Developer Builds Zero-Cost Self-Healing Playwright Tests Using AI and Accessibility Trees

A developer set out to create automated health checks for an e-commerce storefront using plain natural language test descriptions instead of code or selectors. Early attempts using a live LLM to drive Playwright on every test run proved costly, slow, and inconsistent due to API fees and rate limits. The key insight was that AI should generate the test once rather than re-interpret it on each execution, eliminating ongoing inference costs. Subsequent versions explored human-assisted element selection via a bookmarklet and, ultimately, the browser's accessibility tree, which reduced page data from megabytes of HTML noise to a compact, token-efficient snapshot. The final approach uses the accessibility tree to give the model just enough page context to generate a reusable Playwright script, enabling free, repeatable test runs on a cron schedule.