verification-before-completion

Requires running verification commands and checking output before claiming work is complete.

Updated Apr 29, 2026
One-click install
npx skills add https://github.com/fred-meng/harness --skill verification-before-completion-fred-meng
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: verification-before-completion
Source: https://github.com/fred-meng/harness/tree/main/.github/skills/verification-before-completion
Command: npx skills add https://github.com/fred-meng/harness --skill verification-before-completion-fred-meng

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? AI agents and developers often claim work is done, tests pass, or bugs are fixed without actually running the verification commands, leading to broken code, lost trust, and wasted rework. This Skill enforces an evidence-first discipline so no completion claim is made without fresh command output. ## Core Features & Use Cases - Verification Gate: A five-step gate (identify, run, read, verify, claim) that must be completed before any success statement, commit, or PR. - Claim-to-Evidence Mapping: A table mapping common claims (tests pass, build succeeds, bug fixed) to the exact command output required to support them. - Rationalization Detection: Red-flag phrases and excuse patterns ("should work", "just this once", "linter passed") that signal unverified claims. - Use Case: Before committing a bug fix, run the test suite, confirm zero failures in the output, verify the regression test fails without the fix, and only then report completion with the evidence attached. ## Quick Start Before claiming any task is complete, run the relevant verification command, read the full output, and only then state the result with the evidence.

Frequently Asked Questions about verification-before-completion

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I verify work is complete before committing code?

Identify the command that proves your claim, run it fresh and in full, read the entire output including exit code and failure counts, and only then state the result with that evidence. Never commit or open a PR based on assumed success.

How to verify a regression test actually catches a bug?

Use the red-green cycle: write the test, run it to confirm it passes with the fix, revert the fix and confirm the test fails, then restore the fix and confirm it passes again. A test that never fails proves nothing.

Does a passing linter mean the build succeeds?

No. A linter checks style and static issues but does not verify compilation. You must run the actual build command and confirm it exits with code 0 before claiming the build succeeds.

Can I trust an AI agent's report that a task succeeded?

No. Agent success reports must be independently verified by checking the version control diff and confirming the actual changes exist and work. Trusting the report alone is a common failure mode.

When should verification be skipped for small changes?

Never. The rule has no exceptions: partial checks, confidence, fatigue, and "just this once" reasoning are all explicitly rejected. Any claim of success without fresh verification evidence is treated as dishonesty.