval. Make sure Both MongoDB nodes (in ingestion and chat workflows) are connected to the same collection with: An embedding field storing vector data, Relevant metadata fields (e.g., document ID, source), and The same vector index name configured (e.g., data_index). Search Index Example: { "mappings": { "dynamic": false, "fields": { "_id": { "type": "string" }, "text": { "type": "string" }, "embedding": { "type": "knnVector", "dimensions": 1536, "similarity": "cosine" }, "source": { "type": "string" }, "doc_id": { "type": "string" } } } }

📽️ What this workflow does This workflow turns a user-submitted form with country or animal names into a cinematic video with animated scenes and immersive ambient audio. Using GPT-4 for prompt generation, Dumpling AI for visual creation,& Replicate for motion animation, ElevenLabs for sound generation, and Creatomate for video stitching, it fully automates video production — from raw idea to rendered file. 🎯 What problem is this solving? Creating engaging multimedia content can take hours. This workflow automates the entire process of ideation, design, and rendering of high-quality cinematic clips, eliminating the need for manual video editing or audio production. 👥 Who is this for? Content creators and educators Digital artists and storytellers Marketers or YouTubers creating short-form visual content No-code/AI automation enthusiasts ⚙️ Setup Instructions ✅ Step 1: Google Sheet Create a Google Sheet with two columns: Title Generated videos Update the Sheet ID and tab name in the final node. ✅ Step 2: Google Drive Create two folders: One for ambient audio tracks One for final generated videos Update the folder IDs in both Google Drive nodes. ✅ Step 3: Credentials Setup Make sure all your API tokens are saved as credentials in n8n. This workflow uses the following integrations: OpenAI (GPT-4) Dumpling AI (via HTTP header) Replicate.com ElevenLabs Google Drive Google Sheets Creatomate ✅ Step 4: Form Fields Ensure your trigger form includes these fields: Title Country 1, Country 2, Country 3, Country 4 Style (e.g., cinematic, epic, fantasy, noir, etc.) 🧩 How it works User Form Submission Kicks off the workflow with the required inputs. Format Inputs Combines all 4 countries/animals into a single array. GPT-4: Generate Visual Prompts Uses GPT-4 to create rich cinematic descriptions per animal/country. Dumpling AI: Create Images Each description becomes a high-quality visual. GPT-4: Create Motion Prompts Each image prompt is rewritten into motion-based video prompts. Replicate: Animate Prompts and images are sent to Replicate’s model for animation. GPT-4: Generate Sound Prompt Based on the style, GPT-4 creates an ambient sound idea. ElevenLabs: Create Ambient Audio Audio is generated and uploaded to Google Drive. Creatomate: Stitch All Media All 4 motion videos and the audio track are stitched into one cinematic output. Upload to Google Drive + Log to Sheet Final video is saved in Drive and logged in Sheets with its title and link. 🛠️ How to Customize 🎨 Modify GPT prompts for different themes (e.g., horror, fantasy, sci-fi). 🧠 Swap animals for characters, objects, or locations. 🎧 Replace ambient sound with ElevenLabs voiceovers or music. 📂 Add metadata logging (generation time, duration, tags). 🧪 Try using alternative video tools like Pika Labs or Runway ML. ✅ Requirements n8n self-hosted or cloud instance Active accounts for: OpenAI, Dumpling AI, Replicate, ElevenLabs, Creatomate Google credentials set up for Drive + Sheets This is a perfect end-to-end automation that showcases the power of AI + automation for video storytelling.

How it works: Using a Crew of AI agents (Senior Researcher, Visionary, and Senior Editor), this crew will automatically determine the right questions to ask to produce a detailed fundamental stock analysis. This application has two components: a front-end and a Stock Q&A engine. The front end is the team of agents automatically figuring out the questions to ask, and the back-end part is the ability to answer those questions with the SEC 10K data. This template implements the Stock Q&A engine. For the front-end of the application, you can choose one of two options: using CrewAI with the Replit environment (code approach) fully visual approach with n8n template (AI-powered automated stock analysis) Setup steps: Use first workflow in template to upsert a company annual report PDF (such as from SEC 10K filling) Get URL for Webhook in second workflow template CrewAI front-end: Youtube overview video Fork this AI Agent environment Crew Agent Environment Set the webhook URL into N8N_WEBHOOK_URL variable Set OpenAI_API_KEY variable

