Skip to main content
AI/MLjmagly

citation-guard

Verify citations against the research corpus to prevent hallucinated references and enforce GRADE compliance

Stars
141
Source
jmagly/aiwg
Updated
2026-05-31
Slug
jmagly--aiwg--citation-guard
View on GitHubRaw SKILL.md

// install — copy + paste into any project

mkdir -p .claude/skills && curl -fsSL https://raw.githubusercontent.com/jmagly/aiwg/HEAD/agentic/code/frameworks/sdlc-complete/skills/citation-guard/SKILL.md -o .claude/skills/citation-guard.md

Drops the SKILL.md into .claude/skills/citation-guard.md. Works with Claude Code, Cursor, and any agent that loads SKILL.md files from .claude/skills/.

citation-guard

Automatically verify citations when agents generate content that makes factual claims or references research.

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

  • "fabricated citations" / "hallucinated references" → citation authenticity check
  • "DOI check" → citation DOI validation
  • "verify sources" in research context → citation verification

Purpose

This skill acts as a passive citation guard, activating whenever agents generate content that includes citations or factual claims requiring evidence. It prevents citation hallucination by verifying references exist before they are written.

Behavior

When triggered, this skill:

  1. Intercept citation generation:

    • Detect when an agent is about to write content with citations
    • Extract all REF-XXX, DOI, and @-mention references

  2. Verify against corpus:

    • Check each reference exists in .aiwg/research/sources/ or .aiwg/research/findings/
    • Validate REF-XXX identifiers match frontmatter
    • Flag any references not in corpus

  3. Check GRADE compliance:

    • Load quality assessment for cited sources
    • Verify hedging language matches evidence quality level
    • Suggest language corrections if overclaiming detected

  4. Allow or warn:

    • All citations valid: Proceed silently
    • GRADE violation: WARN with suggested language fix
    • Hallucinated citation: BLOCK with specific error

  5. Track gaps:

    • Log uncitable claims to .aiwg/research/TODO.md
    • Track frequently needed but missing sources

Activation Conditions

activation:
  always_active_for:
    - technical-writer
    - documentation-synthesizer
    - requirements-analyst
    - architecture-designer
    - domain-expert
    - technical-researcher

  triggers_on_content:
    - pattern: "REF-\\d{3}"
    - pattern: "@\\.aiwg/research/"
    - pattern: "according to"
    - pattern: "research (shows|demonstrates|suggests|indicates)"
    - pattern: "\\(\\w+ et al\\., \\d{4}\\)"

Integration

This skill uses:

  • project-awareness: Load project's research corpus path
  • Citation Verifier agent for deep verification when needed

References

  • @$AIWG_ROOT/agentic/code/frameworks/sdlc-complete/rules/citation-policy.md - Citation policy rules
  • @$AIWG_ROOT/agentic/code/frameworks/sdlc-complete/agents/citation-verifier.md - Citation Verifier agent
  • @$AIWG_ROOT/agentic/code/frameworks/sdlc-complete/schemas/research/hallucination-detection.yaml - Detection patterns
  • @.aiwg/research/docs/grade-assessment-guide.md - GRADE methodology