Purge unneeded tests — deletion as discipline

Tests are not assets. Tests are liabilities that pay rent by catching real bugs. Volume is not a quality signal — coverage percentage is not a quality signal — only the counterfactual matters: if I delete this test, can a real bug now reach prod? If the answer is no, the test is dead weight, and dead weight slows CI, breeds noise, and trains reviewers to ignore failures.

Modern insight (2025): TDD pairs with purge discipline. The same rigor that earns RED before GREEN earns deletion before keep — a test that cannot describe the bug it would catch should not exist. Mutation testing exposes which tests are actually load-bearing; the rest are cargo cult.

See python for pytest examples (dynamic-language carve-out). See typescript for jest/vitest examples (static-language redundancy). See rust for cargo test examples (compile-time-guaranteed redundancy). See keep-vs-delete-table for the language-agnostic decision rubric.

When to Apply

• Reviewing a legacy test suite where CI takes minutes per run
• Post-refactor sweeps — refactoring made some tests redundant; the refactor commit is the right place to delete them
• PR review where the diff adds tests that assert structure the type system already guarantees
• Onboarding to a codebase — flag suspicious patterns to surface for the original author
• Mutation testing reports — tests that survive every mutation are not catching anything

When NOT to Apply

• Standalone "delete tests" sweep PRs — these become unreviewable and bundle unrelated concerns. Deletions ride alongside the work that makes them safe to delete.
• Behavior-change PRs where you have not separately confirmed the deletion is safe
• Code paths under active migration where coverage is the only safety net
• Any test whose failure mode you cannot articulate — if you cannot say what bug it would catch, you also cannot say it catches none

The four mandates

The principles below are load-bearing. Internalize them; do not paraphrase.

1. A test exists to catch real bugs. If deleting the test would not let a bug reach production, delete the test.

This is the discriminator. Before keeping any test, ask: what bug, specifically, does this fail on? If the answer is "none I can name" or "a bug the compiler would catch", the test does not earn its keep. The bar is a real bug — not a hypothetical, not a "what if the implementation changes" — a concrete failure mode that a real change could plausibly introduce and that this test would catch.

2. Test contracts and boundaries: protocol compliance, error semantics, security invariants, integration across real I/O.

Tests earn their keep at boundaries. Protocol compliance (HTTP status codes, message formats, retry semantics), error semantics (what happens on malformed input, partial failure, timeout), security invariants (authz/authn enforcement, input validation, rate limits), and real-I/O integration (DB transactions, file I/O, network calls) are exactly where bugs hide and where the type system cannot help. These tests stay.

3. Do not test configuration shapes, constructor output fields, or struct assembly — the type system and constructors already guarantee those.

In statically-typed languages, the compiler already proves that a constructor returns the type it claims, that a struct has the fields it declares, and that a config object has the shape its type asserts. A test like assert User(name="x").name == "x" does not catch any bug a compiler does not — the only way it can fail is if User's type signature changes, in which case the test itself fails to compile and the assertion is moot. Delete it.

4. Do not test that a function returns exactly what you passed in.

Identity-passthrough tests (assert echo(x) == x, expect(passthrough(value)).toEqual(value)) prove nothing — they describe the function's signature, not its behavior. If echo is supposed to validate or transform x, test the validation/transformation. If echo is genuinely identity, the function is dead code.

Static-guarantee carve-out (mirror of ~/.claude/claude/system-prompt-baseline.md <directives>)

The carve-out for mandate 3 is language-dependent and must mirror the user's system-prompt-baseline.md testing charter exactly:

• Static-guarantee languages — Rust / TypeScript-strict / Kotlin / Java / C++ / OCaml: structural assertions are redundant. A test that asserts a struct has the fields the compiler proved it has catches no bug. Delete.
• Dynamic languages — Python / JavaScript / Ruby: there is no compile-time guarantee that a function returns the shape the docstring claims. A boundary shape/type test IS a real-bug test — a refactor that silently changes the return shape would slip past type hints (which are advisory, not enforced at runtime). Keep.

The split is not aesthetic — it is about what guarantee the language provides. See references/python.md vs references/typescript.md for the contrasting examples.

Decision rubric (summary; full table in references/keep-vs-delete-table.md)

Workflow

1. Identify candidate — a test that survives mutation, asserts structure, or has no clear failure scenario.
2. Articulate the bug it catches — write one sentence: "this test fails when ___ goes wrong." If you cannot complete that sentence with a real bug, the test is a candidate for deletion.
3. Check the static guarantee — for the test's language, would the compiler/type-checker already catch the bug? If yes, delete.
4. Check the boundary contract — does the test verify protocol/error/security/real-I/O behavior? If yes, keep.
5. Inject the bug — modify the production code to introduce the bug the test claims to catch. Run the suite. If the test still passes, the test does not catch that bug — delete.
6. Delete with rationale in the commit message — record why the deletion is safe, not just that it happened.

Constitutional Rules (Non-Negotiable)

1. Never delete a test as part of an unrelated change — deletions are atomic commits with their own rationale, even if they ride in the same PR as the work that makes them safe.
2. Never delete a test whose failure mode you have not understood — confusion is not a license; if you cannot articulate the bug, you cannot prove its absence.
3. Never reduce coverage of a security invariant — auth, authz, input validation, secrets handling — these stay even when redundant.
4. If conflict with ~/.claude/claude/system-prompt-baseline.md, system-prompt-baseline.md wins — this skill mirrors the user's testing charter; if drift is detected, system-prompt-baseline.md is the source of truth.

Validation Gates