Writing
Which technique finds which bug, how to catch a requirement that cannot be tested, and where AI earns its place in a QA workflow.
A spec in a PDF is not a list of test cases, and most of the work sits between the two. Six steps for getting from one to the other, by hand or with a generator, and where the process breaks down.
One-line prompts return test cases that read well and cannot be run. Here is a prompt structure that fixes most of that, what each part is for, and the failures no prompt removes.
Microsoft now ships two ways for an AI agent to drive Playwright: an MCP server and a command-line tool with skills. Both open the same browser. They differ in what lands in the model's context, and that decides which one fits a coding agent.
Since version 1.56, Playwright ships three agents that plan, write, and repair end-to-end tests for you. What each one takes in and hands back, how to set them up in Claude Code or VS Code, and which step still needs a person.
A test case template is only as good as the fields it forces you to fill. This one has ten, six worked examples for a login page, and a flat layout that imports into TestRail, Qase, Zephyr, and Xray without a cleanup pass. Excel and CSV, free to download.
The plan you approve in Claude Code plan mode is the clearest statement of what a feature should do, and most teams delete it once the code ships. This workflow keeps it: plan, build, turn the plan into test cases, then let Playwright MCP write E2E tests against those cases.
Two quick checks get used interchangeably and shouldn't be. One decides whether a build is even worth testing further, the other confirms a specific fix didn't break on delivery. Here's what separates them and when each one belongs in a pipeline.
More unit tests than integration tests, more integration tests than end-to-end tests: the pyramid shape exists for a cost reason, not a purity rule. Here's what each layer actually catches, and the projects where the classic shape is the wrong one.