rspec-best-practices
RSpec Best Practices
Use this skill when the task is to write, review, or clean up RSpec tests.
Core principle: Prefer behavioral confidence over implementation coupling. Good specs are readable, deterministic, and cheap to maintain.
Quick Reference
| Aspect | Rule |
|---|---|
| Spec types | Model: domain logic; Request: HTTP endpoints (prefer over controller); Job: background processing; Service/PORO: no Rails helpers; System: critical E2E only (slow) |
| Assertions | Test behavior, not implementation |
| Factories | Minimal — only attributes needed; use traits for optional states; prefer build/build_stubbed over create |
| Mocking | Stub external boundaries, not internal code |
| Isolation | Each example independent; no shared mutable state |
| Naming | describe for class/method, context for scenario |
| Service specs | Required: describe '.call' and subject(:result) for the primary invocation |
let vs let! |
Default to let; let! ONLY when object must exist before example runs |
| External service mocking | allow(ServiceClass).to receive(:method) — not instance_double; instance_double only for injected collaborators |
| Example names | Present tense: it 'returns the user', never it 'should ...'; NEVER contains the word "and" — split into separate examples |
aggregate_failures |
Use when asserting multiple related items in one example |
TDD Workflow
When driving new behaviour with RSpec, follow this sequence:
- Write the failing spec — pick the smallest spec type that exercises the intended behaviour (model > service > request > system).
- Run it and confirm the failure message — the error should be about missing code, not a setup problem.
- Implement the minimum code to make the spec pass.
- Refactor — clean up duplication and naming while keeping the suite green.
- Verify — run the full relevant spec file, then the suite, before committing.
Choosing the best first failing spec for a Rails change
| Change type | Start with |
|---|---|
| Pure domain logic | Model or PORO service spec |
| HTTP endpoint behaviour | Request spec |
| Background processing | Job spec |
| Cross-layer user journey | System spec (sparingly) |
Factory Design
Minimal factories only. Never rely on factory defaults for business logic — set explicitly or use traits. Avoid create when build/build_stubbed suffices.
Service Spec (anchor pattern)
RSpec.describe Invoices::MarkOverdue do
describe '.call' do
subject(:result) { described_class.call(invoice: invoice) }
context 'when the invoice is overdue and unpaid' do
let(:invoice) { create(:invoice, due_date: 2.days.ago, paid_at: nil) }
it 'marks the invoice overdue' do
expect { result }.to change { invoice.reload.overdue? }.from(false).to(true)
end
end
context 'when the invoice is already paid' do
let(:invoice) { create(:invoice, due_date: 2.days.ago, paid_at: 1.day.ago) }
it 'does not change the invoice' do
expect { result }.not_to change { invoice.reload.updated_at }
end
end
end
end
→ Full examples: EXAMPLES.md | Copy-paste templates: assets/spec_templates.md
Shared Examples
Use only when the same behavioural contract applies to multiple subjects without per-example let overrides. Avoid when each context needs different setup — that signals a wrong abstraction. → Example in EXAMPLES.md
One Behavior Per Example — NEVER "and" in Example Names
The word "and" in an it / specify description means the example is asserting two behaviors. Split it. One behavior per example. Applies to every spec type — model, request, service, job, mailer, system.
# BAD — two assertions; if the first fails, the second never runs
it 'returns 201 and creates the record' do; end
it 'saves the order and sends the confirmation email' do; end
it 'updates the user and logs the change' do; end
# GOOD — one observable outcome per example
it 'returns 201' do; end
it 'creates the record' do; end
it 'saves the order' do; end
it 'sends the confirmation email' do; end
Self-check before finalizing any spec: scan every it '...' / it "..." / specify '...' string for the word and (case-insensitive, word-boundary). Every hit is a split — no exceptions for "convenience" examples like 'returns nil and does not raise'.
Output Style
When asked to write or review RSpec specs, your output MUST satisfy each rule below. Each is graded independently — one violation drops the whole check.
- Spec file path mirrors the source:
app/foo/bar.rb→spec/foo/bar_spec.rb. # frozen_string_literal: trueas the first line of every spec file.RSpec.describeuses the full constant path (RSpec.describe Module::Class do), not a string.describe '#method'/describe '.class_method'for each method under test.context 'when ...'/context 'with ...'for scenario variations — never usecontextto group methods.letfor test data,let!ONLY when the object must exist before the action under test.- No
let_it_beunless the project already depends ontest-prof(checkGemfile.lockfirst). - NO "and" in any example description — split on every occurrence (see section above). This is the most-missed rule; do an explicit scan before returning the spec.
subject(:result) { ... }for service / PORO specs invoking.call.travel_to/freeze_timefor any time-dependent assertion — never set pastTime.nowor stubTime.currentdirectly.- External boundaries mocked at the class-method level (
allow(SomeClient).to receive(:method)); ActiveRecord finders are NEVER mocked.
Flaky Tests & Deterministic Assertions
| Cause | Fix |
|---|---|
| Time-dependent logic | freeze_time / travel_to; never set past dates as shortcut |
| State leakage | Each example sets up own state; avoid before(:all) |
| Async jobs | queue_adapter = :test + have_enqueued_job; never assert side-effects imperatively |
| External HTTP | WebMock / VCR; never allow real network in CI |
| DB state bleed | Transactional fixtures or DatabaseCleaner; never share let! across contexts |
| Race conditions | Explicit Capybara waits; avoid sleep |
| Imprecise assertions | change.from().to() over final state; exact values over be_truthy/be_falsey; never assert updated_at |