Agent-Skills.md

Agent Skills: Refutation

Verify a finished artifact by dispatching an independent skeptic that never saw it being made, tasked with refuting it, grounding every factual claim in a primary source, and reporting findings with severity and confidence. Invoke before handing over any deliverable that has no automatic falsifier — documents, rules, guidelines, skills, plans, analyses, decisions — above all one produced in the current session. Code already covered by tests, typecheck, or lint has a falsifier; use it there only for claims those checks do not exercise. Not for wording, style, or "does this look OK" passes.

UncategorizedID: hairihou/dotfiles/refutation

Repository

hairihou/dotfiles
hairihou

Install this agent skill to your local

pnpm dlx add-skill https://github.com/hairihou/dotfiles/tree/HEAD/src/.claude/skills/refutation

Skill Files

Browse the full folder contents for refutation.

Download Skill

Loading file tree…

src/.claude/skills/refutation/SKILL.md

Skill Metadata

Name
refutation
Description
Verify a finished artifact by dispatching an independent skeptic that never saw it being made, tasked with refuting it, grounding every factual claim in a primary source, and reporting findings with severity and confidence. Invoke before handing over any deliverable that has no automatic falsifier — documents, rules, guidelines, skills, plans, analyses, decisions — above all one produced in the current session. Code already covered by tests, typecheck, or lint has a falsifier; use it there only for claims those checks do not exercise. Not for wording, style, or "does this look OK" passes.

Refutation

Verify $ARGUMENTS by refutation, not review.

A reviewer who holds the creation context cannot find the flaw the context conceals. Independence is the mechanism, so it is not optional: the skeptic runs as a subagent, or the run is worthless.

1. Isolate the Artifact

Identify what is under test and reduce it to what a stranger could read on its own: file paths, or the artifact text itself. If it exists only in conversation, write it to a working file first.

Pass nothing about how it came to be. No session history, no author intent, no "X was chosen because Y", no summary of your own. A summary carries the author's framing, which is the thing being tested.

When the artifact changes an existing system, the scope is not the change. Name the system it belongs to and say the skeptic may read all of it. A change can be internally consistent and still contradict a third file that carries the same concept, and that flaw is invisible inside the diff.

2. Dispatch the Skeptic

One Agent call. Read the prompt from ${CLAUDE_SKILL_DIR}/agents/skeptic.md and substitute {{artifact}}.

3. Present, Do Not Decide

Report the findings ordered by severity, each with its evidence and confidence as returned. Do not accept, dismiss, merge, or act on any of them, and do not edit the artifact. Adoption, partial adoption, and rejection are the user's call, and the findings are material for that call.

Where a finding is Low confidence, keep the skeptic's note on what would settle it. That is the next action if the user wants one.

Common Mistakes

  • Skipping the run because the artifact looks correct, which is the condition the skill exists for
  • Passing through a factual claim that cites no source consulted