5 resources to learn AI automation in 2026

The fastest way to make an automation course useful is to give yourself one real process to improve while you learn. You’ll understand triggers, data and error messages much faster when they belong to a task you recognise.

Here’s a practical route through the five resources in the carousel. You can choose Make or n8n as your main visual platform, then use the other resources to understand connections, reusable Claude instructions and system design. You don’t need subscriptions to every tool on the list.

Our practice project is an enquiry workflow. A new enquiry arrives, required information gets checked, the complete record is saved and anything unclear goes to review. Later, AI can draft an internal summary. We’ll keep external replies under human approval while learning.

Before you start

Use a practice workspace with sample enquiries rather than customer information. Write down the minimum useful output: a saved record containing the enquiry ID, name, email, request and review status.

Learning pages are generally available to read or watch, but running a workflow can require a platform account, hosting and paid model API usage. A Claude subscription and an API account are different access routes. Check the current pricing and limits for the specific environment you choose before enabling scheduled runs.

1. Make Academy: learn how data moves

Open Make Academy. Its free, self-paced learning is a good starting point if you want to understand automation on a visual canvas. Begin with the fundamentals rather than jumping straight into an elaborate AI-agent example.

As you work through the lessons, connect each concept to your enquiry project. A trigger starts the process. A module performs a step. Mapped fields pass the incoming information into the next step. A filter decides whether that record should continue along a route.

Your first exercise is deliberately small: take one sample input and save it to a test destination. Inspect the name and email in both places. Then add the required-field check.

You’re ready for the next stage when you can point to where each saved value came from. If the output contains an empty field, inspect the incoming sample and the mapping before adding more steps. Course access is free; workflow execution has separate account limits.

2. n8n’s first-workflow guide: inspect each step

Open the official first-workflow guide. It covers a concrete workflow with a trigger, credentials, expressions and a branch. Follow its published example first, then adapt the pattern to your enquiry project. The learning paths provide the next lessons.

You need an n8n environment. Cloud offers a trial; self-hosting involves setup, maintenance and hosting resources. Choose the option you can actually support.

In your practice version, keep the input visible and inspect the data after every node. Build the required-field branch before introducing AI. An empty email address is a straightforward condition, so it does not need a model to decide whether it is missing.

Give every enquiry a stable ID. When you rerun a failed test, check that it updates or recognises the same record rather than creating duplicates. This is a design choice you must implement, not something a diagram automatically guarantees.

3. Nick Saraev’s skills course: package the repeated instructions

Watch Claude Skills Full Course: Automate Your Work, then use Anthropic’s current skills documentation when you implement your own version. The video is the learning resource; product documentation is the reference for current setup details.

A skill stores reusable instructions and can include supporting files. For this project, its job could be turning a saved enquiry into a short internal brief. It does not itself create a form trigger or grant access to your customer database.

First, make the summary work manually with three sample enquiries. Specify the fields it should return: request, missing information, suggested next action and source enquiry ID. Tell it to use “not provided” when information is absent.

Once the result is useful, save those instructions in a project skill, such as .claude/skills/enquiry-brief/SKILL.md. Use the documented frontmatter and test it with a fresh example. Running Claude still requires suitable product access, and an API-based workflow has its own billing.

4. Postman: understand the request and response

Open the Postman quick start. Follow the public example exactly before connecting a private account. The aim is to recognise the request URL, method, parameters and returned data.

Use the exact public request shown in the quick start. Inspect the response and find where the returned values appear. Then repeat the documented example with one permitted input changed so you can see the request and response move together. Keep this exercise on public test data.

When you later connect a real app, look up its own documentation for authentication, required fields and error responses. A successful request to an echo service does not prove you have permission to use a different service.

If you receive an error, capture the status code and message without exposing credentials. Check the URL, request method and required fields before asking AI to diagnose it. That evidence makes troubleshooting much more specific.

5. Building Effective Agents: choose the structure

Read Anthropic’s Building Effective Agents. Use its diagrams to compare a fixed sequence, routing and a writer/reviewer loop. It is an architecture article, so use current product docs for installation instructions.

Apply the simplest pattern to the enquiry example. Missing fields follow a fixed rule. The summary can be one model call. If a reviewer is useful, give it explicit checks and a limited revision loop. You do not need an autonomous agent to perform every part.

Ask where the AI actually adds judgment. If you can express a decision as a dependable condition, implement that condition directly. Save model reasoning for the parts where interpreting language helps.

Turn the idea into a build plan

Paste this into Claude with your chosen platform and a few synthetic examples. It returns a specification you can implement alongside the lessons.

Design a small enquiry-processing workflow for [Make or n8n].

Input examples: [paste 3 fictional enquiries with stable IDs].
Required fields: enquiry_id, name, email, request.
Destination: [name a test table or practice destination].
Goal: save complete enquiries, route missing fields for review, and produce an internal summary. Do not send customer replies.

Return the exact trigger, each step, field mappings and branch conditions. Separate deterministic rules from the one optional AI summary step. Include a proposed output schema, duplicate-handling strategy, execution log fields and where errors go.

Give me 5 test inputs: complete, missing email, missing request, repeated ID and unusually long request. State the expected result for each. Explain which steps need credentials or paid services. If the destination is unspecified, use a local test table in the plan and say it is a placeholder.

Use current official documentation if browsing is available. Flag anything you cannot verify. Finish with the smallest working version I should build first.

Check the workflow before letting it run

Run the five test cases manually. A missing field should reach review, a repeated ID should follow your duplicate rule, and a failed step should be visible in the execution record. Check that the AI summary has not invented facts absent from the enquiry.

If all five tests pass, try a small supervised batch and measure how long review takes. If checking the output takes longer than doing the original job, simplify the output or improve the input rules. Your first win is one dependable handoff you understand well enough to repair.