Verify or refute agent claims with evidence and a confidence score.
npx skills add https://github.com/tkstang/sequence --skill skeptic-tkstang
Please help me install this Agent Skill. Skill: skeptic Source: https://github.com/tkstang/sequence/tree/main/.agents/skills/skeptic Command: npx skills add https://github.com/tkstang/sequence --skill skeptic-tkstang
This Skill helps users verify or refute claims made by agents by gathering evidence from various sources.
Use the /skeptic command to verify a claim made by an agent.
High-intent search queries and answers about installing and using this skill.
To verify agent claims, you can use the /skeptic command to adversarially gather evidence and return a verdict with a confidence score. It searches for contradicting evidence first, then supporting evidence.
Adversarial evaluation is a claim verification method that actively seeks contradicting evidence before looking for supporting evidence. This approach ensures a rigorous confidence score by stress-testing the agent's initial statement.
Yes, claim verification can utilize tools like WebSearch, WebFetch, Bash, Read, Glob, and Grep to gather evidence. These tools allow the system to access files and the web to validate agent statements.
Yes, adversarial claim verification requires access to a Claude Code environment to function. It relies on this environment to execute tools like Bash, WebFetch, and Agent for evidence gathering.
You can get a confidence score for an agent's statement by running the /skeptic command. It evaluates the claim by gathering evidence and returns an inline verdict alongside the score.
The best way to refute an agent claim is through adversarial evaluation, which prioritizes finding contradicting evidence from various sources. It provides a clear verdict and confidence score based on the gathered facts.
tyevansVerify claims against source code, documentation, and external resources.
dalagerProduces numeric truth scores and JSON reports for code and agent outputs.
freedomw1987Compress AI development context and preserve critical task artifacts.
AvihusittonDetect logic flaws, race conditions, and security risks in code reviews.
