Dogwiz
Mar 24, 2026 · 2 min read · 266 words

Research Mode — X Post

Final Tweet (single tweet + graphic)

Anthropic published 7 anti-hallucination techniques for Claude

I keep 3 on permanently. The other 4 I toggle with one command

built it as a Claude Code plugin — open source

3 always-on: say "I don't know", don't contradict yourself, self-review 4 research-mode: citations, direct quotes, chain-of-thought, external knowledge restriction

github.com/dogwiz/research-mode

Graphic

graphics/research-mode-v4.png — 1760x1300, DogWiz paper cutout style

To publish

  1. Create GitHub repo: gh repo create dogwiz/research-mode --public --source ~/LC/Projects/Code/active/research-mode
  2. Push: cd ~/LC/Projects/Code/active/research-mode && git remote add origin ... && git push -u origin main
  3. Post tweet with graphic attached
  4. Reply with source link: docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations

Thread version (backup)

TWEET 1: I keep 3 anti-hallucination rules on ALL the time in Claude Code. most people only toggle them for research. here's why that's a mistake

TWEET 2: Anthropic published a "Reduce Hallucinations" page in their docs. it has 7 techniques — 3 basic, 4 advanced. everyone talks about the 3 basic ones. but they're using them wrong. here's the split that actually works:

TWEET 3: 3 rules that should ALWAYS be on: (1) say "I don't know" (2) don't contradict yourself (3) self-review before responding — zero downside

TWEET 4: The other 4 are a toggle via /research: (4) citations (5) direct quotes (6) chain-of-thought (7) external knowledge restriction

TWEET 5: Why not all 7 all the time? citation requirements slow creative work. step-by-step reasoning is overkill for simple questions. research mode is a scalpel, not a lifestyle

TWEET 6: open-sourced as a Claude Code plugin. github.com/dogwiz/research-mode