hallucination-detective

by harrylabsj

Install this Skill

curl -fsSL -o hallucination-detective.zip https://skill.seembox.com/api/v1/skills/hallucination-detective/download && unzip hallucination-detective.zip

Download hallucination-detective.zip

SKILL.md About

---
name: Hallucination Detective
description: Learn to spot, verify, and handle AI-generated factual claims and confabulations.
version: "1.0.0"
type: prompt-flow
tags:
  - ai-hallucination
  - fact-checking
  - ai-verification
  - critical-thinking
  - ai-reliability
  - misinformation
author: Bell (design)
---

# Hallucination Detective

## Overview

Hallucination Detective is a practical guide to detecting AI hallucinations — those moments when AI confidently produces plausible-sounding but factually incorrect information. It teaches cross-referencing, source verification, confidence-assessment heuristics, and how to design prompts that reduce hallucination risk. Includes case studies of real AI errors.

This skill teaches methodology, not fact-checking as a service. It does not make determinations about the truth of specific claims.

## When to Use

Use this skill when the user asks to:

- Learn why AI hallucinates and how to spot it
- Verify an AI-generated claim they are unsure about
- Develop better fact-checking habits when using AI
- Understand how to reduce hallucinations through prompt design

**Trigger phrases:** "How do I know if AI is making things up?", "AI gave me a fact I'm not sure about", "How to fact-check AI output", "Do AI models lie?", "Why does AI hallucinate?"

## Workflow

### Step 1 — Greet and Set Context

Acknowledge the user's concern. Briefly explain what hallucination means in the AI context: confident-sounding outputs that are factually incorrect, fabricated, or internally inconsistent. Set expectations: this skill teaches detection and prevention methodology.

### Step 2 — Assess the Situation

Ask:
- What kind of AI output are they concerned about? (factual claim, citation, date, statistic, person)
- How confident did the AI sound?
- Have they already tried to verify any part of it?

### Step 3 — Explain Why Hallucinations Happen

Provide a clear, non-technical explanation:
- AI models are pattern predictors, not knowledge databases
- They optimize for plausible-sounding output, not truth
- Training data contains errors, contradictions, and gaps
- Models have no mechanism to "know what they don't know"
- Some topics (obscure facts, recent events, specific numbers) have higher hallucination risk

### Step 4 — Teach Detection Techniques

Walk through the verification toolkit:
- **Cross-reference check:** Does the claim appear in reliable external sources?
- **Specificity test:** Overly specific details (exact dates, quotes, statistics) are higher risk
- **Consistency check:** Does the AI contradict itself within the same response?
- **Source request:** Ask the AI "Can you cite a source for that?" and verify the source exists
- **Plausibility filter:** Does the claim pass basic common-sense checks?
- **Freshness awareness:** Information beyond the model's training cutoff is at higher risk

### Step 5 — Reduce Hallucinations Through Prompting

Teach prompt design strategies:
- Ask for confidence indicators ("Rate your confidence from 1-5")
- Request explicit "I don't know" responses when uncertain
- Ask for sources or reasoning chains
- Use "according to [specific domain]" framing
- Break complex factual queries into smaller, verifiable pieces

### Step 6 — Summarize and Exit

Recap key detection techniques and prevention strategies. Emphasize that healthy skepticism is a skill, not paranoia. Suggest related skills.

## Safety & Compliance

- Does not fact-check claims itself — teaches users methodology, does not make determinations about truth
- Does not encourage distrust of all AI; promotes balanced critical thinking
- Not a replacement for professional fact-checkers or subject-matter experts
- Does not target specific AI models or companies with accusations
- This is a descriptive prompt-flow skill with zero code execution, zero network calls, and zero credential requirements

## Acceptance Criteria

1. User's concern about AI output is assessed and contextualized
2. Why hallucinations occur is explained in accessible terms
3. At least 3 detection techniques are taught
4. At least 2 prevention prompting strategies are provided
5. Does not fact-check specific claims — teaches method, not determination

## Examples

### Example 1: Suspicious AI Output

**User says:** "ChatGPT told me a very specific historical fact with dates and names, but something feels off. How do I check if it's real?"

**Skill guides:** Explain hallucination causes. Walk through the verification toolkit: cross-reference the dates and names, check if sources exist, test for internal consistency. Show how to ask the AI for sources and then independently verify them.

### Example 2: Building Long-Term Habits

**User says:** "I use AI for research a lot. How do I build a habit of not just trusting everything it says?"

**Skill guides:** Focus on the prevention side. Teach confidence-assessment prompting, source-request habits, and the "verify-then-use" workflow. Provide a simple daily checklist for AI-assisted research.

_meta.json Metadata

{
  "ownerId": "kn77zzg9p845zanvy6vrf76k7d81mcnm",
  "publishedAt": 1778122925056,
  "slug": "hallucination-detective",
  "version": "1.0.0"
}