How it works This template uses a slack app to connect with your google calendar, generate an instant google meet link and post it as a message in a slack channel Setup steps Firstly, you'll need to create a slack app Authenticate and connect your slack account Connect and choose the Google calendar you want to generate Google meet links for Customize your slack message Then using a /meet command in slack, you can instantly generate and post your Google meet links

This n8n workflow is a proof-of-concept template exploring how we might work with multimodal LLMs and their multi-image analysis capabilities. In this demo, we compare 2 screenshots of a webpage taken at different timestamps and pass both to our multimodal LLM for a visual comparison of differences. Handling multiple binary inputs (ie. images) in an AI request is supported by n8n's basic LLM node. How it works This template is intended to run as 2 parts: first to generate the base screenshots and next to run the visual regression test which captures fresh screenshots. Starting with a list of webpages captured in a Google sheet, base screenshots are captured for each using a external web scraping service called Apify.com (I prefer Apify but feel free to use whichever web scraping service available to you) These base screenshots are uploaded to Google Drive and will be referenced later when we run our testing. Phase 2 of the workflow, we'll use a scheduled trigger to fire sometime in the future which will reuse our web scraping service to generate fresh screenshots of our desired webpages. Next, re-download our base screenshots in parallel and with both old and new captures, we'll pass these to our LLM node. In the LLM node's options, we'll define 2 "user message" inputs with the type of binary (data) for our images. Finally, we'll prompt our LLM with our testing criteria and capture the regressions detected. Note, results will vary depending on which LLM you use. A final report can be generated using the LLM's output and is uploaded to Linear. Requirements Apify.com API key for web screenshotting service Google Drive and Sheets access to store list of webpages and captures Customising this workflow Have your own preferred web screenshotting service? Feel free to swap out Apify with your service of choice. If the web screenshot is too large, it may prove difficult for the LLM to spot differences with precision. Try splitting up captures into smaller images instead.

What this workflow does This (set of) workflow(s) shows how to start multiple sub-workflows, asynchronously, in parallel, and then wait for all of them to complete. Normally sub-workflows would need to be run synchronously, in series, or, if they are executed asynchronously (to run concurrently, in parallel), there is no easy way to merge/wait for an arbitrary number of them to complete. This is a "design pattern" template to show one approach for running multiple, data-driven instances of a sub-workflow "asynchronously," in parallel (instead of running them one at a time in series), but still prevent the later steps in the workflow from continuing until all of the sub-workflows have reported back that they are finished, via callback URL. There are other techniques involving messaging services, database tables, or other external "flow manager" helpers, but this technique accomplishes the goal fully within n8n. Setup To implement this pattern, examine the nodes in the template and modify the incoming data leading to: A split-out loop to acynchronously execute a sub-workflow multiple times, in parallel. For instance, each sub-workflow might process one of a list of incoming documents. The resumeUrl for the main/parent workflow is provided to all of the sub-workflow executions, along with a unique identifier that can be counted later (e.g. a document file-name). A "wait-for-all" loop that checks whether all sub-workflows have reported back (if node) and builds a unique list of identifiers from the callbacks received from each execution of the sub-workflow. The sub-workflow should be designed to respond immediately (async) and later send a callback request when it has finished processing. The callback request should include the unique identifier value received when the sub-workflow it was started. This is meant to be a possible answer to questions like this one about running things in parallel, maybe this one about waiting for things to finish, this one about managing sub-batches of things by waiting for each batch, or this one about running things in parallel. The topic of how to do this comes up A LOT, and this is one of the only techniques that (so far) seems to work.

