name: verification-before-completion description: "Check evidence before reporting work complete, fixed, passing, or ready. Use at a completion boundary or when later changes may have invalidated an earlier result." license: MIT compatibility: "AgentSkills-compatible verification workflow for coding, skill curation, research, and authorized security work." metadata: author: AeonDave version: "1.2"

Verification Before Completion

Match each completion claim to evidence for the state being delivered. evidence-before-claims owns general evidence quality and uncertainty; this skill checks freshness and coverage.

Completion gate

1. Name the outcome being claimed and the check that can detect its failure.
2. Run the relevant check after the last change that could affect its result. Preserve required repository gates.
3. Inspect exit status, substantive output, skipped cases, and the artifact or state actually checked.
4. Map the evidence to acceptance criteria. Report uncovered requirements and unavailable checks explicitly.
5. Investigate failures, correct in-scope causes, and rerun affected checks. Report unresolved failures and uncertain attribution. Once checks pass, repeat or broaden them only for a new change, failure, or unresolved concern.

A recorded result may be reused when its command, relevant inputs and checked state, environment, and coverage still apply. Compare changes since that result: a new turn, reader, commit, or unrelated edit does not alone invalidate it. Rerun affected checks when relevant inputs change; reverify when provenance or applicability cannot be established.

Select the check

Use diff inspection and required structural checks for low-impact editorial work; do not invent runtime tests that merely mirror text. Distinguish checks not run from checks that failed.

Load references/completion-evidence.md when choosing regression evidence, comparing implementation with an agreed design, or diagnosing an overstated completion claim.