AI for robotframework cases

Hi!
I have around 400 API robotframework cases, and it’s quite tempting to teach an AI to generate new ones via prompts. Has anyone had experience working with AI or automation tools for this kind of task? Where should I start, and are there any recommended approaches?

I’ve done big auto generation of actual robot suites that generated about 2.5k testcases from few excel files and i’ve used AI to generate code for few projects and my personal opinion is that AI will not add anything a capable design and handwritten code would do – unless you actually can prompt with proper design to write the design/architecture in mind and you can prompt it to have both test inputs (endpoints / input data / testase and suite setups) and end results (teardowns, endpoint replies) from the data you already have..

If not, i would just write a some code that that generates the basic testcases/suites by hand and then cover the corner cases by manual work (by not auto genration but manually written testcases/suites)..

1 Like

I was also wondering, especially how it looks in terms of AI assisted IDE, does anyone has some experience (e.g. with Cursor). We have tons of keywords, could be useful to have something that is generating some good working base TC for just small improvements. Which avaialable option woudl be good and safe.

We are having success with vscode and copilot (agent mode with Claude Sonnet 4) for building / fixing API and GUI tests, a few things to look into:

  1. Have the agent create an instructions.md file that explains your repo structure, tech stack, coding practices, etc. This will help keep the agent focused.
  2. Leverage MCP servers for more context. Context7 is very helpful for pulling latest docs for either popular or obscure libraries. We use Atlassian MCP to pull context from tickets as well as some custom ones built to better get context from files or systems we use.
  3. Create / Generate instructions to make the process repeatable and more predictable. Define the creation process you want it to follow (1 test at a time, test all code, debug through failures, wait for user confirmation before proceeding to a new test, etc). Try your instructions, see where it fails, have the agent edit the instructions to make sure that doesn’t happen again, try again, repeat.
  4. Sometimes providing specific examples of structure in your instructions can go a long way if you want things done a certain way each time. API is helpful here as the structure is fairly rigid without a lot of deviation. For GUI I found that having examples for how I want it to do in-place debugging to be really helpful to have it waste less turns figuring it out itself.

Best way to start is to just have the agent make instructions for you, just be sure to provide as much context of what you want to do and iterate on it over and over till it gets to where you want.

I have very little programming experience but using ai I was able to construct a synthetic monitoring login test to a desktop application for our monitoring system.