Build an AI Agent with Claude in Minutes: A Hands-On Workflow Automation Guide

Want to build an AI agent that automatically generates videos and handles emails — all in just a few minutes? This tutorial walks through how to use the Projects feature in the Claude app to turn natural language instructions directly into executable workflows. The core idea is simple: feed Claude a few ready-made workflow templates, describe your automation goal in plain English, and let it generate a complete JSON configuration you can import right away.

The Big Picture: Let AI Write Your Automation Workflows

Traditionally, building automation pipelines means manually dragging and dropping nodes and configuring parameters inside tools like n8n or Make — a process with a steep learning curve. The approach in this tutorial sidesteps all that manual work: hand your requirements to Claude in plain language, and it generates an importable workflow JSON file directly.

The key enabler here is the Projects feature. By uploading a few existing workflow examples to a project upfront, the AI learns the JSON structure and formatting conventions of your target tool, allowing it to produce correctly formatted, immediately usable configurations. This "few-shot demonstration" approach is essentially providing the model with contextual references, which dramatically improves the quality and usability of its output.

n8n and Make (formerly Integromat) are two of the most popular no-code/low-code automation platforms today. n8n is known for being open-source and self-hostable, storing workflows in JSON format with nodes connected via data flows. Make offers a visual "Scenario" editor with integrations for hundreds of third-party services. Both platforms follow strict JSON schema conventions for their workflow config files — node types, parameter fields, connection relationships, and so on. Because of this, giving a large language model enough real-world examples lets it learn this "dialect" and produce platform-compliant configuration files, rather than structurally broken output that can't be imported.

Step-by-Step: From Creating a Project to Generating a Workflow

The entire process breaks down into six steps:

Steps 1–3: Create a Project

Open the Claude desktop app, navigate to the Projects area, and create a new project dedicated to workflow automation. Keeping all your related automation tasks under one project makes it easy to reuse context across conversations.

Step 4: Set Project Instructions

Fill in the Instructions area for your project. If you don't have anything specific yet, simply typing "automation" as a placeholder works fine — it lets the AI understand the purpose of the project. Think of project instructions as setting a "system role" for the AI; they'll shape the direction of all its subsequent responses.

Step 5: Upload Workflow Examples

This is the most critical step in the entire approach. In the Files area, upload 2–3 existing workflow JSON files — things like a Gmail auto-labeling workflow or a Telegram AI assistant template. With these real examples as reference, Claude will strictly follow the correct format and structure when generating new workflows, making the whole process remarkably easy to get started with.

This technique is known in prompt engineering as Few-Shot Prompting: by including a small number of examples in the context, you guide the model to produce output in a specific format or style — without needing to fine-tune the model itself. Compared to the "Zero-Shot" approach of giving only a single instruction, few-shot prompting significantly reduces structural errors in the output. Claude's Projects feature is a natural fit for this kind of persistent context — uploaded files remain available as references throughout the entire project's conversation lifecycle, so you don't need to re-paste examples every time you start a new chat, making reuse far more efficient.

Submitting Your Request: One Sentence, One Complete Workflow

Once everything is set up, simply describe your desired automation in natural language. The example in the tutorial is: "Create a short-form video automation that uses Veo 3 to generate videos for my niche and uploads them to YouTube or TikTok."

Claude returns a complete JSON workflow along with a detailed guide covering the entire pipeline — from concept and script generation, image generation, voiceover and music, all the way to final video rendering and multi-platform publishing. In other words, one sentence of requirements yields a structured, actionable automation plan.

The Value and Limitations of This Approach

The biggest appeal of this method is how much it lowers the barrier to entry. For anyone unfamiliar with the underlying syntax of workflow tools, producing configuration files through conversation saves a tremendous amount of learning overhead. And the "upload examples, then generate" pattern is a universally applicable technique for getting reliable structured output from large language models — one worth adapting to any scenario that requires structured output.

That said, it's important to be clear-eyed about the limitations. AI-generated JSON workflows still need to be imported into the actual tool and validated. Node configurations, API keys, trigger conditions, and other details typically require manual review and debugging. Tasks like Veo 3 video generation and multi-platform auto-publishing are also subject to each platform's API permissions and terms of service. The tutorial demonstrates the generation step — but actually running a complete end-to-end automation involves significant follow-up work that shouldn't be underestimated.

Overall, this tutorial offers an efficient starting-point approach: treat AI as the "first-draft writer" of your workflows, use existing examples to constrain output quality, and then have a human handle final validation and deployment. For creators who want to quickly experiment with AI automation, it's a shortcut well worth trying.

Veo 3 is a video generation model from Google DeepMind that can generate short video clips — complete with native audio including dialogue, sound effects, and background music — directly from text prompts. It's one of the most capable text-to-video models publicly available. However, Veo 3 API access is currently provided through Google's Vertex AI or Flow platform, with regional restrictions, quota application processes, and usage-based pricing. Integrating it into an n8n or Make workflow also requires handling OAuth authentication, API call formatting, and the asynchronous retrieval logic for video files — none of which an AI-generated draft JSON can automatically resolve. Developers will need solid API debugging skills to make it work end-to-end.