Bug Issue の再現手順を Claude Code Skill と Playwright MCP で自動化 This article describes how to combine Claude Code Skills with the Playwright MCP server to…

Manually following the reproduction steps written in a GitHub bug issue is a tedious but unavoidable part of triage and QA work. This article walks through an experiment that offloads that work to Claude Code by combining its Skill feature with the Playwright MCP server, letting an LLM read an issue and drive a real browser to verify the reported behaviour.

Claude Code Skills are reusable bundles of instructions and context that Claude can invoke on demand for specific tasks. Here, the author defines a Skill whose job is to ingest a bug issue, parse the reproduction steps from the body, and then execute them through Playwright MCP. The Playwright MCP server, maintained by Microsoft, exposes Playwright's browser automation primitives such as navigation, clicking, typing and screenshotting via the Model Context Protocol, so any MCP-aware client like Claude Code can drive a browser in natural language.

In practice, the workflow appears to take an issue URL, extract the ordered steps and expected results, and then call the Playwright MCP tools one by one. The agent compares the observed behaviour with the expectation in the issue and reports whether the bug reproduces. The benefit is concrete: developers no longer need to click through the same flow repeatedly when verifying a report or checking a fix, and the same Skill can be reused across projects.

This pattern fits into a broader industry trend of letting LLM agents operate browsers directly. Anthropic's own Computer Use, OpenAI's Operator, and agent-style coding tools such as Cursor and Devin all attempt similar autonomy, though each draws the boundary between deterministic scripts and model-driven exploration differently. Playwright MCP is interesting because it keeps the underlying automation layer stable and battle-tested while letting the model focus on intent.

Skills themselves are a relatively recent addition to Claude Code and are likely to become a common way to package team-specific runbooks, from test automation to operational procedures. One caveat worth noting is reliability: natural-language-driven browser automation can be brittle when UI selectors are ambiguous or when steps in an issue are under-specified, so this kind of setup may work best as an assistive first pass rather than a full replacement for written E2E tests. Still, as a way to shorten the loop between a filed bug and an initial reproduction attempt, the approach looks promising and is straightforward enough that other teams could adopt it with modest effort.