Automatically backs up your workflows to Github and generates documentation in a Notion database. Weekly run, uses the "internal-infra" tag to look for new or recently modified workflows Uses a Notion database page to hold the workflow summary, last updated date, and a link to the workflow Uses OpenAI's 4o-mini to generate a summarization of what the workflow does Stores a backup of the workflow in GitHub (recommend a private repo) Sends notification to Slack channel for new or updated workflows Who is this for Anyone seeking backup of their most important workflows Anyone seeking version control for their most important workflows Credentials required N8N: You will need an N8N credential created so the workflow can query the N8N instance to find all active workflows with the "internal-infra" tag Notion: You will need an Notion credential created OpenAI: You will need an OpenAI credential, unless you intend on rewiring this with your AI of choice (ollama, openrouter, etc.) GitHub: You will need an GitHub credential Slack: You will require an Slack credential, recommend a Bot / access token configuration Setup Notion Create a database with the following columns. Column type is specified in [type]. Workflow Name [text] isActive (dev) [checkbox] Error workflow setup [checkbox] AI Summary [text] Record last update [date/time] URL (dev) [text/url] Workflow created at [date/time] Workflow updated at [date/time] Slack Create a channel for updates to be posted into Github Create a private repo for your workflows to be exported into N8N Download & install the template Configure the blocks to use your N8N, Notion, OpenAI & Slack credentials for your own Edit the "Set Fields" block and change the URL to that of your N8N instance (cloud or self-hosted) Edit the "Add to Notion" action and specify the Database page you wish to update Edit the Slack actions to specify the Channel you want slack notifications posted to Edit the GitHub actions to specify the Repository Owner & Repository Name Sample output in Notion Workflow diagram

⚠️ Important Disclaimer: This template is only compatible with a self-hosted n8n instance using a community node. Who is this for? This workflow is ideal for digital content creators, marketers, social media managers, and automation enthusiasts who want to produce fully automated vertical video content featuring inspirational or motivational quotes. Specifically tailored for Thai language, it effectively demonstrates integration of AI-generated imagery, video, ambient sound, and visually appealing quote overlays. What problem is this workflow solving? Manually creating high-quality, vertically formatted quote videos is often repetitive, time-consuming, and involves multiple tedious steps like selecting suitable visuals, editing audio tracks, and correctly overlaying text. Additionally, manual uploading to platforms like YouTube and maintaining accurate content records are prone to errors and inefficiencies. What this workflow does: Fetches a quote, author, and scenic background description from a Google Sheet. Automatically generates a vertical background image using the Flux AI (txt2img) API. Transforms the AI-generated image into a subtly animated cinematic vertical video using the Kling video-generation API. Generates an immersive, ambient background sound using ElevenLabs’ sound generation API. Dynamically overlays the selected Thai-language quote and author text onto the generated video using FFmpeg, ensuring visually appealing typography (e.g., Kanit font). Automatically uploads the final video to YouTube. Updates the resulting YouTube video URL back to the Google Sheet, keeping your content records current and well-organized. Setup Requirements: This workflow requires a self-hosted n8n instance, as the execution of FFmpeg commands is not supported on n8n Cloud. Ensure FFmpeg is installed on your self-hosted environment. API keys and accounts setup for Flux, Kling, ElevenLabs, Google Sheets, Google Drive, and YouTube. Google Sheets Setup: Your Google Sheet must include these columns: Index** Unique identifier for each quote Quote (Thai)** Quote text in Thai language (or your chosen language) Pen Name (Thai)** Author or pen name of the quote's creator Background (EN)** Short English description of the scene (e.g., "sunrise over mountains") Prompt (EN)** Detailed English prompt describing the image/video scene (e.g., "peaceful sunrise with misty mountains") Background Image** URL of AI-generated image (updated automatically) Background Video** URL of generated video (updated automatically) Music Background** URL of generated ambient audio (updated automatically) Video Status** YouTube URL (updated automatically after upload) A ready-to-use Google Sheets template is provided [here (provide your actual link)]. To help you get started quickly, you can use this template spreadsheet. Next steps: Authenticate Google Sheets, Google Drive, YouTube API, Flux AI, Kling API, and ElevenLabs API within n8n. Ensure FFmpeg supports fonts compatible with your chosen language (for Thai, "Kanit" font is recommended). Prepare your Google Sheets with desired quotes, authors, and image/video prompts. How to customize this workflow to your needs: Fonts:** Adjust font type, size, color, and positioning within the provided FFmpeg commands in the workflow’s code nodes. Verify that selected fonts properly support your target language. Media Customization:** Customize the scene descriptions in your Google Sheet to change image/video backgrounds automatically generated by AI. Quote Management:** Easily manage, add, or update quotes and associated details directly via Google Sheets without workflow modifications. Audio Ambiance:** Customize or adjust the ambient sound prompt for ElevenLabs within the workflow’s HTTP Request node to match your video's desired mood. Benefits of using AI-generated content and localized fonts: Leveraging AI-generated visual and audio elements along with localized fonts greatly enhances audience engagement by creating visually appealing, professional-quality content tailored specifically for your target audience. This automated workflow drastically reduces production time and manual effort, enabling rapid, consistent content creation optimized for platforms such as YouTube Shorts, Instagram Reels, and TikTok.

