Last week, I came across a n8n hack so bad, my eyes rolled to the back of my head. I knew I had a new blog topic. 

It was an Instagram reel (I know…) in which the creator fed a screenshot of a workflow in Claude and asked it to generate a working JSON for the workflow. It sounds like a great hack: if you tell an LLM exactly what you want, it gives you best results. 

I am not against using LLMs to learn automation, but there are multiple reasons why this ‘hack’ is pretty much useless. 

In this blog, I’ll explain why it doesn’t work, and the correct way to use LLMs to learn automation.

Why the Hack Doesn’t Work

Reason 1: A screenshot of a workflow doesn’t have enough context.

Incomplete context = poor results.

Think of n8n like a big Lego spaceship. It is built with multiple blocks, which are called ‘nodes’. 

A screenshot of a workflow is like a photo of a finished Lego model - it shows you which nodes exist and how they’re connected, but hides what’s inside each node — the credentials, parameters, filters, expressions, and little toggles you set.

Without knowing this, you’ll have a workflow that looks right, but doesn’t work.  can’t make a workflow work. 

Example: you screenshot a lead gen workflow which is supposed to scrape linkedin profiles to add to a list and then qualify them based on certain parameters before sending a custom outreach message. If you just copy the screenshot, you’ll have no idea which fields are being scraped (job title, company size, industry, seniority, tech stack), how they’re scored to qualify leads, or which variables are driving the personalization in the outreach — and that’s the actual engine of the workflow.

Reason 2:  LLMs are not perfect with code, even if it’s just JSON 

JSON is generally considered easier than other data formats and, in some ways, easier than some programming languages. But while the current AI models can reliably output small, well-structured JSON objects, building complete workflows demands end-to-end logic and complex orchestration, which they haven't mastered. 

Reason 3: Credentials & other prerequisites 

Setting up credentials for every node in the workflow is time consuming, and often more complicated than configuring the nodes themselves.

Some platforms simply don't play well with others - X (Twitter's) API is notoriously difficult to work with, and their API scopes change frequently, breaking existing integrations.

Beyond credentials, workflows often require extensive prerequisites before they'll function. You might need to create database tables in Supabase, configure Google Sheets with specific column structures, set up webhooks in third-party services, or establish proper permissions across multiple platforms. 

The video made it seem like getting the perfect JSON is all you need. This is very misleading in terms of the actual work that goes behind setting up an automation. 

A Better Way to Learn With AI

Just because Claude can’t one-shot your workflow doesn’t mean you can’t save HOURS with the help of LLMs. 

I usually have a tab open while building workflows. 

It helps brainstorm different paths. It writes small snippets of JSON flawlessly. But it hallucinates. A LOT. 

At time of writing this, the knowledge cut-off for Claude Sonnet 4.1 is March 2025. Since n8n releases frequent updates with new nodes, features, and API changes, Claude might suggest node configurations that no longer exist or miss entirely new capabilities that could simplify the automation.

So it ignores any evolutions on the n8n platform, and there are MANY. Hence leading to a lot of hallucination. 

GPT-5 has a knowledge cut-off in 2024, so it’s even worse. 

Here’s the thing though. AI won’t build your workflows for you, but it can be a powerful ally. 

After building dozens of workflows with AI in the loop, I’ve figured out how to leverage its strengths while working around its blind spots.

I swear by this system, and I use it for every single workflow I build. 

Before beginning: 

Start by getting a rough structure of your workflow. Describe your use-case to the LLM with as much context as possible. Don’t get attached to this first draft, things will change.

If the model dumps a giant blob of JSON on your screen, ignore it. Tell it explicitly not to write JSON until you ask.

Use this prompt to get best results:

I want to build an n8n workflow for [specific use case]. Here's the context:

- What I'm trying to achieve: [detailed description]

- Data sources: [list sources]

- Expected outputs: [describe outcomes]

- Any constraints: [timing, data limits, etc.]

Please help me break this down into logical steps.

DO NOT write any JSON code yet. 

I only want you to help me think through the logic and structure. We'll build this step by step later.

During building: 

Once you have a rough plan of the workflow, it’s time to start building. LLMs get overexcited and start outlining the execution steps at once. But you need to take it methodically, one node at a time, otherwise your workflow will break. 

There are two parts to this:

Part 1: Use node documentation as the source of truth

Before asking the LLM to guide you through configuring a node, copy-paste the official documentation for that specific node. Documentation gets updated frequently to reflect platform changes. In my experience, this single step makes the LLM’s guidance nearly 95% accurate.

Node Documentation Prompt:

I'm about to configure the [Node Name] node. Here's the official documentation:

[PASTE DOCUMENTATION HERE]

Please read this documentation thoroughly first. Then explain how to configure this specific node for my use case. Do not start explaining until you confirm you've reviewed the docs I provided.

Step-by-Step Configuration Prompt:

Now let's configure the [Node Name] node. Based on the documentation I provided:
1. What fields do I need to fill out?
2. What values should I use for my specific use case?  
3. Are there any common gotchas I should watch out for?

Only focus on THIS node - don't jump ahead to other steps.

Part 2: Test as you go 

Every node is required to give you a specific output, which in turn becomes to input for the next node. Always confirm if it is working correctly before moving on to the next step. 

Testing Prompt:

Before we move to the next node, let's test this one. Explain how to test this node in n8n so I can confirm it’s behaving as expected before we connect the next one. Explain what success should look like.

You don’t want to end up with all the nodes configured only to realise the workflow is not working. 

Use Screenshots for Visual Debugging

When something doesn't work, share a screenshot with Claude. Visual context helps it diagnose issues far more effectively than trying to describe error messages or interface problems in text.

Debugging Prompt:

Something isn't working with this node. I'm attaching a screenshot of the error/issue. Please analyze what you see and suggest specific troubleshooting steps.

[ATTACH SCREENSHOT]

At the end: 

Once your workflow is running successfully, ask Claude to create comprehensive documentation. This step is crucial—it transforms your one-off build into a reusable reference for similar projects.

Comprehensive Documentation Prompt:

Perfect! The workflow is working. Now please create a comprehensive summary including:

1. Workflow Overview: What this automation accomplishes
2. Node-by-Node Breakdown: Each node's purpose and key settings  
3. Troubleshooting Notes: Issues we encountered and how we solved them
4. Prerequisites: External setup required (credentials, databases, etc.)
5. Testing Steps: How to verify each part works

Format this as a reference guide I can use for similar projects in the future.

Lessons Learned Prompt:

Looking back at this workflow build, what were the key decisions we made and why? What alternatives did we consider and reject? This will help me make better choices on future workflows.

Warning: Never share API tokens, keys, or secrets with any AI model. This poses a major privacy and security risk.

In Conclusion…

Automation isn't going anywhere, and neither are the complexities that come with it. APIs will keep changing, platforms will keep evolving, and your workflows will keep breaking if you don't understand how they actually work. 

The Instagram hack represents everything wrong with how people approach learning automation—looking for shortcuts instead of building genuine understanding. 

The method I've shared does the opposite: it uses AI to accelerate your learning while ensuring you understand every piece of your workflow. 

Master this approach, and you'll be building robust automations long after the next shiny "hack" comes and goes.

If you'd rather outsource the building to someone who's regularly in the weeds, I work with businesses as a Fractional Automation Expert, and I'm currently taking on new clients. Let's discuss how the right automation strategy can give you back 10-20 hours per week to focus on what actually grows your business.