VerificationQualityAnti-hallucinationAI Agent Skill
verification-before-completion
Forces a verification pass before declaring anything "done" - cross-checking every claim against its original source instead of trusting the summary.
Install Options
Install into Claude Code, Cursor, Codex, or Antigravity with:
$ npx skills add obra/superpowers@verification-before-completion
Real Work Situation Solved
Filter out hallucinations and over-extrapolations in AI feedback summaries
“AI summarized 100 user tickets into a neat report - how do you verify it didn't invent trends?”
5-Step Action Framework
- 1Require the AI to attach direct ticket IDs or exact verbatim quotes to every listed insight.
- 2Spot-check 5 random citations against raw logs - if even one is fabricated, reject the summary.
- 3Force the AI to output exact mention counts rather than vague qualifiers like 'most users'.
- 4Compare summary points against lowest-rated (1-star) feedback to ensure critical negative signals were not smoothed away.
- 5Retain only insights backed by at least 3 distinct, independent feedback instances.
Before
"The AI summary reads smoothly and makes sense, forward it straight to leadership."
After
"Caught AI fabricating a 'huge demand for dark mode' from a single mention. Enforced ticket citations and counts: only 3 issues with >10 occurrences made the roadmap."
Related Course Lesson
AI Literacy • Lesson 3
SKILL.md Source Instructions
SKILL.md • Read-only preview# Verification Before Completion Skill Before accepting a summary or report as final, spot-check its claims against the underlying raw data; if even one claim doesn't hold up, the whole summary is rejected.