The big picture
I grabbed my engineering friend Ramya to discuss Cursor. But we ended up talking about Claude skills instead. What can you do?
I told Ramya how I’ve used plugins and standalone skills in Claude within the recruitment processes (sourcing, research, hiring manager packets, etc) and asked her what she’s doing. She described one workflow she designed that: understand the feature, write test cases, automate them, file bugs, open PRs, review PRs, and triage nightly failures.
The workflow:
| Stage | What the skill does | Output | Human checkpoint |
|---|---|---|---|
| 1. Explore | Reads Jira tickets, product code, and any other available docs, launches the UI. | What I would describe as a full knowledge-transfer | Implicit |
| 2. Test design | Writes test scenarios from that understanding and pushes them (her in case to Notion) | create test cases | Explicit review point. Ramya flagged this as a key place she reviews. As a recruiter, this is where I’d probe a candidate: how do they review AI-generated test cases, and what do they look for? I wonder who is cancelling out productivity gain by spending too much time reviewing AI output.. |
| 3. Automation | Separate skills for UI and API as an example. Reads the test cases, writes Playwright tests, runs them, and self-iterates on failures until they pass | Passing automated tests | Review the tests to confirm they check what you intended |
| 4a. Bug path | If the run exposes a real product defect, a separate skill logs the bug with full details | Bug ticket | Verify it’s a real bug |
| 4b. Happy path | If all is green, a skill opens a PR automatically | Pull request | See next stage |
| 5. PR review | The PR is posted to the QA channel, and a reviewer skill (another bot) reviews it and leaves comments | Review comments | Human reads the comments and judges whether they make sense before approving |
| 6. Nightly test review | A skill analyzes the nightly run to identify what failed and why (that’s some work off the plate and in my opinion it’s a simple but good example of how QE is changing) | Failure analysis | Engineer acts on findings |
I asked how her team handles skills. Her answer: everyone uses the same ones, and she recommends it as a best practice for standardization and consistent quality.
The key hiring signal: human-in-the-loop
I told her everyone describes their workflows with “human in the loop” and she suggested asking a candidate to describe an example of an agent or workflow they built and just listen to how they describe it. Her example is the best screening anchor. She gave me a simple example of something that happened today she had Claude running automated tests and it told her the task was completed successfully and everything was saved. But the data never persisted and she only found out because she went to check. Wouldn’t have caught the bug without having the check point.
The red flag: a candidate describing an AI workflow with no specific mention of where they review. Her rule of thumb is that if they’re signing off, they need to have looked.
Suggested screening questions
- “Walk me through a workflow where you use AI for testing. Where do you review, and what do you look for?”
- “Tell me about a time the AI got something wrong. How did you catch it?”
- “What skills, plugins etc have you built? What did they do?”
- “How does your team share and standardize these skills?”
On tool-specific experience: Since Ramya has worked across Copilot, Cursor and Claude I asked her about the differences. She mentioned they’re fairly interchangeable. She said for example if you are working on a role that requires Cursor and you find a candidate who has the exp you need but from an AI perspective has only used Copilot, then don’t rule them out.