This template can be used to find the content gaps in PDF documents using the InfraNodus knowledge graph / GraphRAG text representation and then generate ideas / questions / AI prompts that bridge those gaps based on optimizing the knowledge graph's structure. Simply upload several PDF files (research papers, corporate or market reports, etc) and generate an idea in seconds. The template is useful for: generating ideas / questions for research generating content ideas based on competitors' discourse finding blind spots in any discourse and generating ideas that address them. avoiding the generic bias of LLM models and focusing on what's important in your particular context What are Content Gaps and Knowledge Graphs? Knowledge graphs represent any text as a network: the main concepts are the nodes, their co-occurrences are the connections between them. Based on this representation, we build a graph and apply network science metrics to rank the most important nodes (concepts) that serve as the crossroads of meaning and also the main topical clusters that they connect. Naturally, some of the clusters will be disconnected and will have gaps between them. These are the topics (groups of concepts) that exist in this context (the documents you uploaded) but that are not very well connected. Addressing those gaps can help you see which groups of concepts you could connect with your own ideas. This is exactly what InfraNodus does: builds the structure, finds the gaps, then uses the built-in AI to generate research questions and ideas that bridge those gaps. How it works 1) Step 1: First, you upload your PDF files using an online web form, which you can run from n8n or even make publicly available. 2) Steps 2-4: The documents are processed using the Code and PDF to Text nodes to extract plain text from them. 3) Step 5: This text is then sent to the InfraNodus GraphRAG node that creates a knowledge graph, identifies structural gaps in this graph, and then uses built-in AI to generate ideas or research questions / prompts (if you use the InfraNodus question module instead). 4) Step 6: The ideas are then shown to the user in the same web form. Optionally, you can hook this template to your own workflow and send the idea / question generated to your own AI model / agent for further processing. If you'd like to sync this workflow to PDF files in a Google Drive folder, you can copy our Google Drive PDF processing workflow for n8n. How to use You need an InfraNodus GraphRAG API account and key to use this workflow. Create an InfraNodus account Get the API key at https://infranodus.com/api-access and create a Bearer authorization key. Add this key into the InfraNodus GraphRAG HTTP node(s) you use in this workflow. You do not need any OpenAI keys for this to work. Optionally, you can change the settings in the Step 4 of this workflow and enforce it to always use the biggest gap it identifies. Requirements An InfraNodus account and API key Note: OpenAI key is not required. You will have direct access to the InfraNodus AI with the API key. Customizing this workflow You can use this same workflow with a Telegram bot or Slack (to be notified of the summaries and ideas). You can also hook up automated social media content creation workflows in the end of this template, so you can generate posts that are relevant (covering the important topics in your niche) but also novel (because they connect them in a new way). Check out our n8n templates for ideas at https://n8n.io/creators/infranodus/ Also check the full tutorial with a conceptual explanation at https://support.noduslabs.com/hc/en-us/articles/20454382597916-Beat-Your-Competition-Target-Their-Content-Gaps-with-this-n8n-Automation-Workflow Also check out the video introduction to InfraNodus to better understand how knowledge graphs and content gaps work: For support and help with this workflow, please, contact us at https://support.noduslabs.com

