Anti-hallucination plugin for Claude Code.
One command. Four constraints. Zero hallucinated "facts" in your research.
Type /research and Claude stops guessing.
| Constraint | What happens | Why it matters |
|---|---|---|
| Verify with citations | Every claim needs a source. No source = [retracted] |
Removes confident-sounding fiction |
| Direct quotes for factual grounding | Extracts exact quotes before analyzing | Kills paraphrase-drift |
| Chain-of-thought verification | Step-by-step logic before conclusions | Surfaces bad assumptions |
| External knowledge restriction | Only uses info from provided documents | --docs flag only |
Type exit research mode to go back to normal.
Plugin (recommended):
claude plugin add ibliminse/research-mode
Manual:
git clone https://github.com/ibliminse/research-mode.git ~/.claude/skills/research-mode/research
This activates 3 constraints: citations, direct quotes, and chain-of-thought. Claude can still use its general knowledge — it just has to be rigorous about it.
With a topic:
/research what caused the Change Healthcare breach
When you're analyzing specific documents and need to know exactly what came from your source vs. Claude's training data:
/research --docs
This adds a 4th constraint: external knowledge restriction. Claude will ONLY use information from documents you provide. No general knowledge mixed in.
/research --docs review this contract for liability issues
When to use --docs: Legal review, financial analysis, compliance audits — any time you need to trust that every insight came from YOUR documents, not Claude's training data.
When to skip --docs: General research, fact-checking, exploration — you want Claude's knowledge, just with citations and reasoning.
Anthropic published 7 anti-hallucination techniques. I split them:
3 should be always on (no downside):
- Say "I don't know" — never fill gaps with fiction
- Don't contradict yourself — catch your own conflicts
- Self-review before responding — one mental pass before answering
4 are a toggle (this plugin):
4-7 add rigor that's valuable for research but would slow down creative work. Citation requirements are overkill for "should I use flexbox or grid." Step-by-step reasoning is noise when you're brainstorming. Sometimes you want Claude to use general knowledge.
Research mode is a scalpel, not a lifestyle.
Add this to ~/.claude/rules/ or your CLAUDE.md:
## Anti-Hallucination (Always On)
1. If you don't have a credible basis for a claim, say so.
Never fill knowledge gaps with plausible fiction.
"I don't have enough information" is always a valid answer.
2. When making multiple related claims, verify they're consistent.
If you catch a contradiction, flag and resolve it before presenting.
3. For non-trivial responses, do one mental pass:
"Can I actually back up each claim I'm about to make?"
Cut anything you can't.This is a single markdown file (commands/research.md) that Claude reads and follows. No dependencies, no build step, no API calls, no external services.
When active, Claude structures output as:
Finding: [claim]
Source: [file path, URL, quote number, or named source]
Confidence: high / medium / low
Unsupported claims are retracted:
[retracted — no source found]
Based on Anthropic's official documentation: docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
MIT -- do whatever you want with it.