Claude Code Agentic Development Skills

A set of 8 global skills for Claude Code that enforce disciplined, test-driven agentic development. Install once, use in any project.

What This Solves

Without skills, AI coding agents write tutorial-quality code. They use brittle CSS selectors, skip tests, claim "done" without evidence, and guess at bug fixes. These skills fix that.

Problem	What the skills do
Agent writes fragile page.locator('.btn') selectors	Forces getByRole() locators that survive redesigns
Agent skips E2E tests ("unit tests already cover this")	/start makes Playwright E2E mandatory — literally cannot skip it
Agent claims "done" after writing code	/verify-done requires fresh test output as proof
Agent guesses at fixes and goes in circles	/debug forces 4-phase root cause analysis before any fix
No visual regression testing	toHaveScreenshot() with baselines, masking, CI font rendering
Tests break in CI but pass locally	CI pipeline patterns with sharding, artifacts, Docker for consistency
Agent writes "page loads" smoke tests and calls it coverage	Quality gates reject smoke-only coverage — demands real user workflow tests
Agent test.skips its way to green	Skipped tests count as failures — gates + a CI grep reject permanently-skipped tests
Agent fakes auth (.or(signIn) assertions, skip-guards)	Real storage-state auth or stop-and-ask — feature tests must run signed in
Rules obeyed only while the prompt is on screen	/start scaffolds GitHub Actions CI on day one — the suite runs even when nobody remembers
Having to invoke a skill every single time	Opt-in hooks put the rules in every prompt automatically, and run your tests before "done" is allowed
No screenshots or traces for debugging	Screenshots every test, video on failure, traces on retry — all automatic

What's included

• 8 orchestrated skills that route tasks through the right pipeline
• 17 deep-dive reference guides (6,400+ lines) covering locators, auth, fixtures, mocking, visual regression, screenshots, CI/CD, debugging, flaky tests, Next.js, POM, test data, clock mocking, iframes, API testing, and test organization
• Production-tested Playwright patterns adapted from the TestDino Playwright Skill
• Autonomous loop testing — write tests, let the agent iterate until green, walk away

Skills

Skill	Command	What it does
Start	/start	Master orchestrator — routes tasks through the full pipeline with guaranteed test infrastructure and E2E coverage.
TDD	/tdd	Enforces strict red-green-refactor. No production code without a failing test first.
Systematic Debugging	/debug	4-phase root cause investigation before any fix attempt.
Verification	/verify-done	Requires fresh evidence before any completion claim.
Brainstorming & Planning	/brainstorm-and-plan	Design-first workflow with specs and implementation plans.
E2E Playwright	/e2e-playwright	Battle-tested Playwright patterns, locators, fixtures, and debugging.
Test Automation Loop	/test-loop	Autonomous test-fix iteration — write tests, let the agent loop until green.
Code Review	/code-review	Two-stage review: spec compliance, then code quality.

Install

git clone https://github.com/burhankhatri/e2e-testing.git
cd e2e-testing
bash install.sh

This installs all skills globally to ~/.claude/skills/ and sets up the orchestrator CLAUDE.md at ~/.claude/CLAUDE.md.

If you already have a ~/.claude/CLAUDE.md, the installer offers to append just the rules and routing tree (about 20 lines, from claude-md-essentials.md) to the end of it — your existing content is never changed or removed, and a backup is written first. Say no and it prints the one-line command to do it later.

Don't skip this part. CLAUDE.md loads in every session with nothing to invoke, so it's what actually routes work into the skills; without it they rarely fire on their own.

Enforcement hooks (optional, recommended)

Skills are prompts — they only apply when they load, and the model grades itself against them. Hooks are run by the harness instead, so they hold even when nobody remembers to invoke anything. The installer copies two into ~/.claude/hooks/ and asks before enabling them:

Hook	Event	What it does
tdd-remind.sh	UserPromptSubmit	Carries the TDD rules into every prompt, so you never have to type a command to get them
tdd-verify.sh	Stop	Runs your test suite before work can be called done. Failing or skipped tests send the agent back to fix them

Both stay silent when no code changed and when a project has no runnable suite, so ordinary conversation costs nothing. Enabling them merges into ~/.claude/settings.json (backed up first, never overwritten) and needs jq. Re-run bash install.sh any time to turn them on later; running it twice won't duplicate anything.

Turning them off:

bash ~/.claude/hooks/toggle.sh off      # switch off — nothing is deleted
bash ~/.claude/hooks/toggle.sh on       # switch back on
bash ~/.claude/hooks/toggle.sh status   # check which it currently is

This works even after the repo itself is gone — it just drops a marker file in ~/.claude/hooks/ that both scripts check before doing anything else, so toggling never touches settings.json and can't corrupt it. off is instant and total: no reminder, and Claude is never blocked from finishing, no matter what your tests say.

If you already have a ~/.claude/CLAUDE.md, the installer will warn you and skip overwriting it. You can merge manually or replace it.

Usage

Once installed, skills are available in any project:

/start <task description>

That's it. /start handles everything — it routes your task through the correct skill chain, creates testing.md if it doesn't exist, guarantees Playwright E2E tests, and won't claim done without verification.

What /start does

1. Creates testing.md + CI — auto-detects your project's test setup, writes a testing guide, and scaffolds a GitHub Actions workflow so the suite runs on every push (if they don't exist)
2. Routes the task — feature, bugfix, refactor, or e2e-only each get the right skill chain
3. Runs TDD — every implementation step follows red-green-refactor
4. Writes E2E tests — Playwright tests for every user-facing flow (mandatory, not optional)
5. Verifies — runs the full test suite and shows evidence before claiming done

Individual skills

You can also invoke skills directly:

• New feature — /brainstorm-and-plan then /tdd
• Bug fix — /debug then /tdd
• Refactor — /tdd (tests for existing behavior first, then refactor)
• E2E tests — /e2e-playwright
• Autonomous iteration — /test-loop
• Review — /code-review
• Before claiming done — /verify-done

How It Works

The /start skill orchestrates the full pipeline:

/start <task> → testing.md → Route → TDD → E2E → Verify

Routing examples:

• /start add user auth → brainstorm → tdd → e2e → verify
• /start fix login loop → debug → tdd → e2e → verify
• /start extract auth middleware → tdd (existing behavior) → refactor → e2e → verify
• /start add checkout e2e tests → e2e → verify

Philosophy

• Tests over prompts — if the agent is going in circles, it needs a test, not more instructions
• Evidence over claims — no "should work", no "looks correct", run the command and show the output
• Root cause over quick fix — trace the bug to its source, don't patch symptoms
• Design before code — even "simple" projects get a design review

License

MIT