Claude Code for QA: Skills, Agents and Workflows

Claude Code is Anthropic's agentic coding tool. Its overview says it "reads your codebase, edits files, runs commands, and integrates with your development tools." Three of its extension points matter most for QA work: skills, subagents, and plugins. Together they let a team hand Claude a testing method once, instead of re-explaining it in every prompt.

Testland's position: Claude Code is useful in QA when it carries a method a tester already trusts, such as a charter template, a boundary-value set, or a flake bisection, and much less useful as a generic "write me some tests" button. The value is in the method. The model applies it faster.

Skills, subagents and plugins in Claude Code

A skill is a folder of instructions. Anthropic's skills documentation puts it in two lines: "Skills extend what Claude can do. Create a SKILL.md file with instructions, and Claude adds it to its toolkit." Claude reads each skill's description to decide when it applies, and "Claude uses skills when relevant, or you can invoke one directly with /skill-name." Only the description sits in context until then: "a skill's body loads only when it's used, so long reference material costs almost nothing until you need it."

A subagent is a helper with its own working space. Per the subagents documentation, "Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions." Claude delegates to one when a task matches its description, or you ask for it by name.

A plugin is the unit you install. The plugins overview defines it as "a directory of skills, agents, hooks, MCP servers, or other components that Claude Code installs and loads as one unit," fetched from a marketplace, which is a catalog of plugins. Skills that come from a plugin are namespaced as /plugin-name:skill-name.

In Testland's marketplace, the split is practical. A skill teaches Claude one technique or one tool, such as risk scoring or Playwright authoring. An agent runs one focused job, such as reviewing a manual test script or bisecting a flaky test, and returns findings.

Installing a role bundle in two commands

Add the marketplace once, then install the bundle for your role:

/plugin marketplace add testland/qa
/plugin install qa-starter@testland-qa

qa-starter is the essentials bundle. It pulls in ten plugins that apply on any stack: test process, test data, test environments, test-code review, reporting, flake triage, bug repro, defect management, CI integration, and manual and exploratory testing. The role bundles build on the same idea for a specific job: manual tester, automation engineer, SDET, frontend, backend, mobile and desktop, security, performance, AI and data, and QA leadership. Every plugin in the full catalog also installs on its own with /plugin install <plugin>@testland-qa.

Outside Claude Code, npx skills add testland/qa --skill <skill> installs individual skills into other coding agents. That route carries skills only: agents and role bundles need Claude Code.

Eight QA workflows and the prompts that start them

Each workflow below starts from a plain-language request. Claude picks the matching skill from its description, so the prompt doesn't have to name it.

Scripted test cases from a feature

Write test cases for the password reset page

manual-test-script-author drafts scripted cases as a step table, Gherkin, a UAT script, or a checklist. To audit cases that already exist, test-script-quality-critic flags vague preconditions, bundled steps, and missing expected results.

A chartered exploratory session

Plan a one-hour exploratory session on checkout

exploratory-charter-author writes the mission, scopes the areas, and sets the time-box. The exploratory-testing skill carries the session sheet and PROOF debrief described in session-based test management.

A bug report someone can act on

The nightly checkout run failed. Write it up from the attached result file before standup.

bug-report-template fills in steps, environment, expected versus actual, and a severity rationale, and it converts a single failure record (JUnit XML, Allure JSON, a pytest log, or a Playwright report) into a ready-to-file bug.

A test that fails only on CI

Why does this test fail only on CI?

flake-axis-bisection varies one condition at a time (isolation, order, worker count, viewport, latency) and checks whether the gap exceeds sampling noise. The e2e-flake-bisector agent runs that loop, and flaky-test-quarantine takes the test off the gating path with an expiry date. The method behind it is in How to Fix Flaky Tests.

Playwright tests for a repo

Set up Playwright tests for this repo

playwright-testing scaffolds the project and writes tests with role-based locators and web-first assertions. playwright-codegen-reviewer turns a raw codegen recording into page objects and arrange-act-assert tests.

Edge cases for an input

The help center says avatars can be "up to 5MB" but the validator works in bytes. Give me boundary cases for it.

boundary-value-generator produces the six-point boundary set (one below, at, and one above each bound) plus equivalence-class representatives, emitted as parameterized test inputs for pytest, Jest, or xUnit.

An API change before release

Did this API change break any clients?

openapi-contract-diff runs oasdiff breaking between two spec versions and classifies each finding by severity. When a contract test is already red, contract-drift-investigator works out which side drifted.

Release risk for a QA lead

Which parts of this release are riskiest?

risk-matrix scores likelihood against impact and maps the top risks to test coverage. test-strategy-author carries those risks into a test strategy, and release-quality-report-agent assembles a GO, NO-GO, or CONDITIONAL release report from the evidence.

The D1 to D6 review every component passes

Every skill and agent in testland-qa is reviewed by hand before merge, against six dimensions published in the repository README:

Citations are the hard floor. The README calls uncited "citation theater" "the dominant failure mode and a hard reject." It's also explicit that the bar is "reviewer judgment," with no automated rating gate.

Where Claude Code stops in QA work

• Installed plugins cost context. For every skill and agent Claude can invoke on its own, "the name and description are in Claude's context on every turn," according to the plugins overview. Install the bundle for your role, not the whole catalog.
• A plugin runs as you. The same page is direct about it: "what the plugin runs, it runs as you." Read a plugin's components before installing it in a sensitive repository.
• Cloud sessions are separate. "A cloud session, including one in the browser at claude.ai/code, doesn't load the plugins in your local settings."
• A reviewed skill isn't a reviewed result. The D1 to D6 review checks the method a skill teaches, not each output it produces. A drafted charter, risk score, or test still needs a tester to decide whether it matches what users care about.

The judgment calls stay with the tester: what to explore, which risk matters, when a flaky test is worth fixing. The rest of Testland covers those calls in depth under exploratory testing, flaky tests, and Playwright in CI.