Why Claude Code Skips Tests (And How to Fix It)

You give Claude Code a task. It writes clean, working code — feature complete, well-structured, maybe even elegant. Then you check the test file: empty, or one trivial assertion.

🧠 The Completion Bias Problem

Claude Code has a completion bias — it rushes to finish the most obvious goal first. "Build X" means build X — tests are a secondary task, and you have to ask for them. Add the pressure to keep responses short (tests can triple output length) and the AI wraps up early every time.

👀 What Skipping Tests Looks Like

🚧 The Prompt Engineering Dead End

Adding "write tests" to your prompt gets you from zero tests to maybe three shallow ones. Being more specific helps a little more, but you're now spending mental effort specifying test requirements for every single task — defeating the purpose of an AI assistant.

✅ The Real Fix: Rules That Apply Every Time

The solution isn't better prompts — it's permanent rules that apply every time, regardless of what you ask. Claude Code calls these skills — reusable instruction files that change how the AI behaves.

# Testing Protocol

For EVERY code change:
1. Write tests BEFORE or ALONGSIDE implementation
2. Cover: happy path, error cases, edge cases, boundary values
3. Test actual behavior, not mocks
4. Minimum: one test per public function/method
5. Run all tests. If any fail, fix before completing.

NEVER mark a task as done with failing or missing tests.

Now the AI treats testing as a mandatory checkpoint, not an optional extra. It can't declare the task complete until tests pass.

🏗️ The Layer Architecture Approach

Testing instructions work even better as part of a layered system. In the 8-layer execution protocol, testing sits as a checkpoint between writing code and finishing up:

1. Deep Context — Read and understand the codebase
2. Architecture — Plan structure before coding
3. Implementation — Write the code
4. Testing — Write exhaustive tests (the mandate lives here)
5. Edge Cases — Hunt for what was missed
6. Security — Check for vulnerabilities
7. Verification — Run everything, confirm it works
8. Documentation — Document what was built

The AI can't skip testing because the later steps depend on it. Edge case analysis needs tests to check against. Security scanning needs a test setup to probe.

📊 What Good AI Test Coverage Looks Like

Before (Default Behavior)

// auth.test.js
describe('auth', () => {
  it('should login successfully', async () => {
    const res = await login('user', 'pass');
    expect(res.status).toBe(200);
  });
});

After (With Testing Protocol)

// auth.test.js
describe('auth', () => {
  describe('login', () => {
    it('returns token for valid credentials', async () => { ... });
    it('returns 401 for wrong password', async () => { ... });
    it('returns 401 for nonexistent user', async () => { ... });
    it('returns 400 for missing email', async () => { ... });
    it('returns 400 for missing password', async () => { ... });
    it('returns 429 after 5 failed attempts', async () => { ... });
    it('locks account after 10 failed attempts', async () => { ... });
    it('handles SQL injection in email field', async () => { ... });
    it('trims whitespace from email', async () => { ... });
    it('is case-insensitive for email', async () => { ... });
    it('rejects expired passwords', async () => { ... });
  });

  describe('token validation', () => {
    it('rejects expired tokens', async () => { ... });
    it('rejects malformed tokens', async () => { ... });
    it('rejects tokens with wrong signature', async () => { ... });
    it('refreshes tokens within grace period', async () => { ... });
  });

  describe('logout', () => {
    it('invalidates the session token', async () => { ... });
    it('clears refresh tokens', async () => { ... });
    it('returns 401 for already-logged-out user', async () => { ... });
  });
});

Not two tests versus twenty — a total shift in what the AI considers "done."

📶 Three Levels of Testing Discipline

💡 The Deeper Principle