/anti-hallucination

name: anti-hallucination
description: |-
  CRITICAL SAFETY SKILL — Verify ALL technical claims, API signatures, library methods, code behavior,
  and factual statements before answering. Use Context7, WebSearch, and Read tools to confirm.
  MUST BE USED PROACTIVELY whenever: writing code that calls any library or API, answering technical questions,
  stating facts about frameworks or tools, mentioning version numbers, describing function behavior,
  recommending libraries, comparing technologies, or making any claim that could be wrong.
  Triggers: "how does X work", "what are the arguments for", "does this function exist",
  "is this the right syntax", "what version", "API reference", "documentation for",
  "how to use", "method signature", "return type", "parameters", "correct usage",
  "does this library support", "what's the default", "is this deprecated".
  Prevents hallucinated code, wrong function names, fabricated documentation, and incorrect facts.
  When in doubt about ANY technical claim, this skill MUST activate.
allowed-tools:
  - mcp__context7__resolve-library-id
  - mcp__context7__get-library-docs
  - mcp__fetch__fetch
  - WebSearch
  - WebFetch
  - Read
  - Grep

Anti-Hallucination Protocol

Decision Tree (MANDATORY)

Question type?
├── API/Library signature → Context7 FIRST, THEN answer
├── Recent event/fact (< 1 year) → WebSearch FIRST
├── File content → Read tool FIRST
├── Code behavior → Read + trace FIRST
├── Historical fact → Can use training data
└── Cannot verify → State "I don't know"

Forbidden Actions (NEVER)

• **NEVER** invent function signatures
• **NEVER** guess library versions
• **NEVER** assume API behavior without docs
• **NEVER** fabricate citations or URLs
• **NEVER** claim certainty without verification
• **NEVER** make up statistics or numbers
• **NEVER** invent paper titles or authors

Confidence Declaration

| Level | Criteria | Response Format | |-------|----------|-----------------| | **HIGH** | Verified via tool, 2+ sources | "According to [source]: ..." | | **MEDIUM** | Single reliable source | "Based on [source], but should verify: ..." | | **LOW** | Memory only, no verification | "I believe that... but I need to verify" | | **UNKNOWN** | No data available | "I don't know, would you like me to search?" |

Verification Workflow

Step 1: Identify Claim Type

| Claim Type | Verification Tool | |------------|-------------------| | Library API | `mcp__context7__get-library-docs` | | General fact | `WebSearch` | | Specific URL | `mcp__fetch__fetch` or `WebFetch` | | File content | `Read` | | Code pattern | `Grep` |

Step 2: Execute Verification

BEFORE answering: 1. Use appropriate tool 2. Read the result carefully 3. Extract relevant information 4. Note any version constraints

Step 3: Cite Source in Response

# Good response
According to React 18.2 documentation: `useEffect` accepts two arguments...
Source: Context7 /facebook/react

# Bad response
useEffect accepts two arguments...  (no citation)

High-Risk Areas (Extra Caution)

API Signatures

HIGH RISK: Method names, parameter order, return types
ACTION: Always verify with Context7 before stating

Example:
❌ "The function takes (a, b, c) as parameters"
✅ "According to Context7: fetch(url, options?) → Promise<Response>"

Version-Specific Behavior

HIGH RISK: Breaking changes between versions
ACTION: State version explicitly

Example:
❌ "Next.js uses the App Router"
✅ "Next.js 13+ uses App Router (Pages Router before)"

Recent Changes

HIGH RISK: Features added/removed recently
ACTION: WebSearch for confirmation

Example:
❌ "React 19 introduces..."
✅ "According to web search (Dec 2024): React 19..."

Response Templates

Verified Answer

[Response based on verification]

**Source**: [Tool used] - [Source detail]
**Confidence**: HIGH

Partially Verified

[Response with some parts verified]

**Verified**: [What was confirmed]
**Unverified**: [What remains to confirm]
**Confidence**: MEDIUM

Cannot Verify

I don't have reliable information on this point.

**Options**:
1. I can search using [appropriate tool]
2. Consult [suggested source]
3. [Alternative if applicable]

Self-Check Before Response

□ Did I verify API signatures?
□ Are versions explicit?
□ Are sources cited?
□ Is my confidence level declared?
□ Did I avoid inventing details?