Verification Before Completion Skill

Verification Before Completion

Overview

Never claim something works without fresh evidence proving it.

Core principle: NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE.

The Rule

Before saying anything is done, working, passing, or fixed:

Identify what command proves the claim
Execute the complete command freshly (not from memory)
Read the full output, check exit codes
Verify the output actually confirms your claim
Then make the claim

Skip any step = lying, not verifying.

Forbidden Language (Without Evidence)

These phrases are BANNED unless you just ran verification:

| Phrase | Problem | |--------|---------| | "Tests pass" | Did you run them? Right now? | | "Should work" | Should != does | | "Build succeeds" | Show me the output | | "Fixed" | Prove it | | "Done" | Verified how? | | "Looks good" | Based on what evidence? | | "Probably works" | Probably != verified | | "That should do it" | Run the tests |

This applies to paraphrases and synonyms too. Any implication of completion without evidence violates this rule.

What Counts as Evidence

| Claim | Required Evidence | |-------|-------------------| | "Tests pass" | Fresh test run output showing all pass | | "Build succeeds" | Fresh build output with exit code 0 | | "Bug is fixed" | Test that reproduced bug now passes | | "Feature works" | Test exercising the feature passes | | "No regressions" | Full test suite output, all green | | "Linting clean" | Fresh linter output, no violations |

Partial evidence is NOT evidence:

Linter passing != build succeeds
One test passing != all tests pass
Agent reporting success != independently verified
"It compiled" != "it works"

Common Violations

The Premature "Done!"

# BAD: Claiming done after writing code
"Added the validation. Done!"

# GOOD: Verify first
*runs tests*
"Added the validation. Tests pass (47/47). Done."

The Optimistic Fix

# BAD: Assuming fix works
"Fixed the null check. That should resolve the crash."

# GOOD: Prove it
*runs failing test*
"Fixed the null check. The reproducer test now passes, full suite green (128/128)."

The Stale Memory

# BAD: Remembering a previous run
"Tests were passing earlier."

# GOOD: Run again
*runs tests now*
"Tests pass as of right now (output above)."

Quick Reference

About to say "done" / "works" / "passes" / "fixed"?
  ├── Did you JUST run the verification command?
  │   ├── Yes → Did the output confirm your claim?
  │   │   ├── Yes → Make the claim (with evidence)
  │   │   └── No → Fix the issue, verify again
  │   └── No → Run it now. Then decide.
  └── Not sure what command to run?
      └── Figure that out first. Can't verify = can't claim.

Red Flags

Stop if you catch yourself:

Saying "done" without running tests
Using words like "should", "probably", "likely" about completion
Relying on a test run from 5 minutes ago
Trusting a subagent's report without checking
Claiming something works because "the code looks right"
Expressing satisfaction ("Great!", "Perfect!") before verification

Honesty is a core value. Claiming success without evidence is dishonest, even if you believe it's true.

Agent Skills: Verification Before Completion

Install this agent skill to your local

Skill Files