second-look

second-look

A Claude Code plugin for writing, auditing, and maintaining tests. Give your tests a second look before they earn their place.

Every test is a bet that its future signal will exceed its future maintenance cost. Write tests that win the bet. Delete the ones that don't.

Why "second-look"

Named for Andrew Look, whose advice on testing shaped the philosophy: signal over coverage, maintenance cost as the real constraint, integration-first, mock at boundaries, table-driven tests for combinatorial cases. A test audit is literally a second look at your suite. The practice the plugin encodes is the habit of giving each test a second look — against the five audit questions — before it earns its place.

What's in the plugin

Component	What it is	How it activates
test-quality skill	Full philosophy, five audit questions, patterns, smells, decisions	Auto-triggers when you talk about tests, coverage, mocks, flaky tests, etc.
test-auditor agent	Structured five-phase audit of an existing test suite	Invoked automatically for audit tasks, or manually via /agents
/second-look command	Explicit trigger to audit tests in a specific path	Type /second-look src/auth (or any path)

The CLAUDE.md snippet is included separately (CLAUDE-snippet.md) for users who want always-on testing principles in their project CLAUDE.md files. Most users won't need it — the skill covers the common cases.

Install

From GitHub (recommended)

# In Claude Code — two steps: register the marketplace, then install
/plugin marketplace add jmilinovich/second-look
/plugin install second-look@second-look

The repo doubles as a single-plugin marketplace, so /plugin install needs the <plugin>@<marketplace> form rather than a bare GitHub path.

Scope defaults to user (available across all your projects). Add --scope project to commit it to a specific project.

From a local directory

For testing before publishing, or if you'd rather keep it local:

# Clone/copy the plugin directory somewhere stable
cp -r second-look ~/plugins/second-look

# Start Claude Code with the plugin loaded for this session
claude --plugin-dir ~/plugins/second-look

Verify

# In Claude Code
/plugin list

You should see second-look listed. Try:

• Ask Claude "what makes a good test?" — the skill should inform the answer.
• Run /second-look tests/ in a project with tests.
• Run /agents — you should see test-auditor.

Usage

Writing new tests

Just work normally. The skill auto-triggers when you ask Claude to write tests, and it applies the five audit questions before committing any test. No special invocation needed.

Auditing an existing suite

Three ways, most explicit to least:

1. /second-look <path> — runs the full five-phase audit, returns a prioritized plan. Best for "give me a concrete action list."
2. Ask conversationally — "these tests are brittle, help me figure out what to do." The skill triggers and applies the framework.
3. Invoke the agent directly — "use the test-auditor to review src/auth/tests/" — useful when you want the structured subagent output without the command wrapper.

Adding always-on guidance (optional)

If you want the core principles in your project's CLAUDE.md so agents see them on every task without invoking the skill:

# From inside the plugin directory
cat CLAUDE-snippet.md  # copy the ## Testing section
# Paste into your project's CLAUDE.md

This is redundant with the skill for most workflows. Use it only if you've found the skill under-triggers in your project.

The philosophy in one page

Five audit questions, applied to every test:

1. Signal — does a CI failure immediately tell you what broke?
2. Maintenance — does the test need incidental updates when unrelated code changes?
3. Uniqueness — does this cover something no other test does?
4. Behavior, not implementation — does this survive refactors that don't change external behavior?
5. Readability — can someone grok it in 30 seconds?

A test that fails two or more is a candidate for rewrite or deletion.

Defaults:

• Integration tests over unit tests, most of the time.
• Mock at external boundaries only (DB, HTTP, time, randomness) — never internal collaborators.
• Table-driven tests (pytest.mark.parametrize) for combinatorial cases.
• Name tests after behavior, not functions.
• AAA structure (Arrange / Act / Assert).
• Fix flaky tests at the root cause — never retry.

Full details in skills/test-quality/SKILL.md and its references.

Updating

/plugin update second-look

Uninstall

/plugin uninstall second-look

Source material