TDD-Verified — Proof-First Development with Evidence Anchors

Register: DISCIPLINE — Iron Laws apply. Non-negotiable.

This skill enforces proof-first development. For behavior code and bug fixes, the default proof mode is RED-GREEN-REFACTOR with VHEATM evidence anchors at every gate. For migrations, config, docs, or spikes, declare an alternate proof mode before implementation.

Core principle: Evidence before claims. If you didn't watch the test fail, you don't know if it tests the right thing. If you didn't run verification, you cannot claim it passes.

Violating the letter of these rules is violating their spirit.

The Iron Laws

1. NO IMPLEMENTATION WITHOUT A DECLARED PROOF MODE
2. NO BEHAVIOR CODE WITHOUT A FAILING TEST FIRST
3. NO COMPLETION CLAIM WITHOUT FRESH VERIFICATION EVIDENCE
4. NO FIX CLAIM WITHOUT A FIX ANCHOR (file:line re-read after change)

For behavior code, writing code before the failing test means delete it and start over. For other work, the declared proof mode defines the evidence gate.

Proof Mode Declaration

Declare one before implementation:

No completion claim is allowed until the proof mode has T1/T2 evidence.

RED-GREEN-REFACTOR with Evidence Gates

RED — Write Failing Test

Write one minimal test showing exactly what should happen.

✅ Clear name: "rejects empty email address"
✅ Tests behavior: assert result.error == 'Email required'
✅ Real code: no mocks unless unavoidable
✅ One thing: no "and" in test name

Evidence gate — verify RED:

# Run test. Must FAIL. Document the failure output.
<test command> <test file>::<test name>

Evidence Anchor: test fails with "<exact failure message>"

If test passes immediately → you're testing existing behavior. Fix test. If test errors → fix error, re-run until failure is intentional.

You cannot proceed to GREEN without documented failure evidence.

GREEN — Write Minimal Code

Write the simplest code that makes the test pass. Nothing more.

YAGNI: no extra parameters, no future-proofing, no "while I'm here" changes.

Evidence gate — verify GREEN:

# Run test. Must PASS. Run full suite. Must stay green.
<test command>

Evidence Anchor: test passes — N/N suite green

If test fails → fix code (not test). If other tests break → fix immediately before proceeding.

You cannot proceed to REFACTOR without documented passing evidence.

REFACTOR — Clean Up

After GREEN only: remove duplication, improve names, extract helpers. Keep tests green. Add zero new behavior.

Evidence gate — verify REFACTOR:

<test command>
Evidence Anchor: suite still N/N green after refactor

Fix Anchor Requirement (VHEATM integration)

For every bug fix, before claiming "fixed":

Fix Anchor: <file>:<line> — <one-line description of change>
Action: re-read that file section AFTER the fix
Status: VERIFIED | ASSUMED

ASSUMED = you didn't re-read the file. Mark explicitly. VERIFIED = you re-read after the fix. Required before marking fixed.

After fix → trigger pattern-globalize: does this bug class exist elsewhere?

Verification Checklist

Before marking ANY work complete:

• [ ] Proof mode declared before implementation
• [ ] Every new behavior/function/method uses TEST_FIRST unless another proof mode is explicitly justified
• [ ] Watched each test FAIL before implementing (have failure output)
• [ ] Each test failed for expected reason — feature missing, not typo
• [ ] Wrote minimal code to pass each test
• [ ] All tests pass — ran full suite, have output
• [ ] Fix Anchor documented for every bug fix — re-read performed
• [ ] pattern-globalize run for every confirmed bug
• [ ] verification-before-completion checked before claiming done

Cannot check all boxes? You skipped proof-first discipline. For behavior code, start over with TDD. For other modes, stop and gather the required proof evidence.

Common Rationalizations — All Rejected

Red Flags — Stop and Start Over

• Code written before test
• Test passes immediately after writing
• Can't show test failure output
• "Should work now" without running verification
• Fix claimed without Fix Anchor re-read
• Completion claimed without evidence
• Any variation of "probably", "seems to", "should"

For behavior code, these mean: delete, start over with TDD. For non-behavior work, stop and re-enter through the declared proof mode.

Pressure Tests — Common Rationalizations

Những reasoning patterns dẫn đến skip skill này. Nhận ra chúng. Reject chúng.

• "Tôi sẽ viết test sau khi code xong để nhanh hơn" → Không có "sau." Code trước test = không có failing test = không biết test đang test đúng thứ gì. Delete và start over.
• "Feature này quá đơn giản để cần test trước" → Simplicity bias. Đơn giản nhất là lúc TDD dễ nhất. Nếu không thể viết test trước, design chưa đủ rõ.
• "Test đang pass rồi nên không cần chạy lại" → Evidence Anchor yêu cầu fresh run trong message này. "Was passing" ≠ "is passing." Chạy lại.
• "Mock quá phức tạp nên skip" → Mock complexity là design smell. Refactor interface trước. Nếu không thể test, không thể implement đúng.

Gotchas — Observed Failure Points

<!-- Populated by knowledge-compound skill after cycles where this skill underperformed --> <!-- Format: [YYYY-MM-DD] What failed | Root cause | What to do instead --> <!-- DO NOT pre-populate with speculation — real observations only --> <!-- Example: [2026-06-01] Agent skipped risk scoring for "trivial" CSS task that touched auth config | Triage gate not read carefully | Triage gate now checks file content, not just task description -->