Verification Before Completion

Core principle: Evidence before claims. No exceptions.

The Rule

BEFORE claiming any status:
1. IDENTIFY what command proves this claim
2. RUN the command (fresh, complete, not from cache)
3. READ full output — check exit code, count failures
4. Does output confirm the claim?
   YES → State claim WITH evidence (paste output)
   NO  → State actual status with evidence
5. ONLY THEN make the claim

What Counts as Verification

Work Type	Verification
Code changes	Run tests: <test command per PROJECT.md> → show "0 failures"
Bug fix	Reproduce original bug → confirm it's gone + regression test passes
Lint/style fix	Run linter: <lint command per PROJECT.md> → show "0 errors"
Build change	Run build → show exit code 0
Documentation	Re-read the doc — does it match the current code?
Config change	Restart/reload → confirm new behavior
Refactor	Run FULL test suite — refactors must not change behavior
Project without tests	Manually verify the feature works — describe what you did and saw

Output Evidence Format

PASS Tests pass: <test command> → 34/34 passed, 0 failures (exit code 0)
PASS Lint clean: <lint command> → 0 errors, 0 warnings
PASS Build: <build command> → exit code 0

Never:

FAIL "Tests should pass now"
FAIL "Looks correct"
FAIL "I'm confident this works"

Key Patterns

Tests:

PASS [Run test command] [See: 34/34 pass] "All tests pass"
FAIL "Should pass now" / "Looks correct"

Regression tests (Green-Red-Green verification):

PASS Verify regression coverage (assumes fix is already in place — for writing the test itself, follow TDD red-green cycle):
   With fix applied → Write test → Run (PASS confirms test covers the fix)
   → Revert fix → Run (MUST FAIL — confirms test catches the bug)
   → Restore fix → Run (PASS — confirms fix works)
FAIL "I've written a regression test" (without revert verification)

Build:

PASS [Run build] [See: exit 0] "Build passes"
FAIL "Linter passed" (linter ≠ compiler)

Requirements:

PASS Re-read plan → Create checklist → Verify each item → Report gaps or completion
FAIL "Tests pass, phase complete" (tests ≠ requirements)

Agent delegation:

PASS Agent reports success → Check VCS diff → Verify changes independently → Report actual state
FAIL Trust agent report (agents lie)

Red Flags — STOP and Verify

You're about to say...	Instead...
"Should work now"	Run the command. Paste the output.
"Done!" / "Great!" / "Fixed!"	Run verification FIRST, then celebrate.
"I'm confident"	Confidence ≠ evidence. Run it.
"The agent reported success"	Verify independently. Agents lie.
"Partial check is enough"	Run the FULL command. Partial proves nothing.
"Just this once"	No exceptions. Ever.

Evidence Strength

When providing gate evidence, note the evidence type (see GATES.md for full classification): (C) command evidence is strongest, (D) diff evidence is medium, (S) self-attestation is weakest. If a stronger evidence type is available, use it — don't downgrade to self-attestation.

For projects without test suites: "I manually verified" is insufficient. Execute the code and provide observable output (command results, HTTP responses, REPL output). The Manual verification gate in GATES.md requires this when the Tests gate is SKIP N/A.

Relationship to Gates

This skill is the implementation of GATES.md post-task gates. When GATES.md says "Tests pass", this skill defines HOW to verify that. Don't duplicate gate checks — use this skill's process to provide evidence for each gate.

Transitions

Invoked by: mbscode:executing-plans, mbscode:subagent-driven-development (post-task gates), and mbscode:brainstorming (trivial tasks) This is a cross-cutting obligation, not a sequential phase. It applies within every other skill whenever a completion claim is about to be made. After verification passes, continue with whatever skill you were using.