This template can be used to generate research ideas from PDF scientific papers based on the content gaps found in text using the InfraNodus knowledge graph GraphRAG knowledge graph representation. Simply upload several PDF files (research papers, corporate or market reports, etc) and the template will generate a research question, which will then be sent as an AI prompt to the InfraNodus GraphRAG system that will extract the answer from the documents. As a result, you find the gap in a collection of research papers and bridge it in a few seconds . The template is useful for: advancing scientific research generating AI prompts that drive research further finding the right questions to ask to bridge blind spots in a research field avoiding the generic bias of LLM models and focusing on what's important in your particular context Using Content Gaps for Generating Research Questions Knowledge graphs represent any text as a network: the main concepts are the nodes, their co-occurrences are the connections between them. Based on this representation, we build a graph and apply network science metrics to rank the most important nodes (concepts) that serve as the crossroads of meaning and also the main topical clusters that they connect. Naturally, some of the clusters will be disconnected and will have gaps between them. These are the topics (groups of concepts) that exist in this context (the documents you uploaded) but that are not very well connected. Addressing those gaps can help you see which groups of concepts you could connect with your own ideas. This is exactly what InfraNodus does: builds the structure, finds the gaps, then uses the built-in AI to generate research questions that bridge those gaps. How it works 1) Step 1: First, you upload your PDF files using an online web form, which you can run from n8n or even make publicly available. 2) Steps 2-4: The documents are processed using the Code and PDF to Text nodes to extract plain text from them. 3) Step 5: This text is then sent to the InfraNodus GraphRAG node that creates a knowledge graph, identifies structural gaps in this graph, and then uses built-in AI to research questions, which are then used as AI prompts. 4) Step 6: The research questino is sent to the InfraNodus GraphRAG system that represents the PDF documents you submitted as a knowledge graph and then uses the research question generated to come up with an answer based on the content you uploaded. 4) Step 7: The ideas are then shown to the user in the same web form. Optionally, you can derive the answers from a different set of papers, so the question is generated from one batch, but the answer is generated from another. If you'd like to sync this workflow to PDF files in a Google Drive folder, you can copy our Google Drive PDF processing workflow for n8n. How to use You need an InfraNodus GraphRAG API account and key to use this workflow. Create an InfraNodus account Get the API key at https://infranodus.com/api-access and create a Bearer authorization key. Add this key into the InfraNodus GraphRAG HTTP node(s) you use in this workflow. You do not need any OpenAI keys for this to work. Optionally, you can change the settings in the Step 4 of this workflow and enforce it to always use the biggest gap it identifies. Requirements An InfraNodus account and API key Note: OpenAI key is not required. You will have direct access to the InfraNodus AI with the API key. Customizing this workflow You can use this same workflow with a Telegram bot or Slack (to be notified of the summaries and ideas). You can also hook up automated social media content creation workflows in the end of this template, so you can generate posts that are relevant (covering the important topics in your niche) but also novel (because they connect them in a new way). Check out our n8n templates for ideas at https://n8n.io/creators/infranodus/ Also check the full tutorial with a conceptual explanation at https://support.noduslabs.com/hc/en-us/articles/20454382597916-Beat-Your-Competition-Target-Their-Content-Gaps-with-this-n8n-Automation-Workflow Also check out the video introduction to InfraNodus to better understand how knowledge graphs and content gaps work: For support and help with this workflow, please, contact us at https://support.noduslabs.com

This n8n template demonstrates a simple approach to using AI to automate the generation of blog content which aligns to your organisation's brand voice and style by using examples of previously published articles. In a way, it's quick and dirty "training" which can get your automated content generation strategy up and running for very little effort and cost whilst you evaluate our AI content pipeline. How it works In this demonstration, the n8n.io blog is used as the source of existing published content and 5 of the latest articles are imported via the HTTP node. The HTML node is extract the article bodies which are then converted to markdown for our LLMs. We use LLM nodes to (1) understand the article structure and writing style and (2) identify the brand voice characteristics used in the posts. These are then used as guidelines in our final LLM node when generating new articles. Finally, a draft is saved to Wordpress for human editors to review or use as starting point for their own articles. How to use Update Step 1 to fetch data from your desired blog or change to fetch existing content in a different way. Update Step 5 to provide your new article instruction. For optimal output, theme topics relevant to your brand. Requirements A source of text-heavy content is required to accurately breakdown the brand voice and article style. Don't have your own? Maybe try your competitors? OpenAI for LLM - though I recommend exploring other models which may give subjectively better results. Wordpress for blog but feel free to use other preferred publishing platforms. Customising this workflow Ideally, you'd want to "train" your agent on material which is similar to your output ie. your social media post may not get the best results from your blog content due to differing formats. Typically, this brand voice extraction exercise should run once and then be cached somewhere for reuse later. This would save on generation time and overall cost of the workflow.