27 Agent Skills, One Chained Testing Pipeline: From Requirements to UI Reports
Agent skills have moved from isolated prompt tricks to a whole workflow: one chain can parse requirements, generate and review cases, run API and UI suites, diagnose failures, and produce reports without a human switching tools. Teams using Claude Code or Codex can adopt this pattern immediately because the skills directory convention is already supported.
27 self-built Agent Skills treat test automation as three chains rather than one-off prompts. The requirements chain turns messy input into structured user stories, XMind test points, and Excel case sheets, then reviews and fills gaps. API automation parses schemas from Swagger, Postman, HAR, or YApi into a shared definition file, generates and tags runnable Python+Requests+Pytest+Allure scripts, executes them from natural-language commands, diagnoses failures into environment/data/script/defect buckets, cleans up test data, and emits HTML reports with Allure integration. A separate scenario track models multi-step business flows like register, login, add-to-cart, order, pay as YAML orchestration files and executes them with variable injection and database assertions.
UI automation builds page objects from URLs or case descriptions, generates Playwright+POM+Pytest scripts, bolsters them with waits, captcha handling, iframe switching, retries, and visual assertions, then runs, tags, diagnoses, and reports under a single scheduler. Every skill is independently invocable, but the wiring between stages is the point: each stage emits a standard file the next stage consumes.
Distribution runs through GitHub and a skill store at testfather.cn/skills. Installation uses `npx skills add`, a git clone, or copying the skill folder into `~/.claude/skills/` or `~/.workbuddy/skills/`.
Standardized intermediate files make this more than a collection of prompts. Each link consumes a file produced by the previous link, which is what allows agent tools to chain steps reliably.
The system treats hallucination as an explicit risk: the API scenario modeler matches every endpoint against api_definitions.json, injects dynamic values rather than hard-coding them, and lists uncertain points for human confirmation. That is a practical answer to the main objection to AI-generated test code.
Distribution doubles as a product funnel: GitHub and a skill store offer the skills, while a paid knowledge community promises the full pack. The technical content is real, but the commercial ecosystem is part of the story.