How to Build Your First OpenClaw Workflow From Scratch

Quick Answer

Building your first OpenClaw workflow means giving your agent a clear goal, the right skills, and a well-written SOUL.md file that defines its personality and behaviour. Start with something simple — like a daily news summariser or an email drafter — then layer in complexity as you get comfortable. The whole process takes about 30 minutes for a basic workflow, and you don't need to write any code.

What Is a Workflow in OpenClaw?

A workflow in OpenClaw isn't a rigid flowchart like you'd build in Zapier or Make. It's more like giving a capable assistant a job description and a set of tools, then letting them figure out the best way to get the job done.

The "workflow" is really the combination of three things: your agent's SOUL.md file (its instructions and personality), the skills you've installed (its tools and capabilities), and the model powering it (its brain). Together, these define what your agent does, how it does it, and how smart it is about doing it.

This is fundamentally different from traditional automation. Instead of mapping out every if-then-else branch, you describe the outcome you want and let the AI figure out the steps. It's more flexible, handles edge cases better, and doesn't break when something unexpected happens.

Step 1: Define What You Want Your Agent to Do

The most important step is the one people skip. Before you touch any settings, write down in plain English exactly what you want your agent to accomplish. Be specific about the outcome, not the process.

Good examples: "Every morning at 8am, check the top AI news from the last 24 hours and send me a summary on Telegram." Or: "Monitor my inbox for invoices and save them to a specific folder with the amount and sender extracted." Or: "When someone messages me on WhatsApp with a question about my business, draft a helpful reply for me to approve."

Bad examples: "Be a helpful AI." Or: "Do stuff with emails." The more specific your goal, the better your agent will perform.

Step 2: Choose and Install the Right Skills

Skills give your agent its abilities. Based on what you want your workflow to do, you'll need to install the relevant skills from ClawHub. OpenClaw has over 13,700 skills available, each one an MCP server that connects your agent to a specific tool or service.

For a morning briefing agent, you'd install web browsing and search skills. For an email agent, you'd install email and file management skills. For a customer service agent, you'd install the relevant messaging platform skill plus any knowledge base or FAQ skills.

Start with just the skills your workflow actually needs. Adding too many skills can confuse the model — it's like giving someone a toolbox with 500 tools when they only need a screwdriver. You can always add more later as you expand your workflow.

Step 3: Write Your SOUL.md File

The SOUL.md file is the single most important part of your workflow. It's the instruction manual for your agent — written in plain English, it tells the agent who it is, what it should do, and how it should behave.

A good SOUL.md has a few key sections. Start with a role description: "You are a morning briefing assistant that helps Ewan stay informed about AI industry news." Then add specific instructions: "Every morning, search for the top 5 AI news stories from the past 24 hours. Summarise each in 2-3 sentences. Focus on practical developments, not hype. Send the summary via Telegram by 8am."

Include guidelines about tone and style: "Write in a casual, friendly tone. Be concise. If a story isn't genuinely important, leave it out rather than padding the briefing." And add any constraints: "Never include cryptocurrency news. Always include links to the original articles."

The SOUL.md is where you shape your agent's personality and judgement. Think of it like writing a brief for a new team member — the clearer you are, the better they'll perform from day one.

Step 4: Choose Your Model

The model is your agent's brain. For simple workflows with straightforward tasks, a smaller model like Qwen 3.5 14B running locally via Ollama is perfectly capable. For more complex workflows that require nuanced judgement, research, or multi-step reasoning, a larger model like Qwen 3.5 72B or Llama 4 Scout on Tulip will give better results.

If you're not sure, start with a cloud model on Tulip. You can always switch to a local model later once you've confirmed the workflow works as expected. It's easier to debug a workflow when you know the model isn't the bottleneck.

Step 5: Test With a Simple Task

Before you set anything to run automatically, test your workflow manually. Give your agent the task through the OpenClaw interface and watch what it does. Does it use the right skills? Does it produce the output you expected? Is the tone right?

This is your debugging opportunity. If the agent calls the wrong tools, your SOUL.md needs to be more specific about which tools to use when. If the output quality is poor, you might need a more capable model. If it misunderstands the task entirely, your instructions need to be clearer.

Run the test three or four times with slightly different inputs to make sure it handles variation well. Agents are probabilistic — they might take slightly different paths each time, which is usually fine as long as the outcome is consistently good.

Step 6: Set Up Triggers and Scheduling

Once your workflow produces good results manually, set it up to run automatically. OpenClaw supports several trigger types: scheduled triggers that run at specific times (great for daily briefings), message triggers that activate when someone sends a message (great for customer service), and event triggers that fire when something happens in a connected service.

For a morning briefing, set a daily schedule. For a customer service agent, enable the messaging channel trigger. For a monitoring agent, set an interval — every hour, every 30 minutes, whatever makes sense for your use case.

Step 7: Refine and Expand

Your first version won't be perfect, and that's fine. Run your workflow for a few days and note what works well and what doesn't. Update your SOUL.md to address any issues. Add or remove skills as needed. Adjust the model if the quality or speed isn't right.

Once your basic workflow is solid, you can start layering in complexity. Add a second step that processes the output of the first. Connect additional skills. Create conditional behaviours in your SOUL.md ("if the news is about a competitor, flag it as high priority"). Build memory so your agent learns your preferences over time.

Example: Building a Daily Research Briefing

Let's walk through a concrete example. Say you want an agent that delivers a daily summary of what's happening in your industry.

Install the web search and web browsing skills from ClawHub. Write a SOUL.md that describes the agent's role, the topics to monitor, the format you want the summary in, and where to deliver it. Connect a Telegram skill for delivery. Set the model to Qwen 3.5 on Tulip for reliable performance. Test it manually a couple of times, tweak the SOUL.md until the summaries are the right length and focus, then set it to run every morning at 7:30am.

Total setup time: about 30 minutes. Ongoing maintenance: almost none, unless you want to change what topics it monitors.

Common Mistakes to Avoid

Writing vague SOUL.md instructions is the number one cause of poor agent performance. Be specific. Instead of "research the topic thoroughly," say "find 3-5 recent articles from reputable sources, summarise each in 2-3 sentences, and highlight any conflicting information."

Installing too many skills at once is the second most common mistake. Each skill is another option the model has to consider, and too many options lead to confusion. Start minimal and add as needed.

Using a model that's too small for your task is the third. If your agent consistently produces poor output or makes wrong decisions, try a larger model before you rewrite everything else.

Frequently Asked Questions

Do I need to know how to code?

No. OpenClaw workflows are configured through the interface and SOUL.md files, which are written in plain English. The most technical thing you'll do is install Docker, and even that is a one-time setup.

How many workflows can I run at once?

As many as your hardware or Tulip plan supports. Each workflow is an independent agent with its own configuration. On Tulip, you can scale to dozens of concurrent agents.

Can I share my workflows with others?

Yes. Your SOUL.md file and skill configuration are portable. You can share them with the OpenClaw community or within your team. The person receiving them just needs their own OpenClaw instance and model access.

What happens if a workflow fails?

OpenClaw logs all agent activity. If a workflow fails, you can check the logs to see what happened, which skill failed, and what the model was trying to do. Most failures come from external services being unavailable or SOUL.md instructions that don't cover an edge case.