@Sana Fatima In my opinion, the advantages of Straico are the bundling of LLMs and other AI services from a wide range of providers such as Anthropic, DeepSeek, Google, OpenAI, Perplexity, Swen, Meta, Mistral, xAI and others. See: https://straico.com/multimodel/

@Eduard Friesen No, unfortunately not. I've had access for years and love the freedom to use the AI model I need.

Great post! I’d add that the real magic often happens when you treat AI not as a “one-shot answer machine” but as part of an interactive process. Prompt engineering is the steering wheel, yes—but driving well also means checking your mirrors, adjusting speed, and taking feedback from the road. In AI terms, that’s a cycle of: - Critical evaluation of the output - Iterative refinement of the prompt/context - Requesting improvements or alternatives This back-and-forth transforms results from “good” to excellent. On top of that, multi-agent setups—where different AI agents take on distinct roles or perspectives (e.g., critic, strategist, creative)—can raise the quality even further. It’s like having a team of experts in the room instead of just one voice. So for me, the best results don’t come from a single perfect prompt, but from collaborating with AI in dialogue and sometimes letting multiple agents challenge each other’s outputs. That’s when it stops being just a tool—and becomes an actual partner in thinking.

@Eduard Friesen Moin in den Norden. I live on Magdalensberg and today I went for a lovely walk with my wonderful Doberman, »Juna«.

@Okolie Uchenna Ẹ káábọ̀!

To connect Google Sheets, follow these instructions: https://www.youtube.com/watch?v=FBGtpWMTppw&t=4s Overview of the workflow: Load webhook & questions: These steps remain unchanged. The workflow is started via webhook and the questions are loaded from Google Sheets. Ask questions and collect answers (in the loop): You continue to run through the loop to ask each question and capture the technician's answer via speech-to-text. Instead of validating each individual answer immediately, save the question and the corresponding answer together. The Set node is ideal for this. Within the loop: Add a Set node after the speech-to-text node. Configuring the Set node: Save an object containing the current question and the given answer and add it to an array that collects all previous conversation pairs. This allows you to gradually build up the entire dialogue. Summarise all answers (after the loop): After the loop has been run for all questions (at the ‘Done’ output of the loop node), you have an array with all questions and their corresponding answers. Use another Set node or a Code node to put this data into a clear format for the AI agent, e.g. simple text or a JSON object. The AI agent with system prompt (the core of validation): Add an AI agent or OpenAI node. Connect it to the node that summarised the responses. Configuration of the AI agent node: System prompt: Here you define the role and exact instructions for the AI. This is the crucial part. User Prompt / Input: Insert the previously summarised text with all questions and answers here. Decision based on the AI evaluation: Add an IF node after the AI node. This node now checks the AI's response. Condition: Check whether the isValid field from the AI agent's JSON response is true. You can read the JSON string returned by the AI directly with an expression such as {{ $json.body.isValid }}. Two paths after the decision: True path (data is valid): The answers are correct. Insert the Google Sheets node here to write the collected answers to the responses sheet.

Possible system prompt: You are a quality control assistant for safety reports. Your task is to validate the responses of a field service technician. The responses must be complete, accurate and relevant to the question. Analyse the following conversation and assess whether all questions have been answered satisfactorily. A response is insufficient if it is too short (e.g. only one word where a description is expected), evasive or clearly does not answer the question. Respond exclusively in JSON format using the following keys: - ‘isValid’: a Boolean (true if all answers are valid, otherwise false). - ‘feedback’: a short text explaining which question was answered inadequately if isValid is false. Also include a note for the technician on what they should do better.

To start with, I would install n8n locally with ngrok. Whether I use Docker or not depends on whether I use community nodes or not.