Skip to main content

I built xcepto-skill so AI agents can write better Xcepto tests

· 2 min read
Daniel Maß
Daniel Maß
Maintainer of Xcepto

Xcepto is still new, and that created an unexpected problem: AI agents were bad at writing Xcepto tests.

Not because the tests were impossible — but because the model had almost no reliable context. When I asked Codex to write Xcepto tests, it usually tried to infer the framework from whatever it could find. Sometimes it checked out Xcepto.NET somewhere temporary. Sometimes it guessed APIs through reflection. Sometimes it copied local fragments without really understanding the test model.

The result technically looked like tests, but they were not good Xcepto tests. They were hard to read, awkward to maintain, and often missed the main idea: Xcepto tests should describe observable behavior as a clear Given-When-Then flow.

The implementation detail is not the point. The behavioral contract is.

So I put the missing context directly where the agent can use it.

What the skill teaches​

The xcepto-skill explains how Xcepto tests are supposed to be structured:

  • scenarios own setup and cleanup
  • adapters expose readable domain actions
  • builders collect meaningful test data
  • states express lazy, condition-driven expectations
  • tests verify observable behavior instead of implementation wiring

The difference is very visible. Before, AI-generated Xcepto tests felt like generated code. Now they look much closer to tests I would actually write myself: readable, maintainable, and behavior-focused.

That matters a lot for a framework like Xcepto. If the test reads well, the system behavior is easier to reason about. And if AI agents can preserve that style, they become much more useful in real projects.

Originally published on themassiveone.net.